Moving a website from one platform to another used to take weeks. You rebuild every page, copy every post and fix every broken link. That cost is the reason many designers stay on a platform they stopped liking years ago.
Dries Buytaert, the founder of Drupal and co-founder of Acquia, thinks that reason is about to disappear. In Website migration costs are going to zero, Buytaert writes: "Within three years, I expect the technical work of moving a website from one platform to another to cost close to nothing."
We agree. We build Splitleaf on the same bet. But the post asks what happens when you can move a site between platforms for free. We think the better question is why your site has to live inside a platform at all.
The short answer
When AI does the migration, switching costs stop protecting the vendor. Just look at the mass migration from Webflow in recent months that is well documented on Reddit and LinkedIn. What is left to hold you in place is either a contract, complex dynamic app features or hosting. I
Why we agree that migration is going to zero
The prediction rests on a method that works, a method we use. An agent builds a copy of a site on the new platform, then compares it with the original, page by page using both screenshots and code. Testers call this differential testing: give two systems the same input and look for differences. William McKeeman described it in 1998, in "Differential Testing for Software" in the Digital Technical Journal (dblp).
Agents do their best work when they can check their own result. Anthropic's Claude Code best practices put this at the top of the list: give the agent a way to verify its work, with tests, screenshots or expected output.
A website is easy to check that way. The old page exists. The new page exists. Open both at the same screen width and compare them.
For marketing sites, most of it works today
Buytaert gives it three years. For a marketing site, a portfolio or a company blog, much of it already works.
- From a Webflow export. Splitleaf imports the pages, classes, components, interactions and CMS collections. See how.
- From a live address (beta). Give the agent the address of your site. It reads the pages and the stylesheets, downloads the images and rebuilds the site as a folder of HTML, CSS and content files. Then it opens the rebuilt page in a real browser, at the screen width you choose, and checks it against the source.
You see a preview before anything changes. One click undoes the full import.
It is not free of work yet. Sites that build their content with a lot of JavaScript do not import perfectly from a live address. Webflow's code export needs a paid Workspace plan, and it leaves out ecommerce, user accounts and translated content. Forms and site search stop working outside Webflow (Webflow help). The technical work is shrinking fast. The decisions are still yours.
The ten predictions, and our answer to each
The post makes ten predictions. Most of them describe a vendor that has to work harder to keep you. Each one also shows something that can still hold a site in place.
| Buytaert's prediction | Our view |
|---|---|
| 1. Switching costs stop protecting vendors | Agree. Then the subscription is the only thing left to defend. |
| 2. Proprietary vendors sell hosting and operations, like Open Source companies | Partly. A static marketing site needs little of that. |
| 3. Agents favor a CMS they can try | Agree. The easiest thing for an agent to try is a folder of files. |
| 4. Vendors pay for agents to recommend them | Like marketing took over Google searches it will infiltrate agent chats as well |
| 5. The next lock-in is the agent's memory | Agree. The most important point in the post. That is why Splitleaf lets people own their own agent connection and chat. |
| 6. Technical requirements become evals | Agree. |
| 7. Vendors compete with contract buyouts | Agree, for subscriptions. |
| 8. Approved-vendor lists become a moat | Agree. |
| 9. Agencies will recommend a CMS they don't know | Agree. It is a real risk. |
| 10. A CMS no longer has to be a ten-year bet | Agree. |
Three of these deserve more than a table row.
Agent memory belongs in your repo
Buytaert warns that an agent which migrates and maintains a site collects its requirements and past decisions. "If that memory stays inside one AI provider, changing agents means rebuilding that context. Site owners will want to keep that context portable."
We agree, and the fix is simple. Keep the memory in plain text files, next to the site. In Splitleaf, the agent's memory is:
- Site rules in an
AGENTS.mdfile. AGENTS.md is an open format that the Linux Foundation's Agentic AI Foundation looks after. Codex, Cursor, Gemini CLI and GitHub Copilot's coding agent read it. - A site profile: what you sell, to whom, and how you write.
- Writing rules for each collection, such as your blog or your case studies.
- A brief for each post, with its sources, search terms and plan.
All of it is Markdown in your site folder, with history in a GitHub repo you own. Change agents, and the new agent reads the same files. Change from Splitleaf to another tool, and the memory goes with the site.
The services that are left are smaller than predicted
Prediction 2 says that when code gets cheap, "customers will keep paying for the services that keep their code running: deployment, hosting, security, scaling, and monitoring."
That is true for a web app, or for a CMS with a database and server code. A marketing site needs little of it. A static site is files. There is no database to patch and no server code to attack, and file hosts already handle the traffic. We wrote more about this in Time to reduce your website's attack surface.
So for marketing sites, selling operations is a thin business. What is left is the tool you design and edit with. We think you should buy that tool once, like a camera.
The best exit path is no migration
Prediction 10 ends with this sentence: "A tested exit path becomes more important than long-term stability and reliability."
Here is a simple test. Can you open your site without the vendor? A Splitleaf site is HTML, CSS and Markdown. It opens in any code editor. It publishes to any host. You can test the exit on the first day, because the exit is the folder you already have. More on owning your code.
Cheap migration is good. No migration is better.
A note on paid recommendations
Prediction 4 is partly here. OpenAI announced ads in ChatGPT in January, and they now run in more than 40 countries. OpenAI says the ads are labeled, kept apart from the answer, and do not change what ChatGPT recommends. Today that makes them a paid slot next to the answer, like search ads. It is not yet a paid answer.
The defense is the same either way. Choose the model yourself, and ask the agent to show a source for each claim. Splitleaf runs on your own key from Anthropic, OpenAI or OpenRouter, and every brief it writes lists its sources.
Who is talking
Buytaert founded Drupal and co-founded Acquia. Several of the predictions favor open, self-service platforms like Drupal. We build Splitleaf, a desktop app with a one-time price, and several of our answers favor us. Read both with that in mind.
The predictions we agree with most are the ones that cost vendors the most: no lock-in by contract, by hosting or by memory.
What to do before your next renewal
- Export what you can now. The code, and the CMS content as CSV files before they are locked in behind another tier. (This has already happened on multiple platforms.)
- Ask where the agent's memory lives. If the answer is "in our platform", it is lock-in.
- Test the exit. Can you open and publish your site without the vendor?
- Run a test migration. Import a copy of your site somewhere else, and compare it page by page with the original.
Download Splitleaf and import your site for free. You need a license only when you publish.