From August 18, a prosecutor in one EU country can order your data from a company in another EU country. Directly. In a lot of cases without a judge in the second country looking at it first.

Ten days to hand it over. Eight hours if they call it an emergency.

That's called the e-Evidence Regulation, passed in 2023. And it's huge… and completely flying under the radar right now.

Ten days

Two new instruments are key. A European Production Order, which makes a provider hand over data. And a European Preservation Order, which makes them freeze it so it's still there when the first one arrives.

The data comes in categories, and the bar rises with them. Subscriber data (who is behind this account) is an easy one. Traffic and content data need a criminal offense that carries a maximum sentence of at least three years in the issuing country, or one from a short list: terrorism, child sexual abuse, payment fraud, some cybercrime.

If a provider doesn't play along, the penalty is up to 2% of total worldwide annual turnover. Late and partial answers count too.

And companies outside the EU are not out of reach. If you offer your service to people in the Union, you have to name an establishment or a legal representative inside it, and that representative is the address these orders go to. Banks and other financial services are the one big carve-out.

So, every country is a potential target here.

Ten days for replacing a process that averaged about ten months before. Wow.

Mutual legal assistance between two European countries could take the better part of a year, and half of it was paperwork moving between ministries. For an investigation, that's basically a failure from the start.

Same shape

The reason so many European governments started moving off American cloud services was one sentence in a French parliamentary hearing. Microsoft could not guarantee that data on French servers would never go to US authorities. Because of the CLOUD Act, an American law that reaches American companies wherever their servers happen to stand.

That answer turned digital sovereignty from a policy preference into a security question, at least for procurement departments. I've written about it more than once.

The two laws are not the same thing, and I don't want to pretend they are. An e-Evidence order comes from a judicial authority, not an agency. There are grounds to refuse: immunities and privileges, a manifest breach of fundamental rights, a conflict with the law of a third country. There are legal remedies for the person affected, at least on paper.

But the shape is the same. An authority in one country reaches data held in another, and the second country's courts are not necessarily part of the decision.

The question was never only which country the servers sit in. It's which authorities can reach them.

Scope

Telcos and hyperscalers are the obvious targets. The definition goes further: electronic communications services, domain name and IP services, and other information society services that either let users talk to each other or store data for them.

Read that slowly and you land on the places most of us publish. A newsletter platform stores your posts and your subscriber list, and lets people comment. A hosting company keeps your site. A forum does both.

I don't know which legal representatives the platforms I use have designated, and I couldn't find it published anywhere. Maybe that's coming next week. Maybe it's already in a filing.

One thing I got a bit wrong before (in hindsight now). I wrote about kSuite as the Swiss answer to Google Workspace, and part of the argument was that Swiss hosting sits outside the CLOUD Act. That's still true. It doesn't sit outside this one, though, if the provider offers services in the EU. Different law, different door.

The catch… there always is one

There is a notification mechanism, and it was supposed to be the safeguard. A second authority sees the order and can object before rights get run over.

It got thinner in negotiation, apparently. Preservation orders don't trigger it at all. Neither do production orders for subscriber and traffic data when they're used to identify a user, which is more or less the exact move you make when you want to know who a journalist has been talking to. EDRi has been saying this since 2021. It's coming though.

Some things are unclear, still. Not every EU judiciary is in the same state, and some are under Article 7 proceedings for exactly that reason. An order from one of them lands with the same ten day deadline as an order from anywhere else.

The professions this bites first are: Journalists and their sources. Lawyers and their clients. Doctors. Whether an immunity applies is a question somebody has to notice, and the notification may not reach anybody in a position to notice it. Fast enough.

And none of this replaces the CLOUD Act. It's still there. European data now has two doors instead of one.

To be fair

The European Investigation Order runs on a 120 day deadline, mutual legal assistance on nothing you could plan around, and evidence that sits on a server in a neighboring country was often just gone by the time the request cleared.

Prosecutors have been asking for this for a decade. The three year threshold is a limit that holds. So is keeping content data behind a higher bar than subscriber data, and the refusal grounds are more than most cross-border regimes come with.

You can think the old process was unusable and still think this version gave away too much. Both things fit.

The Bottom Line

I've spent a fair bit of the last two years writing that Europe should run its own infrastructure, and I still think that. Nothing here changes it.

But the worry underneath those articles is not the flag on the company. It is how few steps stand between an authority and my data, and how little I'd hear about it.

From Tuesday, there are fewer steps inside Europe too.

Is a European prosecutor better than an American one? Don't know. Maybe.. Is that what we meant by sovereignty? I highly doubt it...

If you're not already, follow me on Medium.

Facebook removed about 3.5 billion fake accounts last year. Billion with a B. That is almost ten million a day. Numbers we can't grasp.

On Deezer, more than half of all new songs uploaded every day are fully AI-generated. Around 90,000 tracks. Daily. Who uses Deezer, by the way...

Numbers like these used to sound like a conspiracy theory. Now they come from transparency reports.

I don't think the internet is dead. Not yet, not really. But it's starting to crumble a bit, isn't it?

That feeling has a name. The Dead Internet Theory.

The theory

It started in 2021, on a forum called Agora Road's Macintosh Cafe. The thread: "Dead Internet Theory: Most of the Internet is Fake." The claim: most of what we see online isn't made by humans anymore, bots create the content, bots click it, and a shadowy group steers the whole thing to manipulate us.

The Atlantic picked it up the same year and filed it, politely, under conspiracy theory.

The shadowy-group part still is one. Nobody is steering this… I hope...

But the other half of the theory, machines making content for machines, stopped being a theory a while ago. Large parts of it are just measurement now.

We notice that as writers too. Writing is done by AI. Reading is increasingly done by AI too. So where are we headed?

The numbers

Imperva, a security company that counts bot traffic for a living, published its report for 2024: automated traffic passed human traffic for the first time.

51 percent of all web traffic is bots. 37 percent is malicious bots.

Music: Spotify deleted 75 million spammy tracks in twelve months. On Deezer, AI uploads went from 10 percent of new tracks in January 2025 to over half today. And in a Deezer and Ipsos survey, 97 percent of listeners couldn't tell AI music from the human kind.

Couldn't tell… is the key here.

Books: researchers tracked 14,419 self-published genre novels on Amazon. One in five contained substantial AI text. Among new entrants into the top-25 bestseller ranks, it was 31 percent by 2026. Not one of these books disclosed it.

The readers still lean human, but how long? The AI books are 20 percent of the catalog but only about 12 percent of the sales.

For now.

Bots liking bots

The theory's darkest scenario, sort of, is weirdly mundane. A bot wishes you a happy birthday. Another bot likes the post and says thanks. Somewhere a dashboard counts two engagements, and no human was involved at any point.

I see the small version of this under my own articles. Polite, well-structured comments, vaguely related to the piece, and clearly written by a machine. I wrote about those comments a while ago.

The feeds probably make it worse. They surface whatever people interact with, and AI slop is strange enough that we stop and stare. Every stare tells the algorithm to show it to more people. And some (or more) of those "people" are bots too.

See the problem...

Model collapse

There's a second loop underneath. AI models train on internet text. The internet fills with AI text. So the models increasingly train on their own output.

Researchers showed in Nature what happens then: the models degrade. The output gets less diverse, more repetitive, more wrong. They call it model collapse.

Nobody knows how far that goes in practice. The labs know about it and filter their training data. But the direction is set: the more the internet fills with synthetic content, the harder it gets to train on... the internet.

There is something almost fair about that.

Human corners

Detection tools won't save us. Nope. They are easily fooled, or wrong in the first place. Watermarks may help for a while, until they don't.

So the likely future is the reverse: we won't label what's fake. We'll verify what's real.

I've said before that it doesn't matter if AI wrote it. If a story is a good read, it's a good read. I still think that.

But the Dead Internet Theory isn't asking who wrote the story. It's asking whether anyone is reading it afterwards.

People might turn out to want content from other people. If that's true, either the existing platforms find ways to surface verified humans, or new networks appear that do (which is a cool thing to think about).

If not, the split becomes financial too: some pay for human work, the rest scroll whatever the feed pours out.

The Bottom Line

I make a big part of my living on this internet today. Real readers, as far as I can tell. Thank you!

"As far as I can tell" is the new wording, though. Because sometimes I can't tell.

The internet isn't dead. The humans are all still here somewhere, writing, reading, arguing, buying odd stuff. We're just getting harder to find these days. And we're losing on numbers, for sure.

Worth looking for each other, though… don't you think?

If you're not already, follow me on Medium.

Get Substack SEO

Hi friends,

somewhere between exporting my Medium archive and my Substack archive, this became backup week.

Neither export is a real backup. One doesn't know my subscribers, the other has no images. So I wrote down what to save on top, for both platforms. Two more posts fit right in: my addresses on my own domain, and my earnings in one place where I can see them. Plus one about the automatic transmission, which has nothing to do with any of this.

Here's my week.

How to Really Back Up Your Substack (Posts, Subscribers, and URLs)

Substack's export has no images, no drafts, no Notes, and 84% of it is mail tracking. A real backup takes a few more steps, and your own domain is the biggest one.

Read it on Medium or on Substack.

How to Really Back Up Your Medium (Posts, Images, and Subscribers)

Medium's export knows every IP address I ever logged in from. My images and my newer subscribers? Not in there. Here is what to save on top, and where it hides.

Read it on Substack.

Use Subdomains to Put Your Domain on Every Platform

stories.byburk.net for Medium, letters.byburk.net for Substack, shop.byburk.net for the shop. One domain, all the addresses. This one started as a reader question in the comments.

Read it on Medium or on Substack.

I Built a Menu Bar Revenue Dashboard With Claude Code

I got tired of opening six dashboards to check what I earn. So I used Claude Code to build a Mac menu bar app that shows Medium, Substack, Gumroad, and PayPal in one click. About 600 lines of Python, no dependencies.

Read it on Medium or on Substack.

Who Invented the Automatic Transmission?

Two brothers who were both named Thomas, a gearbox that sometimes exploded, and the $57 option that finally got it right. It took 40 years.

Read it on Medium or free on fun-facts.net.

📎 From the Archive

Still the question I get when I talk about backing up a Substack: why not just a blog then? This one from last September is my answer, and it hasn't moved much.

Read it on Substack.

💡 Fun Facts

MySpace lost every song uploaded between 2003 and 2015 in a server migration. Around 50 million tracks from 14 million artists. The company confirmed there is no way to recover them.

GitLab deleted its own production database in 2017. They had five backup methods. None of them worked. What saved them was a copy an admin had made by hand, six hours earlier.

World Backup Day is March 31. On purpose the day before April Fools'.

PS: Fun facts have their own home now. I started a little blog at fun-facts.net.

That's #159. Back soon. ✌️

Get the Substack Automation Blueprint

My Medium friends can read this over on Medium.

In 1998, an animator at Pixar watched Woody disappear. First his hat vanished from the file system. Then his boots. Then the rest of him.

Someone had run /bin/rm -r -f * on the server holding Toy Story 2. The command does exactly what it says: delete everything, no questions asked. By the time somebody pulled the plug, about 90 percent of the film was gone. Two years of work.

Not a problem, right? That's what backups are for.

Pixar's backups had been failing quietly for about a month. Nobody had noticed.

Maternity leave

The movie survived for one reason. Galyn Susman, the film's supervising technical director, had recently had a baby and was working from home. Her computer there held a copy of the film, refreshed regularly so she could keep working.

They drove the machine back to the studio, plugged it in, and copied the film over. Most of it came back.

Toy Story 2 went on to make about half a billion dollars at the box office. And Pixar got the lesson in its purest form: a backup only counts if someone checks that it works.

My favorite detail: the rescued version isn't the one you saw. Months after the file rescue, Pixar decided the movie wasn't good enough and remade most of it in nine months. Toy Story 2 was deleted twice. Once by accident, once on purpose.

It keeps happening

In 2019, MySpace confirmed it had lost every song uploaded between 2003 and 2015 in a server migration. Around 50 million tracks from 14 million artists. The company's statement: there is no way to recover the data.

Ten years earlier, the T-Mobile Sidekick had already shown what happens when the cloud is the only copy. Its servers, run by a Microsoft subsidiary, lost the contacts, photos, and calendars of around 800,000 users. The compensation was a $100 voucher.

And in 2017, GitLab had five backup mechanisms in place when a tired admin deleted the production database. None of them worked. What saved the company was a manual snapshot a colleague happened to have made six hours earlier.

Five backups. Zero backups. Same thing, apparently.

Paper too

This is not a digital invention. In 1973, the US National Personnel Records Center in St. Louis burned for 22 hours and took up to 18 million military service files with it. No duplicates, no microfilm. Archivists still reconstruct veterans' records from scattered sources today.

The US Patent Office burned down in 1836 with the entire record of American invention inside, and the government's restore plan was asking inventors to mail in their own copies. About 2,845 patents came back. Roughly 7,000 are still lost.

There's a second way to lose everything, by the way: keep it safe in a format nothing can read. The BBC put a digital snapshot of Britain on laserdiscs in 1986, and sixteen years later the discs were nearly unreadable because the players had disappeared. The book that inspired the project, written on parchment in 1086, is still perfectly legible.

Three, two, one

I spent last week doing my own small version of all this. I exported my Medium and my Substack archives, and both turned out to be missing more than I expected. I wrote up what a real Substack backup looks like separately.

The fix: Three copies, two kinds of storage, one of them somewhere else. The memorable version of that rule came from a photographer, not an IT department.

There's even a date for it: World Backup Day, March 31. On purpose one day before April Fools'.

If you're not already, follow me on Medium.

My Medium friends can read this over on Medium.

To find out how last month went, I used to open six browser tabs. Or more.

Two Stripe dashboards, Gumroad, RevenueCat, PayPal, and the Amazon Associates page that, by the way, looks like it was designed in 2011. Probably was.

That's annoying.

So I used Claude Code to build a Mac menu bar app. Now there's a number in my menu bar that shows my earnings in detail. It updates every thirty minutes. I click it and see which platform brought what.

The app I didn't buy

There is probably a paid Mac app for this out there. I bet. I didn't want to buy or subscribe to another tool, though.

So my little menu bar app idea was born. It takes my app subscription payments (for the paid apps I have in the App Store). My earnings from Medium, Substack and Gumroad via Stripe, my PayPal income (mostly from affiliate marketing), and a few other sources.

Then, it checks all, because some overlap. And then I have a clean number of my income.

The tool

About 600 lines of Python, no dependencies, nothing to install. It reads the credentials from a folder where all my keys live, hits five APIs, converts everything to euros using the European Central Bank's daily rate, and prints this:

Umsätze 01.07.2026 bis 31.07.2026

Stripe · Writing  (12 Zahlungen)
  Medium                    …
  Substack                  …
  Gumroad                   …

PayPal  (2 Zahlungen)
  WriteStack (Affiliate)    …

RevenueCat  (26 aktive Abos)
  App subscriptions         …

German, because it's for me.

It costs nothing to run and there's no account anywhere. My revenue numbers don't travel to a company I've never heard of. The API keys never leave the laptop, and I gave every one of them read-only permission, so the worst case is somebody learns what I earn.

Menu bar

I wanted a number in the menu bar, without opening anything.

Writing that as a native Mac app is a whole lot of Swift code. Don't need that right now.

Instead I used SwiftBar. It's a free, open source app that runs any script on a schedule and puts whatever the script prints into the menu bar. The filename sets the interval, so a file called revenue.30m.sh runs every thirty minutes. First line of output is what you see up top. Everything after a line with three dashes becomes the dropdown menu.

💶 5.234,56 €
---
Stripe · Writing: … | font=Menlo
-- Medium: …
-- Substack: …
PayPal: …
RevenueCat: …
Updated 12:05 | color=gray
Refresh | refresh=true

Mine shows a rolling 30 days rather than the calendar month, for a reason. On the first of every month a calendar view says zero, and two days of staring at zero is… depressing.

SwiftBar has hundreds of plugins people wrote for other things. Weather, CPU temperature, crypto prices, unread mail. I only use the one, but the point is you can put anything up there without building an app.

Amazon

One thing is still missing: Amazon Associates. There's no API for affiliate earnings right now. At least, I haven't found one.

The Product Advertising API is being replaced by the Creators API, and I set up credentials before checking what it does. Four operations, all product data. Nothing about commissions. Amazon runs an S3 data feed with the real numbers, but you get in by writing to support and asking nicely, and they don't give it to everyone. I sent the request. No answer so far.

So that one number gets typed into a CSV file once a month, by hand. For now.

Downsides… there are always some

Nobody maintains this but me. If Stripe changes something, my script gets it wrong.

The Stripe CLI login expires every 90 days, so twice a year I have to log in again. I'll forget both times, which is why the script now warns me two weeks ahead. There is a way around this. I'll try that later.

The currency conversion is a compromise. Everything becomes euros at today's rate, including a payment from March. Fine for "how are we doing", not so for a tax return.

The Bottom Line

I've wanted this for months and never built it, because a tool that talks to five APIs was a weekend I didn't have.

But that script with SwiftBar is pretty sweet and works great.

Love the quick check of numbers.

If you're not already, follow me on Medium.

I clicked Medium's export button this morning. 32 MB, 4,030 files, five and a half years of writing.

Then I opened the folder to see what I got.

The writing comes out cleaner than Substack's export does. The rest of it is a different story.

Export

Settings, then Security and apps, then "Download your information". You get an email with a link. Mine took a few minutes.

The zip comes with a README listing 16 folders: blocks, bookmarks, claps, highlights, interests, ips, lists, notes, posts, profile, pubs-following, reposts, sessions, topics-following, twitter, users-following.

One of those is your work. posts holds 28 of the 32 MB.

And 4,030 files is not 4,030 articles. Every comment you ever left is its own HTML file in there too, so the count is mostly conversation. The smallest one in my folder is 2021-08-23_----2caadbdea70.html, 2 KB, and the entire content is :-)

Where Medium does better than Substack: every file is a real page. Title in the head, subtitle, author, publish date, and the canonical link in the footer. Substack hands you body fragments with no title element at all, so you have to join them against a CSV to know what you're looking at. Here you double-click a file and it opens like an article.

Images

1,559 image tags in my posts folder. Zero image files in the whole export.

Same problem as Substack. Every one of them points back at cdn-images-1.medium.com, so the pages look complete on your laptop and stay complete right up until the account is gone.

But Medium adds a twist. The links don't point at your originals. 1,103 of mine point at /max/800/, another 421 at /max/1200/. Three made it to 2560. So even if you download every image by hand, you get the web-sized copy Medium built for the feed. Not the file you uploaded.

Which means the backup happens before you publish, not after. Write the draft somewhere local, keep the images next to it, then paste into Medium. Everything of mine lives in my own app first, with a copy in iCloud.

If you didn't do that, the fix while the account is still alive is a script that walks the export, follows the CDN links and pulls the files down. A few of those exist. I haven't used them.

So: posts ✔ (mostly)

Your list

In my eyes this is the part that decides whether any of the above is worth anything, and the export doesn't even contain it. Subscriber emails are a separate download over on the audience page.

Where Medium tells you the catch, right above the Confirm export button: "You can only export email addresses of users who subscribed before April 2025 and users who you imported. If your list only includes users who subscribed after April 2025, the export will be empty."

So anyone who hit "notify me by email" on your profile after April 2025 is a number in your stats. Nothing else. You can't mail them anywhere but here.

I pulled mine to see what that looks like. A few thousand addresses. I have more than 21,000 followers.

Your words come with you. Your newer readers don't.

Pull the old ones anyway. It only gets more out of date, and those people signed up because they liked something you wrote.

Money

The partner-program folder is interesting. Eleven pages, 994 stories, each with a dollar amount next to it. Lifetime earnings per story, in a file, on my own hard drive.

Substack has no equivalent to that.

It's a humbling read too... Plenty of them earned nothing at all.

The billing side: Medium pays out to your Stripe account instead of sitting on top of it, so there's no application fee to negotiate away when you leave and no support ticket to write. You walk, the payouts stop, done.

What isn't in the zip is the tax paperwork. The 1042-S forms come by email once a year. Download them and keep them with the rest.

What it knows about you

Substack's export is mostly tracking data about readers. 61 of my 73 MB there were email opens with cities, devices and user agents.

Medium's is the mirror image. The tracking is about me.

ips.html is a list of every IP address my account was seen from, with timestamps, going back 30 days. 140 lines. Medium notes in the file that it deletes them after that. Then there are 75 sessions, 7,858 stories I clapped for, 340 highlights, 109 bookmarks, 295 X contacts who are also on Medium, and the seven people I blocked… :-D

Don't drop that folder in a shared Cloud drive and forget about it.

Also missing: view counts, read ratios, follower numbers. Stats live in the dashboard.

Addresses

Every Medium URL ends in a hex ID. ...-47b926474ce7. That ID is in every export filename and the canonical link sits in every footer, so you can rebuild the complete list of your own addresses out of the folder in about a minute.

Worth something only if the part in front of the slug belongs to you.

Custom domains are an option on Medium, as long as you have a membership. You point a domain you own at your profile in the settings, and old medium.com links redirect to it. Mine live at stories.byburk.net, which is why I connect a domain to everything I can.

If I want to move from Medium, I can… but I have to do quite a few things.

The redirect work, mainly. That hex ID means nothing to WordPress or Ghost or anything else, so moving means a rule that catches the old shape and maps it to the new one.

Annoying with your own domain, but at least, I can. It's impossible without a custom domain.

Downsides… there are always some

The domain needs a paid membership. If you're in the Partner Program you have one anyway, but it's a subscription you now depend on for your own URLs.

The export is a button you press by hand. No API, no schedule, nothing that runs while you sleep. Same as Substack.

And no backup fixes the list. You can save every word you wrote and still have no way to tell your readers where you went. That is a huge downside now. And definitely a huge plus for Substack. Until they do the same.

The Bottom Line

Export every few months, pull the old subscriber list now, save your images before you publish rather than after, keep the tax forms, and put your own domain in front of all of it.

Medium hands back more than I expected. Just not the readers. Which really sucks!

Get the Medium + Substack Dual

My Medium friends can read this over on Medium.

Substack has an export button. It gives you a zip file, and clicking it may falsely make you feel safe about a complete backup. It's not complete. Issue 1. And it's not all you should do. Issue 2.

With the export button, you did about a third of the thing.

Your posts are in there. But a lot is not. Let's go through that.

Export

Settings, then Import / Export in the left sidebar, then "Export your data".

You get a zip with your posts as a CSV and as HTML files, your subscriber list, and your stats. Substack's own page lists exactly those three things.

Which is useful. Posts, subscribers, stats. But some things aren't mentioned. Notes aren't there. Neither are comment threads or drafts.

I pulled mine to look at it properly. Three things at the top level: posts.csv, email_list.burk.csv, and a posts folder. That folder holds one HTML file per published post, named with the post id and the slug, plus two small CSVs for every post that went out as an email.

The HTML files are not pages. They're body fragments. No html tag, no head, no title, no date, no author. Out of about 1,100 files, not one carried a title element. Titles, subtitles and dates sit in posts.csv instead, matched up by post id, so turning the zip into something readable means joining the two yourself.

Drafts get another treatment. Mine show up as rows in posts.csv with is_published set to false, and there's no HTML file for a single one of them. You get the title of a piece you never finished. Not the piece.

So the zip is a copy of your text. Good to have. Just don't confuse it with a copy of your publication.

There's no official API for any of this either, so it stays a button you click. I built a small command line tool that talks to Substack's internal endpoints for drafts and Notes, but the export is not one of them.

Images

The export contains no media at all. The HTML files still show your images because they link back to Substack's servers, so everything looks fine on your laptop. Michael Macfadden, who documented his own backup routine, says it right: "The export doesn't contain ANY media assets whatsoever."

Molly White said the same thing after moving her newsletter off the platform. If you use the content export as-is, "you will lose a lot" and your posts will be broken.

The zip looks complete right up until the publication is gone. Then every image in it is a grey rectangle or broken link symbol.

Macfadden's fix is a detour through WordPress: import the Substack export with the importer plugin, check the box that says "Download and import file attachments", and the media lands in the library as real files. Then he flattens it to static HTML.

Another version is to mirror your own public site while it's still online. Something like wget --mirror --page-requisites --convert-links --span-hosts --domains=yourname.substack.com,substackcdn.com https://yourname.substack.com will pull the pages and the images they point at. I've read about people doing this. It's rough, and paywalled posts won't come along, but it gets you the pictures.

Probably the best thing (but involves planning ahead) is already having created all your drafts with images locally in an app or tool and saved on your PC or in the cloud, before putting it up on Substack. So you have local copies of your posts. I hope most people do that anyway. I used to write in Notion, so most of my stuff is there too. Now I have my own app that keeps everything saved locally + a second backup in iCloud.

So: Posts ✔

But there's more, right...

Stats

My export came to 73 MB. The posts are 11 MB of that. The other 61 MB is email tracking.

Every post that was sent as an email gets two extra files. A delivers CSV that logs who it went out to, and an opens CSV that logs the opens. And the opens file is detailed: the subscriber's email address, the country, the city, the region, the device type, the operating system, the full user agent string.

That's a lot of data, by the way. Data protection must be top notch...

So the export isn't only your writing. It's also a record of where your readers were sitting when they read you.

Your data, and useful. But it changes where the zip should live. A folder with thousands of email addresses and city-level location logs doesn't belong in a shared Dropbox or a synced Downloads folder. Encrypt it, or keep it somewhere nobody else reaches.

Your list

The subscriber list is in the zip, but it's the thin version. Seven columns: the email address, whether the subscription is active, the expiry, the plan, whether email is disabled, the signup date, and the date of the first payment.

The fuller one lives on the Subscribers page. The three dots above the list, then "Export". You can take all columns or only the visible ones, and all columns goes down to when someone last opened an email. Substack documents it here.

Do this regularly if you care about "owning" your email list. A list you pulled eight months ago is eight months of new readers short, and those are the ones who signed up because they liked something recent.

Put the CSV somewhere that isn't only your laptop.

Subscribers ✔

Paid subscribers

Free subscribers are a spreadsheet. Paid subscribers are a billing relationship, and that's a different animal entirely.

The good news is that Substack runs on your own Stripe account, so the customers sit in a system you control (sort of). Ghost's migration docs say to "use the same Stripe account that is connected to your Substack" and the subscriptions carry over instead of everyone having to re-enter a card.

Sounds great… if it works. I haven't tried that.

The less good news is in the same docs: "Substack will continue to take 10% fees on your existing paid subscriptions." You leave, they keep earning on the people who were already paying. What???

That reads like an error in the documentation. It isn't.

Substack sits on your Stripe account as a connected platform, and every subscription it set up carries a 10% application fee inside the subscription itself. Cancel the publication, delete every post, walk away for good, and those subscriptions keep sending Substack their cut on each renewal. Molly White, who moved Citation Needed to self-hosted Ghost in 2024, put a whole warning around it: "if you don't do this, subscriptions created through Substack will continue to send a 10% cut to Substack, even after you've left the platform!"

Nobody takes it off for you. You have to ask.

And there's a wrong way to ask. Don't hit disconnect in the Substack dashboard. Buttondown's migration guide mentions it: "This will cancel and refund all paid subscriptions." White says the first support rep who answered her suggested exactly that.

The version that works is an email asking Substack to do it from their end. Buttondown hands you the sentence: "disconnect Stripe from my publication and remove the 10% application fee, without cancelling or refunding any existing subscriptions." White needed a second email and five days before someone did it. Ghost offers to chase it for you through concierge@ghost.org if that's where you're going.

Then, once the fee is off, cut the connection from the Stripe side, under Settings, Team and Security, Installed Apps. In that order. Fee first, disconnect after.

The import itself isn't clean either. White had more than 20,000 subscribers, and every paid one landed in a single tier no matter what they'd been paying, so she wrote a script to sort them back without re-charging anyone. Non-US currencies had to be moved by hand. Ghost had no welcome or renewal emails, so she coded those too.

Several weeks of work. Not an afternoon. Yikes...

Addresses

This, in my eyes, is the deciding factor whether any of the above is worth anything at all.

If you publish at yourname.substack.com, every link you ever shared, every Google result, every mention in someone else's post points at a hostname Substack owns.

You can export the words. You cannot export the addresses, and you can't redirect them either. The archive comes with you, the traffic doesn't. All links break once you're moved.

The export helps a little here. Every HTML file is named with the post id and the post's slug, and that slug is the one in the live URL. Substack shortens long titles, so I have files like the-importance-of-a-platform-independent, and that clipped version is exactly what the address bar shows. Which means you can rebuild the full list of your own addresses out of that folder.

Worth something only if the part in front of them is yours.

A custom domain turns that around. One-time $50 per publication, plus whatever your registrar charges for the domain, then you point a CNAME at Substack. I did that a while back and my publication lives at letters.byburk.net. burk.substack.com still 301s to it, so nothing I'd shared before the switch broke. I connect a domain to everything I can for exactly this reason.

The point of the domain is that leaving becomes a DNS change instead of a fresh start.

You keep your domain, URLs don't break (some must be redirected, but at least that's possible with your own domain), and you keep the traffic and domain authority you've built over the years.

About that possible redirection:

Substack puts posts under /p/slug and most other platforms don't, so you need a redirect rule on the new side to catch the old shape. Ghost documents a regex for it. White needed custom nginx rules on top of that, plus a script to fix internal links between her own posts.

AI can help a lot with that as well.

So in the end, it is still a lot better than watching every link you've ever placed go to a 404.

Downsides… there are always some

The domain costs money. Fifty dollars to Substack once, then the yearly bill at your registrar. Small, but it's not free.

Switching to a custom domain also means Google has to re-learn where you live. Rankings can dip for a while before they come back. Worth it, in my eyes, but it can be a true cost in the first months.

And the export stays manual. There's no scheduled backup, no webhook, nothing that runs on its own. If you forget for a year, you have a year-old backup.

None of this makes leaving easy, by the way. But it makes leaving possible. And that's important. Because we never know...

I do those same steps for Medium too, by the way.

The Bottom Line

Export your posts every few months if you don't have local copies anyway, pull the subscriber CSV more often than that, keep both somewhere that isn't one hard drive, and own the domain address.

That's a true backup of your Substack.

If you're not already, follow me on Medium.

Get PROSTACK

My Medium friends can read this over on Medium.

Mark asked about subdomains in the comments of my last post, the obvious question that follows from it. If you're supposed to connect your domain to Substack and to Medium and to your store, do you need three domains?

No. You need one. The rest are subdomains, and you already have them.

What's a subdomain

A subdomain is the word before the first dot. blog.yourname.com. shop.yourname.com. notes.yourname.com.

They're not new domains. Nobody sells them to you. They come with the domain you bought.

One domain, one bill a year, and depending on the provider a few or up to as many subdomain addresses as you have platforms.

Mine

I run one domain, byburk.net. (I do have more, but for this argument here, we're focusing on this one).

From there, I built subdomains that point to my projects.

stories.byburk.net points at Medium. letters.byburk.net points at Substack. shop.byburk.net points at Gumroad. Three platforms, three companies, three completely separate systems, three subdomains, but one domain name across all of them.

The domain is the brand.

The setup

You go into the DNS settings at whoever sells you the domain. You add a CNAME record. The name is the word you want (shop). The value is the address the platform hands you when you add the domain on their side. Copy it in, save, and go do something else, because it can take a day or two to work everywhere.

The subdomain is also the easy version. Root domains are more annoying sometimes. The DNS spec doesn't really allow a CNAME at the root of a domain, so you end up with A records or your provider's flattening trick, and it can break in ways that are hard to see.

blog.yourname.com is one clean record. Which is why many platforms recommend the subdomain in the first place.

Naming

Boring wins. blog, shop, notes, links, letters, stories, store...

Someone should be able to guess it or type it after hearing it once.

And once a subdomain is out in the world, treat it as permanent. Every link, every share, every Google result points at that exact string. Renaming writing to blog two years later costs you all of it unless you redirect properly. Rather take the extra minute now.

SEO

SEO is messier than the setup.

Google's John Mueller has said subdomains and subdirectories are seen the same, and then added that he'd "try to keep things together as much as possible" and use subdomains where things are slightly different. Which is a bit of a hedge. A lot of SEO people will tell you that in practice, folders (yourname.com/blog) still do better than subdomains, and I think there's something to that. Though I don't really know (or care frankly).

But it doesn't matter much here, because it isn't the choice you have.

Medium says it plainly: you can connect a top-level domain or a subdomain, and you cannot connect a subdirectory. Substack is the same. The folder version would need you to host and route everything yourself, which is the whole thing you avoided by using these platforms.

So the comparison isn't subdomain versus folder. It's subdomain versus yourname.substack.com.

Downsides… there are always some

The platforms charge for the privilege. Substack takes a one-time $50 per publication, then you can add and remove domains for free. Medium wants an active membership. Nothing dramatic, but it adds up if you connect four things at once.

DNS changes might take a day, and during that window your address may not work.

And the authority is a slow build. Every subdomain starts near zero with Google, so stories.byburk.net didn't rock on day one. It took months before it went up. But once it goes up, it's all yours.

Last one: one domain is also one point of failure. Miss a renewal and you don't lose one address. You lose all the subdomains as well, at the same time.

The Bottom Line

You don't need one domain per platform unless you want to. You need one domain, and a habit of adding subdomains every time you sign up for some new platforms that allow connecting a custom (sub)domain.

If you're not already, follow me on Medium.

Get Substack SEO

Hi friends,

a lot piled up since my last letter. Eight posts, and four of them turned out to be about the same thing without me planning it: the address, not the site.

A custom domain on Substack. A custom domain on Medium, which is the whole reason search became my biggest traffic source there. The platforms nobody would file under blogging. And whether a personal website is still worth building at all.

Then two on AI tools, one thought experiment about AI writing everything, and 19 conspiracy theories. That last one fits nowhere. I know. But I love it.

Here's my week.

If AI Wrote Everything From Today On… What Would Happen?

Graphite ran 43,000 articles through a detector and found AI passed humans back in November 2024. So half the experiment is already running. What I keep circling is where new information would come from, if every text came out of the same handful of models.

Read on Medium

Read on Substack

Connect Claude to Apple Notes, Reminders & More

Can Claude read a note and add a reminder on its own? On a Mac, yes, through small local MCP servers that already exist. Unofficial, a bit slow, and it works.

Read on Medium

Read on Substack

Blogging Platforms You Don't Think Of 2026

X, LinkedIn, Telegraph, Notion. Places you can publish 4,000 words on but would never call a blog. Which one fits mostly depends on whether you already have readers sitting there.

Read on Medium

Read on Substack

Connect a Custom Domain to Everything You Can

Substack, Ghost, Beehiiv, Gumroad, even a Notion page. Point your own name at each one and the URLs travel with you when you leave. The catch is that a new domain ranks worse at first, so it might get worse before it gets better.

Read on Medium

Read on Substack

19 Conspiracy Theories, Ranked by How Close They Are to True

Every one of them has one second where it almost works. I took each seriously enough to ask what it would actually take, then looked at where it breaks. Some get closer than I expected. Not close, though.

Read on Medium

Read on Substack

Medium SEO in 2026

Search became my biggest source of traffic on Medium, bigger than the algorithm and bigger than shares. What changed wasn't Medium. It was one setting, plus writing boring titles on purpose.

Read on Substack

Who Still Needs a Personal Website

I keep saying writers don't need one, and I still think that. Developers are the exception. GitHub proves you can code. The site proves you can decide what matters and put it first.

Read on Medium

Read on Substack

Is There a Gemini Equivalent to Claude Cowork?

I called Antigravity "Claude Code + Cowork and faster" once. Not true, and that mix-up is everywhere right now. More than 90% of Cowork sessions have nothing to do with coding, which points at a different Google product entirely.

Read on Medium

Read on Substack

📎 From the Archive

Sits right next to the Claude and Apple Notes piece up top, except this one costs nothing. If you turned on Apple Intelligence, there is already a model on your Mac. Apfel is one brew command and you can talk to it, offline, on a plane, with Wi-Fi off.

Read on Medium

Read on Substack

🔧 One tip

If you have a Substack/Medium account, connect a custom domain to your profile. It's one setting, and it's the reason search is now my biggest traffic source there. Google sees stories.byburk.net instead of one page among a few hundred thousand on medium.com.

Do it sooner than later. A new domain starts out ranking worse than the subdomain you're leaving behind, and the only thing you can't buy back is the time it needs to catch up.

That's #158. Back soon. ✌️

Get the Medium + Substack Dual

My Medium friends can read this over on Medium.

Every text you read from today on was written by AI. A human picked the topic and the angle. The machine wrote the words. Nothing else changed.

What would be different?

I keep circling this one because it is barely hypothetical. Graphite ran 43,000 English articles from CommonCrawl through a detector and found that AI-generated articles passed human-written ones in November 2024. Since then it has been roughly fifty-fifty.

So half the experiment is already running. Let's finish the other half in our heads.

The floor comes up

First effect, almost all texts would be clean. No typos, no paragraph that goes nowhere, no wall of text without a break in it. The structure would be well… structured.

That might matter a lot for the person who has a great idea and no idea how to write it down. A mechanic who knows exactly why your car does that thing, an accountant who understands one specific rule better than anyone, a nurse with twenty years of night shifts. They all get to write it all down with ease.

The edges go first

Shumailov and colleagues published a paper in Nature in 2024 on what happens when models train on their own output. Their term for it is model collapse, and the mechanism is that the tails of the original content distribution disappear. The middle, sort of typical survives. The odd stuff does not.

That is a finding about training data, not about prose. But the shape is the same when every text comes out of the same handful of AI models.

The average goes up. The range goes down.

You might not notice it in any single article. But you would notice it after a month of reading. Everything competent, nothing strange. No writer who is "unconventional" in an interesting way. No sentence that only one person on earth would have built like that.

New information

This is probably the most interesting thought here.

A text is a pipe, kind of. Something goes in at one end that one person knew, and it comes out the other end where other people can use it. That is the idea, right? The prose is the pipe.

If every text is written by the same few models, the models are not adding anything new. Are they? They are recombining what is already in there. The only new material entering the system comes from the prompt.

Which means it comes from a human.

Then the writing is not the work anymore. Having something to say is the job.

And that is a much harder job than writing, some would argue. Because ideas are the currency. You cannot batch those easily. You can ask AI for ideas, but you'll get the same pool as everyone else.

So creating a real idea pool requires having done something, gone somewhere, gotten something wrong and noticed. I already ran into a small version of this when voice-to-text changed how I write. The typing got faster. The thinking did not.

Reading changes too

We talk about this as a writing problem… which it certainly is. But it is a reading problem just as much.

If I knew everything was machine-written, I would stop asking whether a piece is "well" written. That question would be dead after all. Everything is grammatically and structurally well written. AI can do that pretty well. But is it "good"?

I would start asking whether anyone actually had the experience.

Names, prices, dates, places, a photo out of a window, a number someone had to measure. Proof. The value moves from the sentence to whatever the sentence can be checked against.

The machines are already doing this, in a way. Graphite looked at 31,493 keywords in October 2025 and found that 86% of articles ranking in Google Search are written by humans. For articles cited by ChatGPT and Perplexity it was 82%.

Half of everything published is AI, and the systems built on AI still mostly point at people.

Substack started scanning last week. Readers can now run a Pangram check on any post and see an estimate of how much was written by hand, and writers can add a "How I make this" statement instead. Medium has had a rule since May 2024 that fully AI-written stories do not go behind the paywall.

Both platforms are betting that readers want to know.

I don't need to know. Also, detection tools are easily fooled, or entirely incorrect in the first place. But in the end, does it matter? If the story is a good read, it's a good read… no matter who wrote it.

What I would lose

I spent ten years as a computational linguist before I did this, and I have been blogging since 2007. So I trained machines and wrote myself.

The pieces of mine that people write back about are almost never the really polished ones. They are more often the ones where I was wrong or incomplete in my thinking, or where I admitted something a bit embarrassing, or where the sentence came out broken because I am not an English native speaker.

An AI writing in my voice does not get those, usually. It gets the rhythm and the tics and the short paragraphs. It can mimic those. It cannot deliver it like I can or you can, because it did not have the afternoon that caused it.

To be fair

I do use AI to work, daily. Mainly for the automation of the boring tasks around writing and my job. But also increasingly for research, drafting, and correcting.

Writing was never sacred in that sense. We accepted spellcheckers for a long time, editors, ghostwriters who put a CEO's name on someone else's sentences, writing tools that correct and "improve".

There is a real case that a text is a delivery mechanism and I am romanticizing the packaging because I happen to like it. But people with something to say and no writing skill have been locked out of publishing for a long time. A machine that helps with that is worth some collapsed variance.

The Bottom Line

The thought experiment does not end where we might have expected. I thought I would land on "the writing gets worse." It does not necessarily, depending on how we measure that. What is now called "AI slop" because of some overused expressions or signs, was totally normal a few years ago. Nobody would have noticed.

Now we do.

If AI Wrote Everything From Today On, the writing would get more competent each day… but also maybe less interesting.

Right now, the ability to write a clean paragraph is still worth something. In an AI-only world it is worth nothing, because everyone has it.

What is left is what you put into the prompt. Which is just what you have seen, done, screwed up, and paid attention to.

So the question is not whether AI can write like you, but rather what would you have to say?

If you're not already, follow me on Medium.

Get Superwriter

My Medium friends can read this over on Medium.

Most of my life sits in Apple Notes, Mail, Calendar and Reminders. Ideas, drafts, the grocery list, the thing I have to do on Friday.

I use Claude Code every day for writing and building. So at some point the obvious question was: Can Claude touch that stuff directly? Not copy-paste. Actually read a note, add a reminder, on its own, as it can via MCP with other productivity tools like Google's tools, Notion, ClickUp, and others?

So I went looking.

Short answer...

Yes!

On a Mac, through small local MCP servers. Several of them exist already (unofficial of course), free and open source. Nobody needs to build this from scratch.

An MCP server is a little translator that runs on your machine. Claude talks to it, it talks to the app, it hands the result back. That is the whole idea.

What you can do

The most complete one, Dhravya's apple-mcp, covers a lot:

• Notes: search and create

• Reminders: list and create

• Calendar: search, list, and create events

• Mail: send and search

• Messages: send and read

• Contacts and Maps: search

Most of it is read and create. Editing and deleting is thinner. Other, more focused servers go deeper on one app. One does full create, edit, and complete on Reminders. Another does proper updates on Notes.

And you can chain them. Read a note, pull out the names, draft a follow-up email. That part works.

How it works underneath

The server shells out to osascript, Apple's built-in scripting tool, and sends a tiny script like "tell Notes to make a new note." That reaches the app through Apple Events, the same plumbing that has been in macOS for decades.

It works. It's not really fast… but it works.

The sturdier servers skip scripting and use EventKit, Apple's real framework for Reminders and Calendar. More robust, but it needs a compiled binary, which is friction if you do not write code or know how to make AI do it.

Either way, the first time Claude reaches for an app, macOS throws a permission prompt. You click allow. That is the security gate.

Downsides

This all runs local. On your Mac, against your Mac. No phone, no cloud Claude, not without extra tunneling you do not want to set up.

It is macOS only.

And the scripting route is a little brittle. It breaks when Apple changes things, and permissions get reset on big updates. Something that worked in June can want a re-grant in October.

Another way

Apple Shortcuts.

The most robust way to give Claude hands on your Mac is not raw scripting. It is to build the action as a Shortcut, out of Apple's own maintained blocks, and let Claude run that Shortcut by name.

Apple keeps those blocks alive. A "create reminder" Shortcut is far less likely to rot than a hand-written script poking at the app. A small server does one job, run a Shortcut, and Apple's stable command does the rest.

Two bonuses. Shortcuts sync to your iPhone, so the same action works on the go (in theory). And a Shortcut can run a shell script, so it can reach past Apple's apps into your own tools.

The trade-off. Shortcuts are great at doing things, weaker at handing structured data back. For reading a note and getting clean text out, a little code still wins.

So the best setup is a hybrid. Shortcuts for the actions, a thin read layer for pulling data.

But why?

Why do it? For me, it's fun, but I also want to automate a few things with AI.

On its own, "Claude can add a reminder" is a demo. It gets cooler when it plugs into something you already do. For me that is writing.

Two ideas:

1. An ideas note as the front door. I drop article ideas into one note on my phone. Claude reads it and turns the good ones into drafts. I already do a rougher version of this by hand. The note would just be the capture layer.

2. Reminders as the human step. My publishing has manual gates. Publish here, grab the link, cross-post there. Claude could drop those into Reminders as it finishes the automated parts, so the leftover human steps sit on my phone.

The Bottom Line

Can Claude run your Notes and Reminders? Yes, today, on a Mac, with a free server and a permission click.

Should you wire your whole life to it? I have not. It is local, a little fragile.

But the other version is quite interesting. Shortcuts for the doing, a bit of code for the reading, and only where it feeds something you already run. That part I want.

If you're not already, follow me on Medium.

Get Superwriter

My Medium friends can read this over on Medium.

Say blogging and most people I know picture the same three names. WordPress. Medium. Substack. I compared those in a separate piece.

This is about the other places. The ones you can publish long-form writing on, but would rarely call a blog.

Twitter / X

Twitter lets Premium users post very long now, into the tens of thousands of characters. The exact cap has moved around over the years, but it sits somewhere around 25,000 [1]. Call it 4,000 words. A full article inside a tweet.

But the timeline still shows about 280 characters and folds the rest behind "Show more." The opening line does the reaching. Everything after it is read by whoever decided that line was worth a tap.

Who is it for: someone who already has a following on X. The long post rides on the audience you built and the hook up top. So it pays off when people already show up for your name. Starting from zero, 4,000 words land in front of almost no one. Then again, that is true almost everywhere, and X at least has some discovery built in through replies and reposts.

LinkedIn

LinkedIn is kind of counterintuitive. A normal post with an outside link gets throttled now, creators report around half the reach or worse [2]. LinkedIn has never confirmed a number, but writers feel it. Articles and newsletters don't get hit. LinkedIn is betting on those as its own growth engines.

A newsletter edition goes straight to your subscribers by email and notification. And LinkedIn has started surfacing long articles again after years of leaving them alone. The long-form side of LinkedIn runs on different rules than the feed everyone pictures.

Who is it for: professionals. If you write about your industry and your readers are clients, colleagues, or recruiters, they are already scrolling here. Consultants, freelancers, B2B founders, people building a name in a field. For a personal essay or fiction, wrong crowd (for now).

Telegraph

Then there is Telegraph, at telegra.ph. Telegram built it. You open the page, type a title, write, and publish. No account, no signup, nothing to manage.

It has been around for years and still runs, though Telegram has tightened it over time, like cutting off new image uploads after the anonymity got abused. For plain text you want live in under a minute, nothing really beats it.

Who is it for: the one-off. A piece you just need a link for, a long text dropped into a Telegram group, something you want online without attaching your name or building a home for it. Nobody subscribes to a Telegraph. It is a publish-and-move-on tool, not a base you grow.

Notion

Notion publishes too. Any page turns into a public website, and since 2024 you can put your own domain on it [3], with SEO basics and author pages. (Proper RSS still mostly needs a third-party tool.) People run real blogs on it now.

If you already draft in Notion, the page you wrote in becomes the page people read. No export, no second tool.

Who is it for: anyone already living in Notion. If your notes, drafts, and docs sit there, publishing is one toggle away, and a low-maintenance personal site or a public knowledge base falls out of it for free. What Notion will not hand you is an audience. There is no discovery, so you bring your own readers.

Dev.to and Hashnode

If any of what you write is technical, Dev.to and Hashnode come with a developer readership already sitting there. Free, decent SEO, and the audience is the whole point. Most non-coders have never heard of either. Which can be a good thing.

Who is it for: this makes most sense if you write code, or about it. A tutorial or a dev log finds readers fast here. A piece on anything else finds none.

Beehiiv

Beehiiv is a newsletter platform from ex-Morning Brew people. No cut of your subscription revenue, a flat monthly fee instead, with an ad network and a referral system built in [4]. Substack, for contrast, takes about 10 percent [5].

It sits next to Substack in most comparisons, so it is less hidden than the rest of this list. But outside the newsletter world plenty of writers have never looked at it.

Who is it for: running a newsletter like a business and combining it with a blog. If you are chasing growth and money past subscriptions, the ad network, the referrals, and the missing revenue cut are good reasons to try it. If you just want to write something and hit send, it is more machine than you need, though.

The Bottom Line

The platforms with an audience built in reward writing made for them. A thread written as a thread, a LinkedIn article written for LinkedIn. If we drop an unchanged blog post in, it tends to land flat. Makes sense, right? Discovery on these is not dead, though, it just doesn't fire as well for long-form as it could.

None of these is your own site either, which is a separate argument.

But when people picture blogging, they rarely think of those above. However, writing fits in a lot more places than we think.

If you're not already, follow me on Medium.

Get the Medium + Substack Dual

Sources

[1] X/Twitter Premium character limit, 2026 — TypeCount

[2] LinkedIn external-link reach penalty — Gromming

[3] Notion Sites, custom domains and web publishing — Notion Help Center

[4] Beehiiv revenue model and Morning Brew origin — eMarketer

[5] Substack's 10% fee — SchoolMaker

My Medium friends can read this over on Medium.

Own the domain, not the website. I've made that point a few times now. It holds up, at least in my eyes. People will disagree. That's fine.

Either way, if the custom domain is where you're headed, then here's the move that follows from it. Connect a custom domain to everything you can. Because it can make a huge difference, even though it might actually be worse in the beginning.

Hear me out!

Everything you can

Most platforms let you connect a custom domain nowadays.

Substack lets you. Medium lets you, although this has changed back and forth over the years, so I am not actually sure if you can right now. Others do: Ghost, Beehiiv, Gumroad, even a Notion page. Point your own name at each one. It's crucial, and many people don't do it for two reasons:

1. Domains cost a bit

2. They don't even know you can

Once you do connect a custom domain, though, the address is "yours". In the sense that you have control over where that domain points.

yourname.com for the writing. Not yourname.substack.com. Not medium.com/@you.

The platform still does the hosting, the design, the sending. You just put your name on the door. And that is a big deal because...

It's Portable

This is the whole point. You can leave any platform and go to the next. And your URLs will follow you.

Say Substack changes something in 2027 you don't like. You move to Ghost. You switch the DNS. Your readers type the same name and land in the same place. The links you spent years building still work. The Google traffic? Stays.

Try that without a domain.

The day you leave, yourname.substack.com is a dead end. Every share, every backlink, every old link sitting in someone's bookmarks, gone to a 404.

That's the "rented land" argument. It's not really true I think, because in the end, everything is rented. A domain too. But the probability of losing it or being banned is lower for a domain than for Medium, for example. People get banned from Medium. They hardly get banned from a domain provider. It's still possible, of course. But rather unlikely.

So in our view now, the domain is the address that stays "yours" when you move out.

SEO

A domain you've held for years collects things.

• Backlinks.

• Mentions.

• Index age.

• Google starts to trust it.

yourname.com pointed at Substack beats yourname.substack.com, even with the exact same words behind it. The authority is yours, not the platform's. And when you move, it comes along. (That's most of what I do in Substack SEO, by the way.)

BUT also… in the beginning a new custom domain is obviously much lower ranked than a Substack subdomain. Which means in the beginning, your traffic might actually be worse. That's a risk you must take. Over time though, your custom domain will be the biggest asset.

About that "ownership"

You don't own the domain either. I mentioned it a second ago.

You rent it. Every year, from a registrar, who answers to a registry, who answers to ICANN. Miss one renewal and it might be gone (or even worse, snatched up by someone else who then enjoys all the work you put in).

So when I say "own your domain," I'm cheating a little. You own it about as much as you own your Substack. You rent it from someone further up the chain. The difference is that this particular rental moves with you.

That's the point. Not ownership. Portability.

Higher up

And it goes higher than the registrar.

.com runs through Verisign, an American company, under US jurisdiction. Domains get seized by court order. It has happened plenty of times. Sanctions can cut a whole country out of the naming system.

So if the US wanted to lock a place out tomorrow, and with the current administration that's not the wildest thing to picture, your "owned" .com is one order away from pointing at nothing. Same as the platform. Same as everything else.

At that point nobody owned a thing. We were all renting from the same chain the whole time.

I mean I wouldn't trust Trump with anything. So in that sense, everything I do online is pretty much "rented" and definitely not owned.

The Bottom Line

Connect the custom domain. To all of it. To everything you can.

It's the most movable piece you've got, and switching platforms without losing your address is worth a lot.

Just drop the word own. You're renting a name that lets you change which other thing you're renting. That's a good deal. The best we have right now.

Until the naming system itself goes dark one day. Then it turns out none of us owned anything to begin with.

My Medium friends can read this over on Medium.

Every conspiracy theory has one second where it almost works. That little moment of… what if...

Let's live in that second for a moment.

The rule is simple. Take each theory seriously enough to ask what it would actually take for it to be true. Then look at what breaks. At the end, a ranking of how close each one gets to real.

Not very. But some get closer than others.

The Earth is flat

The oldest one here, and somehow one of the newest too. Serious people stopped believing it more than two thousand years ago. A British writer named Samuel Rowbotham revived it in the 1800s, and then YouTube revived it again around 2015, which is where most of today's believers come from.

If true: Every image from NASA, the ESA, China's CNSA and Russia's Roscosmos is faked, or part of one agreed lie. GPS built on a curved Earth still works perfectly, and nobody notices the geometry underneath is wrong. Pilots fly globe routes every day and never hit anything that doesn't add up. And millions of people, engineers, astronauts, meteorologists, backyard astronomers with their own telescopes, are either in on it or handing out proof of the lie without knowing.

Why it breaks: That headcount is the whole problem. Rival states that agree on almost nothing, the US, China, Russia, India, the ESA countries, would have to cooperate flawlessly for decades. No leak, no whistleblower, no stray measurement out of Beijing or Moscow, even though the payoff for exposing a Western hoax would be enormous. Meanwhile the curve of the Earth was measured with simple tools back in antiquity, by Eratosthenes around 240 BC, and reproduced by amateurs ever since. The bigger the circle of insiders, the smaller the odds it stays sealed.

The moon landing was faked

It started in 1976, with a self-published book by Bill Kaysing, a former writer at one of the rocket companies, called "We Never Went to the Moon". It landed in the right decade for it. Post-Vietnam, post-Watergate, when trust in the US government was near the floor.

If true: NASA staged the 1969 footage, the radio traffic, the telemetry and the rock samples, with hundreds of thousands of Apollo workers and not one of them ever talking. The shots would have to be simulated on a set, without a single independent expert in optics, physics or film ever confirming the supposed tells, the missing stars, the "waving" flag, the odd shadows.

Why it breaks: There is one witness who would have loved to embarrass the US: the Soviet Union. Mid space race, with its own rocket experts, its own radar tracking of the mission, and every propaganda reason to call it a fraud. And still not one word from Soviet leadership doubting the landing. On top of that, moon rock that has survived every geochemical test for over fifty years, and landing sites since photographed by several countries, India and Japan included, gear still sitting there.

Chemtrails control weather or people

This one is relatively young. It grew in the late 1990s, after a 1996 US Air Force paper on weather modification got read as a confession, and spread from there through internet forums and late-night radio.

If true: Tens of thousands of pilots worldwide, from countries that agree on almost nothing, keep one shared secret for decades. Airlines and plane makers fit the spray gear into passenger jets unnoticed, and the aviation authorities that inspect and certify every aircraft somehow never spot it or are told to "not see it". Plus a volume of chemicals big enough to move weather or health across whole continents, without a single cargo manifest, fuel log or maintenance record giving it away.

Why it breaks: Contrails were explained a long time ago. Water vapor from engines freezes into ice crystals at altitude, and how long they hang around depends on the humidity of that air layer. That is why a trail lasts minutes one day and hours the next, with zero change to the plane. Environmental agencies in several countries have measured ground air for years and never found raised levels of the claimed stuff, aluminum or barium. A theory that needs rival states and corporations to cooperate dies right at the cooperation.

Elections are centrally rigged

As old as elections themselves, and it flares up every time a big one is close or bitterly lost. Now throughout history, there have always been rigged elections. That part's not wrong. But the modern everything-is-rigged version is something else and it travels fastest online, usually right after a result someone didn't want.

If true: Steering entire national results across several cycles would mean fooling or enlisting independent observers, rival parties with opposite interests, thousands of local poll workers in decentralized counting systems, and often courts too, across very different political systems, consistently enough to stay hidden for years.

Why it breaks: Small local vote fraud is there, documented, and sometimes prosecuted. The big version is a different claim entirely. Recounts and forensic audits run after the most disputed elections in several countries have kept confirming the original results, not the fraud claims. A system decentralized enough to produce local errors is a bad fit for central, invisible control. You can't have both.

Vaccines contain tracking microchips

Basically a child of 2020. It spread during the COVID vaccine rollout, often bolted onto older Bill Gates and 5G stories that don't fit together with it at all. That rarely matters to a theory in motion.

If true: An injectable chip fits through a standard needle, inner diameter usually 0.3 to 0.5 millimeters, runs with no power source, carries an antenna for data, and stays invisible in every one of billions of vials through every quality check at independent labs worldwide, across makers, countries and batches.

Why it breaks: This is a physics question. Even passive RFID chips, the kind that need no battery, require an antenna too big to pass through an injection needle, and once in tissue it would show up on ultrasound or an X-ray. If tech that small existed, it would surface first in logistics or the military, where people already pay for it openly. Not hidden in a vaccine dose, where it barely adds anything.

A secret world government runs everything

The granddaddy...

Its bones go back to 18th-century Illuminati panic, and the modern "New World Order" version took off in the 1990s, once the Cold War ended and people went looking for a new single villain. Now, there are powerful people that run a lot of stuff. No argument there. But the world government…

If true: One central group choreographs wars, financial crises and pandemics with a precision that keeps political, economic and media systems in sync for decades, across rival governments, central banks and corporations that otherwise fight in public.

Why it breaks: That elites and connected institutions hold outsized sway is boring and well documented. Lobbying, tangled interests, informal networks, all real. The stronger claim, one group steering it all, explains everything after the fact and predicts nothing you can check. That is the classic mark of a theory that dodges any test. It would also have to explain why the alleged puppet masters keep eating huge losses themselves.

Area 51 hides aliens

I always love a good alien story.

And Area 51 is the most famous, of course. It runs on two real things stitched together: the 1947 Roswell crash story, and a genuinely secret base in Nevada. It got its alien-tech shape in 1989, when a man named Bob Lazar went on TV claiming he had reverse-engineered a saucer at a site near Groom Lake.

If true: The US military has held alien tech far beyond our own for decades and hoards it in a sealed desert base in Nevada, without a single visible jump in military or civilian technology coming out of it.

Why it breaks: Area 51 is real, it is secret, and experimental military tech was tested there for decades, the U-2 and SR-71 spy planes, early stealth. That fully explains the secrecy and the strange lights and craft seen over the base, with no aliens needed. Anyone still holding on to aliens has to also explain why the military got no visible tech leap out of supposedly superior alien hardware, none that "normal" research can't already account for.

Big Pharma is hiding a cancer cure

An old, sticky one. And man, this is one that would probably be the worst if true.

The "they're hiding the cure" idea has floated around cancer treatment for decades, latching onto whatever unproven remedy happens to be current at the time.

If true: Oncologists, nurses and research bodies worldwide, including in countries with state health systems and no profit motive, and individual scientists who lose their own family to cancer, all know about a working cure and stay quiet forever, without one whistleblower from inside science ever going public.

Why it breaks: Patents run out anyway, usually 20 years from filing, which guts the supposed reason to bury something. A company could patent a cure and sell it, rather than hide it. And nonprofit and public research, at universities and state-funded institutes, keeps searching independently of industry, with every scientific and career reason to expose a suppressed cure the moment they found one.

Paul McCartney died in 1966 and was replaced

October 1969. A Detroit radio DJ, Russ Gibb, took a caller who told him to play Beatles tracks backward for "clues." A college newspaper ran with it, and the rumor that Paul had died in a 1966 crash went global in weeks.

If true: The three other Beatles, their management, their families and dozens of session musicians all hid a swap for almost sixty years, through public shows, endless interviews, fingerprints on paperwork, and, since the 2000s, public DNA from his own family. The stand-in would also have to look and age like him for decades, and match his songwriting and his singing voice.

Why it breaks: The supposed proof, the barefoot Abbey Road cover, lyrics said to be hidden in reversed audio, a "telling" number plate on a record sleeve, almost always turns out… let's call it mundane... on a closer look. The plate belonged to a car parked in the street that happened to be in shot, not a planted clue. The more you check each "proof" on its own, the more it dissolves into coincidence and hindsight, instead of adding up to anything.

The internet runs mind control

A newer theory on this list, and the only one that needed the internet to exist at all. It grew through the 2010s as social feeds took over how people get their news, and ordinary fear of the algorithm hardened into something huge.

If true: A few big platform owners use their algorithms not just to farm attention (which is true) but to grow one uniform set of political or social beliefs in billions of users, coordinated enough to land the same way across borders and languages. I mean… theoretically this is possible.

Why it breaks: That algorithms shape behavior is well studied and not in doubt. Platforms actively test what keeps you scrolling and tune the feed for it. Coordinated, targeted opinion steering is a different claim. It would have to explain why rival platforms with opposite business models and owners pull in the same direction, and why single countries, and single platforms, still hold sharply opposed political camps that successful central steering should have erased. The research shows filter bubbles hardening opinions people already had, not redirecting them. With more research and AI in the picture, this whole story will get more interesting in the coming years, I am sure.

Hitler survived and fled to Argentina

Ah, yes. The Nazi stories. As a German, I know those too well.

The Hitler survived story started almost immediately in 1945 after the capitulation, and Stalin himself fed it by publicly suggesting Hitler might still be alive. Real Nazi escapes to Argentina in the following years gave it a believable stage. So, yes, this would have been the place for Hitler. Probably the only one on earth.

If true: Soviet intelligence, first into the bunker in 1945, faked dental IDs and skull fragments as evidence for decades, even though the USSR had every reason to milk a real find for propaganda rather than fake one. An escape network across the Atlantic slipped past British and American intelligence saturating Europe in 1945, and Argentina hid one of the most wanted people alive for decades, without one verified photo, fingerprint or DNA trace surfacing. I mean a few very famous Nazis did exactly that.

Why it breaks: Nazi war criminals like Adolf Eichmann and Josef Mengele did flee to Argentina, which is what makes the country a plausible backdrop at all. For Hitler himself the evidence points the other way. In 2018 a team led by French forensic scientist Philippe Charlier matched tooth fragments and a skull piece with a bullet hole, held in the Russian state archive, to Hitler's known dental X-rays from 1944. Charlier said afterward you could "end all conspiracy theories about Hitler." Worth being fair about a wrinkle: a 2009 DNA test on a separate skull fragment from the Moscow archive found it came from a woman. Conspiracy fans have run with that ever since, but historians put it down to different finds getting mixed up in the chaos after the war, while the 2018 teeth and their skull piece were kept separate and clearly matched. The escape stories themselves trace largely to a deliberate Soviet disinformation campaign after the war, and the FBI and CIA chased the tips without ever confirming them.

The Titanic was sunk on purpose

I had never heard this one before writing this.

Fairly modern, as these go. The idea that the damaged sister ship Olympic was sunk in the Titanic's place got its popular form from Robin Gardiner's 1998 book, Titanic: The Ship That Never Sank?

If true: White Star Line quietly swapped the damaged Olympic, hurt in an earlier collision, for the new Titanic, to sink the battered sister ship for insurance in the North Atlantic. That means over 1,500 dead out of more than 2,200 aboard, including upper-class passengers like John Jacob Astor IV, whose families would surely have gone public with any hint of a con.

Why it breaks: Photos of both ships before and after the supposed swap show clear, unchanged build details, like the number and layout of portholes on the forward A-deck suites, which differed between Olympic and Titanic. And an insurance fraud that accepts thousands of real deaths of prominent, powerful passengers makes little economic sense, when a plain insurance job had far less risky, quieter routes.

Elvis is alive

It began the moment he died in 1977, and never really stopped. Decades of "sightings" and tabloid covers kept it going long after the man was gone.

If true: One of the most famous faces on Earth went fully invisible after his official death in 1977, and in almost fifty years not one verified photo, DNA sample or checkable proof turned up, in a market where tabloids would pay seven figures for the reveal.

Why it breaks: The death certificate, the Memphis autopsy and named eyewitnesses, his doctor, family, the funeral home, are all documented. The money for a real proof would be huge, and the fact that none ever arrived in checkable form, despite decades of sighting reports, points straight against the theory.

JFK was killed by a CIA or Mafia plot

It started the day he was shot in 1963, grew through the Warren Commission and prosecutor Jim Garrison's case, and got a huge second wind from Oliver Stone's 1991 film JFK.

If true: Several intelligence or Mafia players held total silence for decades, in circles better known for turning on each other, cutting immunity deals, and leaking, than for perfect discipline.

Why it breaks: This is the most thoroughly investigated theory of its kind. The Warren Commission concluded in 1964 that Lee Harvey Oswald acted alone. The House Select Committee on Assassinations landed in 1979 on the softer wording of a "probable conspiracy," without ever naming or proving one. In 2025 the US government ordered the release of nearly all remaining JFK files, and after reading them, historians on the case agreed that new details surfaced, but nothing that shakes the core finding of Oswald acting alone. The longer and harder a claim gets searched with no hard proof of a plot, the more its absence weighs. That doesn't make it true. It makes it the least cleanly closed theory on this list.

9/11 was an inside job

It appeared within months of the 2001 attacks and went mainstream with the 2005 film Loose Change, which spread across the early internet faster than almost anything before it.

If true: A controlled demolition in two skyscrapers plus the neighboring WTC 7, rigged unseen inside busy office towers with thousands of daily users, carried out with thousands of engineers, investigators, cleanup crews and politicians across party lines. All for an event that cost the supposedly guilty government dearly: two wars, thousands of its own soldiers dead, trillions in knock-on costs.

Why it breaks: The official US investigation by NIST concluded that fire and impact weakened the towers' steel enough to explain the collapse spreading down from the impact zones, and found no evidence of explosive residue or controlled demolition. WTC 7 was put down to structural failure from office fires. NIST also worked out that a demolition using the often-cited thermite would need thousands of kilos placed unseen on hundreds of columns, which is hard to do without notice. A group of engineers and architects, Architects & Engineers for 9/11 Truth, still disputes this and wants a new inquiry. That shows the technical debate hasn't gone fully silent, but it doesn't change that the overwhelming majority of the field, and several independent bodies, land where NIST did.

Antarctica hides an ancient civilization or a Nazi base

Nazis again. Some of this is true. Germany ran an Antarctic expedition in 1938 and 1939 and named a slice of it New Swabia. After the war that grew into stories of a hidden Nazi base, and folded into older ideas of a lost civilization under the ice.

If true: A working base under the harshest climate on Earth got supplied unnoticed for decades by the more than twenty countries that have run research stations since the 1959 Antarctic Treaty, countries that carry out joint inspections and get to watch each other do it.

Why it breaks: Precisely because the continent was strategically sensitive in the Cold War, it became one of the most closely watched places on the planet. Satellite coverage, ice cores that track climate back hundreds of thousands of years, and international inspection rights leave little room for a base that stays hidden for decades. Picking that spot to hide a secret base runs straight into how densely it is observed.

The Bermuda Triangle is supernatural

This theory even has a coinage date. Writer Vincent Gaddis named the "Bermuda Triangle" in a 1964 magazine article, and Charles Berlitz's 1974 bestseller turned it into a global mystery, Atlantis and all.

If true: Inside one stretch of sea between Florida, Bermuda and Puerto Rico, ships and planes vanish far more often, and far more strangely, than in equally busy waters, through forces no instrument can measure.

Why it breaks: Lloyd's of London, one of the oldest and strictest ship insurers there is, charges no higher premium for this area than for other routes, and the US Coast Guard has said more than once that the incident rate there is not statistically higher than elsewhere in the Atlantic. Author Larry Kusche showed back in 1975, digging into individual "Triangle cases," that many of the underlying reports were simply wrong, bad positions, inflated details, or ships that vanished outside the area. The zone is big, storm-prone and heavily traveled, enough to explain rare losses through chance, weather and human error, no supernatural stuff needed.

Bigfoot or the Yeti exist and are covered up

The name is from October 1958, when a California logger found huge footprints and a local paper called their maker "Bigfoot." The prints were later revealed to be a hoax, carved wooden feet, though the legend easily outran that reveal. The Yeti is older, out of Himalayan folklore.

If true: A breeding, surviving population of large two-legged primates has lived undetected in North American and/or Asian forests for decades, while wildlife agencies, forestry offices and scientists keep the secret or simply miss it.

Why it breaks: A population like that needs enough food, territory and regular breeding, which means physical traces, droppings, carcasses, clear DNA, skeletal finds. Despite trail cameras and smartphones everywhere now, not one of those has ever been confirmed. The most famous supposed proof, the 1967 Patterson-Gimlin film, is read by most experts on the case as a person in a fur suit, even if a few believers still argue otherwise.

The Earth is hollow

It has an oddly respectable start. The astronomer Edmond Halley floated a nested-shells Earth in 1692 to explain magnetic quirks, and in 1818 an American named John Cleves Symmes Jr. pushed the version with openings at the poles, even lobbying Congress for an expedition to find them.

If true: Seismic waves would travel through earthquakes in a fundamentally different way than the thousands of stations worldwide record, and the total mass and density worked out from gravity measurements would be flat wrong, for a supposedly hollow structure that still produces the pull we feel at the surface.

Why it breaks: Those seismic and gravity readings come from dozens of countries with no shared reason to cover anything up, and for over a century they have drawn the same picture, a layered inside, mostly solid to liquid, with a solid inner core and a liquid outer one. This is not really a fight over photos or witnesses anymore. It is over hard physical numbers that reproduce daily, which makes it one of the most clearly disproven theories on the list.

The RANKING

No ranking like this is exact. But I love me some rankings… so, let's do it anyway.

With conspiracy theories you can't cleanly compute odds, only compare them. From least unlikely to basically ruled out:

1. JFK conspiracy. Even after the broad 2025 file release, it's the only one with real historical loose ends, though still no new proof of a coordinated plot.

2. Central election rigging. Local fraud is real, the nationwide version most likely isn't.

3. Internet mind control. The behavior nudging is real, the targeted coordination isn't (yet).

4. Secret world government. Outsized elite influence is true, central steering isn't (yet).

5. Big Pharma hiding a cure. At least a motive you can follow, even if a weak one economically.

6. Area 51 and aliens. The secrecy is real, just military, not alien.

7. Hitler's escape to Argentina. Other Nazis fled there, he didn't, per the 2018 dental match. Also per chance. Would have been hard for the most famous face on earth to flee to a country on the other side of the world in slow trains and planes that could barely cross an ocean.

8. Chemtrails. Explained by physics, but at least you can point at it and argue.

9. 9/11 inside job. Refuted in detail by NIST, but with a dissenting engineers' group still active.

10. Antarctic civilization or Nazi base. Beaten by how densely the place is watched.

11. Bigfoot. No hard biological find after decades of looking, but at least biologically thinkable.

12. Paul McCartney died in 1966. Specific "proofs" exist, and dissolve one by one.

13. Elvis is alive. Certificate, autopsy, witnesses, pretty obvious.

14. Titanic ship swap. Disproven in photos by the build details.

15. Faked moon landing. Never denied by the USSR itself, which had every reason to.

16. Bermuda Triangle. Directly refuted by Lloyd's numbers and the Coast Guard.

17. Microchips in the vaccine. Simply not technically doable.

18. Flat Earth. Independently disproven since antiquity.

19. Hollow Earth. Runs against hard seismic data, so it sits at the bottom.

The thread through almost all of them is the same. They need an unrealistically large, unrealistically disciplined group of insiders who never once leak, often people who would take each other out if they could. The real conspiracies, Watergate, the VW diesel scandal, were almost always smaller, shorter-lived, and far worse kept than these nineteen would need to be. Doesn't make all of them entirely implausible… which is the reason conspiracy exists, right?

In 2024 I wrote a piece called Why I Don't Care Much About SEO on Medium. I meant it at the time.

Then SEO became my single biggest source of traffic on Substack and Medium over the years. Bigger than the algorithm. Bigger than shares.

What changed wasn't Medium. It was one setting.

Custom domain

By default, your Medium articles live at medium.com/@yourname/some-title. A few hundred thousand other writers share that domain. Google sees a Medium article. Not your article.

Connect a custom domain and that changes a lot. Mine live at stories.byburk.net. Same Medium, same editor, same everything. Different address. And Google treats that address as mine.

Numbers

I don't have to guess whether it works. I can see it in Google Search Console.

My Medium pulls millions of search impressions. A few thousand clicks a month, every month, from people typing something into Google and landing on my stuff.

Medium isn’t sending me readers. It does too, of course. But this is entirely Google. It compounds while I sleep, and it doesn't stop when a post drops out of the feed. That is the part the feed can't give you.

Titles

I hate title SEO. Ends up in boring titles. Sometimes it’s okay and I cave in. Often I just ignore it.

A clever title wins claps on Medium. A searchable title wins Google. They're rarely the same title. Pick and choose.

If you want Google, "My Favorite Terminal, Revisited" gets nothing. "Best Terminal for Mac" gets typed into Google a few thousand times a month.

I do it more often now. I resisted this for years.

Evergreen over news

News dies in a day. Usually.

Search rewards the opposite. "How to X", "best Y", "is Z worth it" get searched every month, for years. Usually. My oldest pieces still bring in Google readers I wrote for once and forgot about.

Evergreen is good.

Downsides

The usual: Medium's algorithm, Medium's rules, Medium's paywall deciding who gets to read past the third paragraph. Google sends the reader over, Medium's wall can still turn them away.

AI Overviews are eating search clicks too. Google answers the question itself now, and the reader never clicks through. Less traffic than a year ago, for everyone. Not just Medium.

Which is why I do the same thing on Substack, where I own the email list. Same idea, my list, no paywall.

The Bottom Line

SEO on Medium isn't exciting. It's a domain setting, boring titles, and a lot of waiting.

But I get readers every single day from posts I forgot I wrote. Two years ago I said I didn't care about this.

I care now.

Get Substack SEO

My Medium friends can read this over on Medium.

I've written the same article a few times now. Personal websites are mostly over. Platforms handle the hosting, the design, the payments, the distribution. For a writer, your own site isn't necessary anymore. I still like having a personal website and it does have some unique aspects, but for most creators, it is not needed anymore.

There's an exception, though.

Developers.

This is the article about that exception.

Proof, not a homepage

For a writer & creator, a personal site is a homepage. But hardly anybody browses homepages anymore. Same goes for personal blogs. Mostly. With exceptions of course.

People rather scroll feeds, click recommendations, read in apps. So the writer builds a beautiful page, points to it, nobody clicks, and goes back to publishing on Substack. I do that too.

A developer's site is not a homepage. It's a work sample.

For a developer, the website isn't where you point people. Not just that. It's the thing you're being judged on, often.

A writer shows their writing somewhere else. A developer's "somewhere else" is the site itself, running, in a browser, doing what they said they could build.

The maintenance, the deploy, the little bugs you fixed at midnight. For a writer that's overhead. For a developer that's the point.

GitHub

"But I have a GitHub."

Sure. GitHub is the baseline now. Recruiters check it. Clients check it. If you write code and you're not there, that's the first problem to fix.

But a GitHub profile is a list of repos. Green squares, READMEs, commit history. It shows you show up. It doesn't show taste. Usually.

A personal site is where you frame the work. The one project you're proud of, front and center. The thing running live, not the repo. A screenshot that loads and a link that works. GitHub proves you code. The site proves you can decide what matters and put it first.

Those are different skills. The second one is the one people actually pay for.

Playground

There's a second reason, and writers don't get to use it. Building the site is itself the signal.

You pick a framework. You wire up a deploy pipeline. You connect a domain. You ship something small and real. That's a portfolio piece about building the portfolio. The medium is the message, and the message is "I can ship."

A writer building a website is often… procrastinating. Been there...

A developer building a website is working. In my eyes, at least.

It's also where you get to try the thing you can't try at your day job. A new framework. A weird animation. A stack you've been curious about. Low stakes, your own rules, and the result markets you while you learn.

Over-building

Now the catch.

Many portfolios are over-built. Someone spends three weekends on a Next.js site with scroll animations, a headless CMS, dark mode, particle effects, and a blog they will never post to. For five projects and one contact form. Been there too...

The site becomes the project that eats the time you'd spend on projects worth putting more effort into.

And the maintenance doesn't leave. Dependencies. The framework you picked two years ago wants updating. Your own site starts throwing build errors. This is why I tell writers to not build a website, and developers walk straight into it because building is fun and shipping the boring version isn't.

But ship the boring version first. A clean one-pager is still pretty cool. Who you are, three things you built with links that actually work, and how to reach you.

Hey, even put it on Carrd* to see how it goes. Done in a weekend, not a quarter.

The people looking at your site want to know if you can build their thing. Not whether your homepage has a custom cursor effect. Mine has...

The domain

The one thing worth owning here is the same thing I told writers to own. The domain.

yourname.dev. That's the asset. The site behind it is an implementation detail. It can be a one-pager today, a full portfolio next year, a redirect to your best project the year after. The URL stays. People may remember it.

I made this point about writers too, in the piece on why most of them don't need a website anymore. It holds harder for developers I feel like.

The Bottom Line

For a writer, the personal website is a nice-to-have option. It exists. Some still use it. Nobody really needs it.

For a developer, it's a cool part of the job and a showcase of the abilities. You don't point people somewhere else. You build the place they judge you by, and building it well is half the argument.

*this is an affiliate link

My Medium friends can read this over on Medium.

In an earlier piece I called Antigravity "Claude Code + Cowork and faster." That is not really true, and the mix-up is everywhere right now.

Claude Cowork and Google Antigravity get compared a lot. Search for a Gemini equivalent to Cowork and Antigravity is what shows up. But they are not built for the same person, and once you look at what each tool is for, the answer is somewhere else entirely.

What Claude Cowork Is

Cowork launched as a research preview at the end of January 2026. On July 7 it expanded to mobile and web, with sessions now running in the cloud by default. Start a task on your laptop, check progress on your phone, close the lid, the work keeps going.

Under the hood it runs in a sandboxed Linux VM, powered by Opus 4.6. It reviews a plan with you before it executes, rather than just running off on its own. And it connects to over 50 tools through MCP: Slack, Notion, Jira, HubSpot, Figma, Snowflake.

It is included from the Pro plan ($20 a month) up. Every paid tier gets the same features. Max just gets you more usage per session. Cowork also went enterprise recently, with role-based access, spend limits, and an analytics dashboard for admins.

The 90%

Anthropic published usage data from 1.2 million anonymized Cowork sessions across more than 600,000 organizations. Software development made up 8.7% of it.

More than 90% of Cowork sessions have nothing to do with coding.

Business process work, the largest single category at 33.4%, means status reports, onboarding checklists, reconciling spreadsheets. Content and copywriting came in at 16.4%. Anthropic calls this "the work around the work," the stuff that touches every job but is nobody's actual job title.

That number is the whole point. Cowork was never really built to compete with a coding agent. It was built for the people who never open a terminal.

Why Antigravity Isn't That

Antigravity launched in November 2025 alongside Gemini 3. It is a developer IDE, a fork of VS Code, built for people who ship code. Its Manager View lets you spawn and direct up to 16 agents in parallel. It has an SDK and a CLI for building your own workflows.

That is a coder's tool. Cowork is a knowledge worker's tool. Comparing them head to head is comparing a terminal to a spreadsheet. If you want to know what Google has for the audience Cowork is chasing, Antigravity is the wrong place to look.

The Answer: Gemini Spark

Google announced Spark at I/O 2026 in May. It runs on Gemini 3.5 plus Antigravity's agent infrastructure, but it is aimed at the same non-coder crowd Cowork is. A persistent, 24/7 background agent. Give it a task, it keeps working even with your phone and laptop off.

The integration goes deeper into Google's own stack than Cowork can go into: Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, Maps, plus Tasks, Keep, Canva, Dropbox, Instacart, OpenTable, Zillow, and custom MCP connections on top. A macOS beta rolled out at the end of June, doing file sorting and spreadsheet building right on the desktop.

This is the Gemini answer to Cowork. Same audience. Same "hand it a task, walk away" model.

But right now Spark is Google AI Ultra only, 18 and up, US only. Cowork is a $20 Pro plan add-on available to basically anyone. So the for-now answer to "is there a Gemini equivalent" is yes, it exists, but most people can't use it yet.

Worth a mention too: Google Workspace Studio, which lets you build your own no-code automations ("Gems") inside Workspace. Less ready-made agent, more a construction kit. A third option depending on how hands-on you want to be.

Downsides

Security first, because it's the serious one. Within 48 hours of Cowork's January launch, researchers at PromptArmor got it to exfiltrate financial documents. The trick: a Word file with hidden white-on-white text carrying the injected instructions, which told Cowork to upload the user's files to an attacker's account. The sandbox blocks most outbound traffic, but it trusts calls to Anthropic's own API, and that trust is exactly what got exploited.

There is no fix for prompt injection yet, for any of these agents. Nobody has found the equivalent of what parameterized queries did for SQL injection. If you're handing Cowork or Spark real files with real financial or client data in them, that's a risk.

Price is the other one for Cowork. Pro gets you in the door, but heavy use pushes you toward Max, and Max is not cheap.

And for Spark, the gate itself is the downside. Ultra-only, US-only, adults-only. A lot of the people typing "gemini equivalent to claude cowork" into Google right now are going to find out it exists and still can't touch it.

I have not lived in Cowork day to day myself. Claude Code is still my default for building, as I wrote more recently comparing it against Antigravity and Codex. This one's built from what Anthropic and Google have both put on the record, not a week of hands-on testing.

The Bottom Line

Yes, Gemini has a Cowork equivalent. It's called Spark, not Antigravity.

Google just built it for a smaller audience than Anthropic did. For now. That will change soon enough, I am sure.

Hi friends,

This week I noticed how much of my routine runs without me now. Gumroad shipped a chat agent that handles support tickets on its own. My own Substack workflow runs on a script I built. Even picking an AI tool comes down to three good options depending on where I'm sitting.

And then there's Kimi K3, a new model out of China claiming to match Claude for a third of the price. Turns out that claim needs some math.

Here's my week.

A Chinese AI Model Beat the Mighty Fable 5

Kimi K3 beats Claude Fable 5 on a few real coding benchmarks, and Moonshot is selling it as a third of the price. Run the token math first. K3 needs more than double the tokens to finish the same job, and that number eats the discount fast.

Read it on Medium or on Substack.

The Best Personal Blog Sites in 2026 (In my Opinion)

Substack, Medium, WordPress, or building your own with AI now. I ranked the platforms I'd recommend, plus the new options.

Read it on Medium or on Substack.

Gumroad Shipped a CLI and an AI Agent

Gumroad's new Agent closed 478 of 479 support tickets in one 24 hour stretch. I ran the CLI myself on my own shop, here's what worked and what I'd still double check.

Read it on Medium or on Substack.

How I Automate My Substack Workflow

The writing stays mine. Everything after it, the formatting, the header image, the Note I used to forget about, now runs on a script I pieced together myself.

Read it on Medium or on Substack.

Three Ways I Use AI to Work

Web app, desktop app, or terminal? I use all three for different jobs. Only one of them actually won my daily driver spot.

Read it on Medium or on Substack.

🤔 FAQ

Is Kimi K3 actually a third of the price of Claude Fable 5? Per token, yes. Per finished task, not really. It needs more than double the tokens to get there, so the real gap shrinks fast.

Do I need to know how to code to automate a Substack workflow? I pieced mine together from three different open source projects, since Substack has no public API. Not a weekend project, but doable if you're patient.

Which AI tool should I actually start with? Depends where you already work. Web app for quick one-off tasks, desktop app if you want file and screen context, terminal if you're already comfortable there.

That's #157. Back soon. ✌️

Get the Substack Automation Blueprint

My Medium friends can read this over on Medium.

Moonshot AI dropped Kimi K3 on July 16, and by the next morning half my feed was making the same claim. As strong as Claude. Maybe stronger. Crazy benchmarks. For a third of the price. It's a steal. Let's switch to Kimi K3.

Some of that is true. Not all of it. So hold your horses.

The Numbers

K3 is a 2.8 trillion parameter model, the biggest open-weight release yet. That's pretty remarkable.

Full weights land July 27, but the API went live on release day, and Artificial Analysis already ran it through their Intelligence Index. Score: 57. Rank #4 out of 187 models.

Just behind Claude Fable 5, ahead of Claude Opus 4.8.

On a few specific benchmarks it does better than "just behind." K3 beats Fable 5 on Terminal-Bench 2.1 and SWE-Marathon, both agentic coding tests. Incredible, right? It is!

On the WebDev Arena leaderboard, the one that scores frontend code output, K3 scored 1679 against Fable 5's 1631. A real win, not a rounding error.

So the headline isn't wrong in that sense. K3 is close. On some tasks, closer than close. To Fable 5 and also GPT 5.6.

But there's more to the story.

Speed

Kimi K3 gets less impressive here. K3 outputs 62 tokens per second. The median for models in its price tier is 73.7. Time to first token: 1.99 seconds, not fast for an API call you're about to run in a loop.

Matching Claude on intelligence and trailing the field on speed is a real trade.

I use Claude Code daily, after going back and forth between it, Antigravity, and Codex, and a model that thinks well but drags on output changes how a session feels.

Especially a long one.

The Token Problem

K3 is chatty. And thus, token-hungry.

In Artificial Analysis's eval suite, it generated 130 million tokens finishing the same tasks that averaged 63 million tokens across the field. More than double.

That's actually an improvement. Wild! K3 uses 21% fewer output tokens than its predecessor, K2.6. Moonshot is working on it. But 21% off double is still roughly double.

Now the math.

K3 costs $3 per million input tokens and $15 per million output tokens, up from K2.6's $0.95/$4 and now sitting at Sonnet-tier pricing. Fable 5 charges $10 input and $50 output. On paper, K3 runs about a third of Fable 5's price per token.

But price per token isn't what you pay. Price per finished task is. If K3 needs twice the tokens to get there, that "third of the price" number shrinks toward two-thirds or even less. A third of the price times double the tokens is not a third of the price anymore.

In some tasks with loops or review rounds, it can get even more expensive than Fable 5 and especially GPT 5.6, which is very affordable for the power it packs.

So the speed and token consumption: for a single prompt, that's annoying. For an agentic coding session, the kind K3 is explicitly built for, sustaining long engineering runs and orchestrating terminal tools, it compounds. Every extra token in every tool call adds up over a session that can run hundreds of turns.

The $0.30 cache hit rate, a 90% discount, softens this some. Heavy agentic use leans on caching hard, and K3's 1M token context window gives you room to do it. But caching offsets repeated context, not the extra tokens the model generates fresh every turn. The verbosity problem doesn't go away because you're caching the prompt.

Downsides

Cost first, since that's the whole point of this one.

The "one-third the price" headline is a per-token number, not a per-task one. Run your own numbers before you move a production pipeline over.

Full weights aren't out yet. Anyone wanting to self-host is stuck with the API until July 27.

R&D World reported getting rate-limited testing K3 in the days after launch. Early demand on a brand-new API, probably not permanent, but worth knowing if you're planning to lean on it this week.

The Bottom Line

K3 earns its benchmark wins. Beating Fable 5 on WebDev Arena and Terminal-Bench is a strong result from an open-weight model, not a cherry-picked one.

But the "third of the price" claim needs an asterisk next to it. Run the same job on K3 and Fable 5, and the gap between cheaper and cheaper enough to matter depends entirely on how many tokens it burns getting there. Sometimes that gap disappears pretty quickly.

I'll probably try it once the open weights land. Not switching anything yet.

My Medium friends can read this over on Medium.

"Where should I start a personal blog?" is a question I get a lot.

Since it's 2026 and AI is in full force, the answer got more complicated and at the same time simpler. You can now build your own blog with AI, for close to nothing, and pretty easily.

So let me lay out my picks. The popular sites, the underdogs, and the build-it-yourself route I took for my own sites.

No single winner here. Different people, different picks.

The 4 Everyone Lands On (First)

Start with the names you probably know by now. A quick take on each, 2026 edition.

Substack is the big blog spot for people who want the blog + free and paid newsletters. You own your list, you get paid straight from readers, and Notes feeds you new ones. It's chasing video now, a live studio, a TV app, the lot. It also takes 10% of your subscription money, which is why a few bigger writers may need something else.

Medium is where I've written since 2021. Its 2026 bet is human writing against AI slop. AI filler gets buried (although it doesn't work all too well), curation is done by people (a lot), and the Partner Program pays without ads. Five dollars a month to read and earn. I am still a huge Medium fan and I still believe everyone who writes online should be on Medium.

WordPress runs a huge slice of the web. Still. Total control if you self-host, endless plugins. Customization in any shape or form. WordPress feels like the dinosaur of the bunch, because it is, and that has upsides and downsides. You can build almost anything with WordPress, and often cheaper than with anything else, but WordPress does come with a bit of baggage, slowness, maintenance, and more.

Ghost is the sleek, independent one. The new WordPress if you will. Ghost 6 wired it into the social web, so people follow your blog from Mastodon or Threads, and it takes 0% of your revenue. Starts at $15 a month. Gets more expensive with more readers and subscribers. It has the best backend of the bunch and great features.

If you want my personal ranking of these four, I keep a separate list.

The Underdogs

Now the fun part. Smaller platforms, fewer users, more "niche".

Bear Blog. Fast, no JavaScript, no trackers, no cookies. It has a discovery feed that sends readers to new blogs, which almost none of these have. Free to start, or $5 per month for more. It's super basic but that also makes it quick, reliable, and secure.

Mataroa. Minimalist, built-in newsletter, run on a stay-cheap-forever promise. Export anytime, no lock-in.

Blot. You drop text files in a folder, they turn into your website. Five dollars a month. That's the entire product. Doesn't get much simpler than this.

None of these will hand you Substack's or Medium's audience. That's the trade. What you get back is speed, low cost, and something that feels like yours.

Build Your Own With AI

This option didn't exist for most "normal" people a couple years ago.

I run a few sites of my own. Always have. But now, there is no WordPress, no page builder, no monthly platform fee. Plain files I own, on cheap hosting. A year ago that meant knowing how to code. Now it doesn't.

I build mine with Claude Code. I describe the page, it writes the code, I check it over. What used to eat a week takes an evening now.

You don't even have to hand-roll it. Tools like Claude Code, Codex, Lovable, Bolt, or the gazillion others on the market now take a description and give you a working site in minutes.

And hosting is almost free. A static site on Cloudflare Pages gives you unlimited bandwidth for nothing. Point a domain at it and you're live.

I did a version of this years ago, ditching WordPress for a simpler setup. The difference now is that AI removes the one wall that used to stop most people. The code.

There's a catch, though. You own the upkeep. No built-in readers show up just because you built a nice site.

The Trade-offs

The popular sites hand you an audience but take a cut and set the rules. Substack has discovery and social media, Medium has curation and boost, but both can change the rules tomorrow.

The underdogs give you speed and independence, but the reader has to already know your URL.

Building your own gives you the most control and the lowest cost, and asks the most of you back. You're the platform now. Upside: you can make it look and work the way you want. Downside: you need to know if the AI has done everything correctly.

The Bottom Line

So where do you start?

If you want readers to find you and maybe pay you, go Substack or Medium, or both. That's what I do.

If you want something fast, cheap, and yours, Bear Blog or Mataroa, and you'll be happy.

If you want to build the exact thing in your head and own every pixel, do it yourself with AI. It's finally a normal-person option, not a developer one.

I've got a foot in two of those camps. My writing lives on Medium and Substack. My own little sites run on code I built with AI.

Both feel good. For different reasons.

Get the Medium + Substack Dual guide