Two weeks ago, the Drupal Association convened Drupal Pivot North America, a room full of Drupal agency owners and senior leaders asking what this ecosystem should become next.
It was a room of friendly competitors, or "coopetition" if we want the ugly but useful word: firms that compete for the same work and still know the shared platform matters. Pivot operates under Chatham House Rule, so this is not a transcript. It is my takeaway.
The takeaway was simple: Drupal has an enormous AI-era advantage, but the world will not discover that advantage automatically. Agents will not know it by magic. Buyers will not know it because we say "enterprise CMS" one more time.
Truth is not a distribution strategy.
The short version. AI made confident prose free, so I didn't write a pitch — I shipped the proof on drupal.org and you can install every piece of it. Four new projects, all released on drupal.org and covered by Drupal's security advisory policy: GEO Starter and GEO Starter JSON-LD make governed content legible to answer engines; Contentful Migration is the exit ramp off proprietary platforms; Content Packages lets content move through review like code — and this post shipped through it. Plus upstream contributions to four community modules. Idea to production: about two weeks.
Where do governed facts live?
Dries' public framing is useful and, I think, correct: AI gets teams to a prototype fast; Drupal CMS gives those prototypes foundations that last.
The AI conversation around websites is still too shallow. Too much of it is about page generation, chat widgets, and whether a search result has a little sparkle icon next to it. The deeper question is the one in the heading: where do governed facts live?
If an agent is going to answer on your behalf, summarize your services, cite your evidence, respect permissions, or help an editor act inside a workflow, it needs more than a page. It needs structured content, stable URLs, source visibility, review dates, access rules, and a system where humans can see what changed, who approved it, and what is safe to expose.
That is Drupal CMS. Not theoretically. Architecturally.
The agency question is not "AI or Drupal?" It is "what stays in Drupal, and what goes into the AI harness?" Durable content structure, paragraph architecture, permissions, editorial workflows, review state, and governance belong in Drupal. AI should help create, migrate, test, and operate those systems. Sometimes the best use of AI is using it to remove the need for AI: turn a repeated judgment into a field, a workflow, a recipe, a migration path, or a governed source of truth.
That is also how I think about Drupal Canvas. Canvas is valuable when it preserves structure and reduces waste. If it helps a team move faster while keeping Drupal's content model intact, great. If a visual workflow cannot carry the paragraph structure and governance editors rely on, then it is not a replacement for the system underneath it.
So I built the proof — here is the scoreboard
The golden rule of this moment is not "tell me." It is "show me." So instead of writing the pitch, I shipped the work and put it where you can check it. Here is the whole portfolio at a glance — with honest status: each new project is released on drupal.org and covered by Drupal's security advisory policy, while the upstream work doesn't count until maintainers accept it.
The rest of this post is just those cards, in sentences.
What does "built for answer engines" actually mean?
GEO Starter is a Drupal CMS site-template recipe for teams trying to make their content legible to answer engines and agents. It gives you typed Service, Answer, Article, and Evidence Source content; controlled vocabularies; moderation; reviewed-date provenance; sample content; an editorial dashboard; sitemap support; and documentation that says what is proven now and what is not. It is released on drupal.org and covered by Drupal's security advisory policy.
Its companion, GEO Starter JSON-LD, emits schema.org JSON-LD from the same governed fields the visible page renders. It does not invent claims for the machine layer. It publishes what the site actually says, in a form retrieval systems can parse and cite — and a companion submodule adds an /llms.txt index so crawlers and agents can find the governed content in the first place.
What does that look like in practice? It looks like this page. The card below is the same idea GEO Starter JSON-LD ships: the governed facts of this post, rendered for you on the left and emitted for a machine on the right. Flip between them, copy the structured data, and paste it into any assistant.
This is not a promise of AI rankings, citations, or rich-result eligibility. Nobody honest can promise that. It is a foundation: clean structure, visible sources, canonical pages, machine-readable facts, and editorial governance. We are already applying the same pattern in real evaluation work — a current association demo publishes page-level JSON-LD, exposes an /llms.txt index, and puts a "see how an answer engine reads this page" card on each event page. The point is not that the assistant is magic. The point is that Drupal CMS can publish governed facts in a shape assistants can use without guessing.
How do you get off a proprietary CMS?
Another Pivot thread was more practical: other platforms are winning business. Some look easier to start with. Some have sharper sales motions. Some hide the long-term cost of renting your content model from a proprietary vendor. We can complain about that, or we can make switching easier.
Contentful Migration is my answer. It moves a Contentful space into Drupal on Drupal's Migrate API and handles the parts a plain export drops: rich-text rendered to clean HTML, embedded entries and assets, media de-duplicated by SHA-256 checksum, locales, and reference fields rewritten to links that survive the move. It is released on drupal.org and covered by Drupal's security advisory policy, with the migration runtime — including real two-pass Migrate runs that resolve embeds, stage assets to Media, and attach translations — covered by unit and kernel tests. It is not a live sync product and not a clone of Contentful's delivery API. It is a path off the platform and into an owned Drupal content model. Telling teams "Drupal is where governed content belongs" is marketing; handing them the exit is the argument.
I also contributed back to the broader AI Migration work. If AI is going to change how Drupal sites are built and migrated, the right answer is not a private pile of clever code. It is shared infrastructure, tested in the open.
Can content be reviewed like code?
Content Packages lets selected Drupal content live as canonical files with YAML front matter, then validates, imports, exports, and diffs that content against the live site. It is released on drupal.org and covered by Drupal's security advisory policy, with a kernel test suite behind it.
This post is one of those packages. The argument is running in the system the argument is about: authored as a file, reviewed in a pull request, imported into Drupal, diffed against the node, and kept in version control with the rest of the site. You are reading the proof, not a description of it.
That is the same thesis again. If content is going to be used by agents, search systems, editors, reviewers, and humans, it needs provenance. It needs review. It needs a source of truth.
Sometimes the proof is a number
Every platform now has an essay explaining why it is ready for AI. When a claim costs nothing to make, it stops carrying information — and running code still costs something you can measure. On the AI Context project I measured what it costs to re-send an agent's full context on every loop. On one real task — a single heading edit — more than half the tokens were spent re-sending context the agent already had.
The number is sitting on a public issue where anyone can re-run it and argue with me. That is the difference between a claim and a measurement. I also added deterministic benchmarks and a test suite to AI Google Analytics and filed skill proposals — each with its own eval suite — against the AI Best Practices initiative.
The honest "how"
The honest how is AI coding agents: Claude Code most of all, with Codex and local models through Ollama where they fit.
I am not a developer in the traditional sense, and I did not hand-write every line. I chose the targets, argued with the architecture, read the diffs, ran the tests, filed the issues, and cut the releases under my own drupal.org account. The models multiplied what I could ship. They did not decide what mattered. That distinction is the whole story for Zivtech right now: AI is in our process, human judgment is still the governing layer, and the gains flow back into better client work and better open-source infrastructure.
What I am not claiming
I am not claiming GEO Starter guarantees citations, rankings, or AI visibility. It does not.
I am not claiming every Drupal site should be rebuilt around this recipe. It should not.
I am not claiming AI agents make expertise optional. They do the opposite — they make weak judgment cheaper to express and strong judgment more valuable.
And I am not claiming Drupal wins by default. Drupal has the architecture, but architecture only wins when people can see it, install it, test it, migrate to it, and explain it to buyers.
FAQ
Is Drupal really better than a proprietary CMS for AI? For governed, agent-readable content, the architecture favors it. Drupal already treats structured entities, taxonomy, revisions, moderation, permissions, APIs, and rendered public pages as core publishing concerns. That is exactly the layer an agent needs to answer accurately on your behalf — and it is open, so the rules are inspectable.
What is GEO, and what does GEO Starter do? GEO is Generative Engine Optimization: making your content legible to answer engines and agents, not just to traditional search. GEO Starter is a Drupal CMS site-template recipe that ships a governed content model (Service, Answer, Article, Evidence Source), an editorial workflow with review dates, and a companion module that emits schema.org JSON-LD from the same fields the page renders.
Does this guarantee AI citations or better rankings? No. Nobody honest can promise that, and GEO Starter's own documentation says so. It builds the structural foundation — clean structure, visible sources, machine-readable facts — that citation- and answer-engine paths favor. It does not promise outcomes.
Can I move my content from Contentful into Drupal today? Yes. Contentful Migration handles the hard parts a plain export drops — rich text, embedded entries, media de-duplication, locales, internal links — over the core Migrate API, and it's released on drupal.org under Drupal's security advisory policy. Install it, run it against your space, and file issues where it falls short.
How was all of this built, and by whom? By me, directing AI coding agents — Claude Code most of all. I chose every target, read every diff, ran the tests, and cut every release under my own drupal.org account. The models multiply what I can ship; they do not decide what matters.
The assignment
If you are evaluating CMS platforms for the agent era, ask harder questions. Where do governed facts live? Can the system publish structured data from the same fields humans review? Can it expose content through APIs without losing editorial control? Can you migrate out of proprietary content stores? Can content be reviewed like code? Can your team own the result? Drupal CMS and GEO Starter are where I would start.
If you are in the Drupal community, build the recipes, migrations, and demos that make the advantage obvious. Look at how other platforms are winning, then build the thing that lets Drupal win on better terms.
Since 2008, Zivtech's purpose has been to illuminate. In the age of agents, that means showing our work and letting you check it. Install the recipe. Try the migration. Read the code. File issues. Send patches — PRs and MRs are welcome.
Drupal CMS, with recipes like GEO Starter, was built for the age of agents. Now we have to make sure the world can see it.