Preparing your CMS for AI answer engines takes several key phases: set up crawler access, structured data and tracking; analyze citation gaps; generate content that opens with a direct answer and real expertise; and engineer the cycle so it runs every week. The platform carries setup and engineering. Your editors carry analysis and writing. Neither half works on its own.
That’s the short version of the session Miles De Feyter and I ran at Brightspot’s user conference, Illuminate, this October. The full session is below, along with the workflow we use on our own marketing site, brightspot.com, which moved from the seventh to the second most cited domain in the category we’re tracking against in about five months.
- Publish something worth citing. Structure helps AI engines read a page. Original data and expert insight are what earn the citation.
- Spread your citations. After one ChatGPT change, Reddit’s share of its citations fell 86% in under a week. Build presence across your own site, your documentation and trusted publications.
- Track less, move weekly. AI visibility responds in about a week. Pick a handful of metrics that drive decisions, then test and adjust.
- Split the work. Your CMS owns setup and engineering. Your editorial team owns analysis and writing.
A year ago, about 60% of Google searches ended without a click, and our advice was that success meant being the answer, not the click. This year the data caught up.
Digital Content Next tracked 19 publishers, including The New York Times, Condé Nast and Vox, over eight weeks. Google referral traffic fell 10%, and losses outpaced gains two to one. Non-news content took the bigger hit, down 14% against 7% for news. That matters for most enterprise teams, because evergreen and product content is exactly what AI summarizes well. Pew Research separately found that people click through less after seeing an AI summary.
Google tightened the other side at the same time. It ran 4 spam updates in a year, all aimed at the same thing:
- August 2025: autogenerated and programmatic content at scale
- March and June 2026: thin, low-value and AI-generated content, plus an upgraded SpamBrain in June
- August 2026: scaled content abuse and thin affiliates, with some sites losing rankings for more than 200,000 queries
Google’s official line is that using AI to create content is fine in principle. Its ranking foundation is still E-E-A-T, and AI on its own can only restate what’s already been said elsewhere. The bar for ranking and the bar for citation are rising together.
The measurement picture shifted too. Some sites lost keyword rankings after the May core update and gained AI Overview citations in the same window. If you only watched rank tracking, it looked like a loss. If you also watched citations, some of those sites came out ahead.
Then there’s the clearest warning of the year. On August 8, ChatGPT changed how it scoped site searches in its fan-out queries. Six days later, Reddit’s share of ChatGPT citations had collapsed by 86%. Google AI Overviews barely moved. Last year’s GEO advice said “get on Reddit.” That playbook changed in a week, and much of the share Reddit lost moved to official brand sites and documentation. Both of those run on your CMS.
Setup is the most technical phase, and the one most teams rush. It decides what AI systems can reach, read and trust before anyone writes a word.
Best CMS for LLM SEO: 7 AI-ready platforms for 2026
Decide who gets to crawl you
“Crawler access is now a business decision,” says Miles De Feyter, who leads SEO and GEO for Brightspot Creative Services. Cloudflare now splits AI bot permissions three ways: search, agent and training. Licensing is starting to matter as well. Otterly found that publishers with direct OpenAI licensing deals see about 48% more citations inside OpenAI, and Google has started a pilot that pays publishers when their content shapes AI answers.
The trap is the blanket setting. “Multi-purpose crawlers follow your strictest rules,” De Feyter says. Googlebot handles both search and training, so a blanket “block AI” rule can quietly cost you search visibility, or block Google altogether. Most of our customers want to be indexed in as many places as possible. Whatever you decide, decide it per bot type, on purpose.
Model content as typed fields
Most CMSs store a product page as one big rich-text field. Headline, price, features and use cases all sit in the same HTML blob, formatted for a browser, not parsed for meaning. An AI answer engine can’t reliably tell a price from a feature in that blob. It guesses, and it often guesses wrong or skips the page.
Model the same content as typed fields instead: name, price as a number, features as a repeatable list, use cases as their own array. An AI system asking “what does this cost” pulls the price field directly instead of parsing a paragraph to find it.
A media company modeling a news article hits the same problem in a different shape. Headline, byline, dateline and body copy usually get typed and separated. Related topics, sourcing and a summary field often don’t, because they read as extras when a human is the only audience. An AI answer engine needs that summary field more than the body copy. It’s the field most likely to be quoted directly.
What GEO actually asks of your CMS
Expose clean markup, schema and APIs
AI systems reach your content two ways: as rendered pages that answer engines crawl and cite, and through APIs that RAG pipelines and agents query directly. Both need to be clean.
On the page side, “LLMs are also very good at just parsing HTML,” says Lee Teague, Brightspot’s Chief Technology Officer. “HTML is a very machine-readable language, and there’s a lot of structure that’s inherent in HTML.” Add article, FAQ and author schema so the engine knows what it’s looking at before it reads a word of copy. Generate that schema from your typed fields, so it stays accurate without anyone editing it by hand.
The dependency chain runs one way. Structured content enables clean API output. Clean output enables reliable citation. Reliable citation enables inclusion in RAG pipelines and the agentic workflows built on them. Break the chain here, and nothing after it happens.
Treat markdown and llms.txt as experiments
At Brightspot’s Ocotber 2026 user conference, an attendee asked whether Brightspot would offer markdown versions of web pages. “This is not natively an out-of-the-box capability,” De Feyter said. “But we have multiple customers that have built it out in their specific Brightspot instance.” The SEO community is still skeptical of llms.txt, with none of the major AI companies reliably supporting or crawling it yet. Where markdown has shown traction is on documentation sites, where the content is highly structured and every page has a clear place in the taxonomy.
Set up tracking before you publish
A year ago we tracked prompts in massive spreadsheets and ran them through each chatbot by hand. That’s matured. Bing Webmaster Tools added a free AI performance report in February that shows citations, cited URLs, sample queries and trends, though it can’t tie citations to traffic yet. Profound, Semrush and BrightEdge all have mature AI citation tracking now.
The term worth learning is share of model: the percentage of AI answers in your category that mention your brand. It sits next to share of voice, the keyword measure most teams already know. The risk is overload. Overlapping platforms use inconsistent methods, and you can end up with hundreds of prompts across several tools. To solve for this, pick two or three measurements that drive decisions and keep the rest in your back pocket to avoid analysis paralysis. On brightspot.com, we track about 20 topics and 250 prompts through the AI tracking tool we’re currently using, Profound.
We are using Profound’s SAGE model (setup, analyze, generate, engineer) to run this work. Credit where it’s due: SAGE is Profound’s framework, not ours. Analyze is where our editorial team takes over.
Every week we pull citation gaps from Profound. That’s the easy part, because it’s a data pull. The hard part is deciding which gaps deserve resources. Before a new brief gets the green light for brightspot.com, it has to pass 3 questions:
- Is there a real practitioner or subject matter expert behind it? If we can’t add a perspective of our own, we’d only be producing something derivative.
- Is it net new, or a refresh? Content we published years ago may already be earning citations and only need updating for answer engines.
- Does it map to a prompt we already track? If it doesn’t, we can’t tell whether it worked.
Sometimes the data points at the fastest way to close a gap, like a ten-platform ranking listicle with no unique perspective, so we say no. A fast fix that doesn’t build E-E-A-T credibility isn’t worth the slot.
The CMS has a role here too. You can only make the refresh-or-new call if you can see what you already have, which content type it is and how it’s performing.
Real statistics, unique research, customer testimonials, insights from subject matter experts. They’re the pieces of your content strategy that AI cannot create. That’s the gold dust.
Generate is where structure and substance meet. Across pages that ChatGPT, Perplexity and Google AI Overviews cite, we see 6 patterns again and again:
- A direct answer opening. Answer the likely prompt in the first 40 to 60 words. That’s the chunk AI lifts.
- Question-shaped subheads. Write H2s the way people ask AI: “How does X work?” instead of “Overview.”
- Scannable body copy. Paragraphs of 2 to 4 sentences, bullets and tables. Easy to extract, easy to quote.
- Real numbers and real sources. A specific stat with a linked source beats a vague claim every time.
- Self-contained FAQs. Each answer stands on its own. No “see above.”
- Structured data. Article, FAQ and author schema.
“None of this is new to SEO,” De Feyter says. “It’s just good information architecture.” Engines are just rewarding it more directly now.
Your CMS should make these patterns the default, not something each writer remembers. A summary field that feeds the direct-answer opening. A reusable FAQ component that outputs its own schema. Author fields linked to real bios. On brightspot.com, we check each page in the Brightspot toolkit, a panel that shows SEO, GEO and accessibility guidance on the live page, one click from the asset in the CMS.
Why structure alone won’t get you cited
Format without substance is exactly what the spam updates are built to catch. In the session I called the substance “gold dust": real statistics, original research, customer testimonials and insight from subject matter experts. They’re the parts of your content strategy that AI can’t create, and they’re where your expertise, authority and trust come from.
You need to have a unique viewpoint. You need to be talking about and presenting something that is worth citing.
Asked for the one must-do thing in AEO, De Feyter didn’t mention a format.
“You need to have a unique viewpoint. You need to be talking about and presenting something that is worth citing,” he advises. If I write a generic article on just why SEO matters, but I don’t add why I think that matters and reference things that we have seen or proven within data, it’s empty and it’s useless. It’s not going to get cited.”
Teams that skip this usually have the modeling right and the substance wrong: perfectly typed fields wrapped around content that could be anyone’s.
Do you have to choose between depth and short answers?
One Brightspot Illuminate attendee raised a valid point: their user research says visitors want deeper content, not short blurbs and bullets. How do you match that with the short answers AI engines lift?
You do both, and the content model is how. Think of each page as modular. The parts that answer a question directly can be parsed and quoted. Around them sits the original research and editorial insight that makes you the authority. Typed fields let one page serve both readers, human and machine, without compromising either.
Engineer is the last phase of SAGE, and the one that turns a project into an operating model. The goal is a process that runs every week without someone pushing it along.
On brightspot.com, the moment a page publishes, we add it to Profound as a newly tracked prompt. The next weekly pull tells us whether the approach worked. We run the steps through Esca automations inside our marketing workflow, mixing automated steps with human-in-the-loop editorial review at every stage, so we gain efficiency without pumping out content for the sake of it.
Speed is what makes the weekly cadence worth it. In Google, a content change used to take a month to six weeks to show results. Now we see pages start to track against targeted prompts within a week or so. The flip side is volatility, so you need to keep watching.
Profound has a name for the default pattern: citation decay. Most cited pages spike, then fade. One brightspot.com article we published this summer ramped up in June, held a plateau, then climbed again, and that second climb lines up with optimizations we made along the way.
Governance doesn’t have to be heavy. Assign an owner per content type, not per department. Set a review cadence tied to how fast that content actually changes: quarterly for pricing pages, annually for evergreen explainers. Review your most-cited pages first, because they have the most to lose.
And don’t get stuck in analysis paralysis. Pick a topic and a prompt for the week, vet it, ship it, then check what happened and adjust.
We started this approach on brightspot.com about 5 months ago. Across our tracked topics and prompts, we now appear in about 21% of answers in our category, up from 13% in May. Our citation share more than doubled in the same window, and brightspot.com moved from the seventh to the second most cited domain in our category.
To be honest about it, that’s growth from a smaller base. Our category leader still shows up in over two-thirds of answers, so we have catching up to do. What the data does show is that the approach works, it repeats and it scales.
- Setup: Developers and SEO, with a content architect owning the content model
- Analyze: Content strategy and SEO, with subject matter experts consulted before any brief is approved
- Generate: Editorial and subject matter experts, working inside templates the CMS enforces
- Engineer: Content operations, with an owner per content type for reviews
This is an operating model, and most enterprise content organizations don’t have one yet. The teams that build it now are building the content infrastructure their AI strategy will depend on, whether that’s a chatbot, a RAG pipeline or an agent nobody has built yet. That’s the difference between a content organization that can say yes when leadership asks “can our content support this?” and one that’s still auditing HTML blobs to find out.
If you’re mapping this against your own content model, Brightspot’s SEO, AEO and GEO services team runs a GEO-readiness assessment against your actual content types, not a generic checklist.
Click here to learn more and find a time to talk about your content-strategy needs!
FAQs: How to prepare your content for AI engines
A CMS that separates content from presentation, models it as typed, structured data and exposes it through clean markup, schema and documented APIs. Chatbots, copilots and AI writing assistants are features. This is the architecture those features depend on.
Decide per bot type, not with a blanket rule. Cloudflare now separates search, agent and training bots. Googlebot handles both search and training, so blocking “AI” broadly can cost you search visibility.
Not yet for most sites. As of fall 2026, none of the major AI companies has said its answer engines use llms.txt, and studies show AI crawlers rarely request it. Where llms.txt and markdown versions have shown value is on documentation sites, where AI coding assistants read them and content is highly structured. Some Brightspot customers have built markdown output into their own instances.
Good writing still matters. But an AI engine has to find the facts before it can cite the writing around them. Typed fields make facts extractable. A rich-text blob makes the engine guess, and when it guesses, it often skips you for a competitor that didn’t make it guess.
In our experience, within a week or so. Google ranking changes used to take a month to 6 weeks. That speed is why a weekly test-and-adjust cycle pays off.
Contentful, Sanity, WordPress VIP, Drupal, Strapi, Sitecore and Brightspot all show up in this conversation. They differ in depth: how far each platform’s content modeling, API layer and editorial governance go, not just whether it checks an “AI-ready” box.
No. Headless architecture makes API-first delivery easier, which helps with setup. It says nothing about whether your content is modeled into typed fields, backed by real expertise or kept current.
Partway. You can add schema markup, restructure headings and build FAQ sections inside most platforms. What’s harder without deeper platform support is turning rich-text blobs into typed, queryable fields at scale. Teams on legacy platforms often find the cost of manual fixes adds up faster than a migration would have.
Audit one high-value content type first, the one that answers your most common customer question. Model it as typed fields, expose it cleanly, back it with a real expert and put it on a review cadence. Then repeat the pattern across your content model.
