At the end of September, my colleague Karla Ivancok and I (Diana Skof, Head of Marketing at Improove) teamed up with LSZ Future Connections to host the webinar "Content-Cluster statt Content-Chaos" (Content Clusters Instead of Content Chaos). The webinar was held in German. For a full hour, we tackled one question: why do so many brands simply not show up in ChatGPT, Perplexity and Google's AI Overviews? Since 60 minutes didn't leave room for every detail, we're recapping the key takeaways here: what AI systems do structurally differently from classic search engines, how a citable cluster is built, how you measure its success, and the most common questions from the Q&A.
The key points up front
- AI systems don't evaluate whole pages, they pull individual passages (query fan-out). A page that answers several closely related questions offers more points of connection than lots of thin subpages.
- Authority is built per topic, not globally. A well-known brand is still unknown to an AI system in a new topic area, and only 38 percent of AI citations come from the top 10 of classic search at all.
- You measure success at the level of the cluster as a whole, not the individual URL, following a four-stage logic (mentioned, cited, recommended and, looking ahead, the system takes action), and as a trend over weeks rather than a snapshot.
Why classic search logic is no longer enough in AI search
Around 40 percent of Google searches now show an AI Overview. Where one appears, the click-through rate on organic results drops by up to 61 percent, according to Seer Interactive. But the real problem isn't the lost click, it's whether your brand appears in the answer at all: only 38 percent of pages cited in AI Overviews still rank in the top 10 (Ahrefs), down from 76 percent in July 2025. Citations and rankings are increasingly decoupling, and what decides this is the overall structure of a topic area, not the ranking of individual pages.
AI systems pull passages, not pages
AI systems don't evaluate pages. They pull passages from various sources and assemble an answer from them. This technique is called query fan-out: a query is broken down into many sub-questions (price, team size, alternatives, migration), and the best available passage is sought for each one.
A page that answers only one sub-question offers one point of connection. A page that answers five of them offers five.
Authority is topical, not global
“The AI already knows us, we're a well-known brand” is unfortunately not automatically true. A mobile provider that has been publishing about plans and roaming for years gets mentioned for those topics. If it launches a new product without a supporting topic environment, the same brand is completely unknown to AI systems in that area. Brand awareness doesn't carry over. Entity Authority has to be built anew for each topic, and clusters are exactly the tool for that.
How visible is your brand in AI search already?
We analyze your visibility in ChatGPT, Perplexity, Gemini, Copilot and AI Overviews and show you where entities, structure and content readiness are still missing.
Free initial consultationThe most expensive mistake: more articles instead of better structure
The obvious reaction is to write more. That usually doesn't help, for three reasons: more of the same adds no information gain, because the model already knows it from its training data. The entity gets diluted when twenty thin posts cover the same topic with shifting terminology. And orphaned pages without links contribute little to topical coverage.
Even the classic cluster model is no longer quite up to date
One pillar page plus lots of narrow subpages, each targeting a keyword: for classic search, that was spot on. For AI search, something else works better.
Classic vs. AI-ready cluster model
| Classic cluster model | AI-ready cluster model 2026 |
|---|---|
| Pillar page plus lots of narrow subpages | Fewer but deeper pages |
| Each subpage targets one keyword | One page answers several closely related questions |
| Clustering by keyword overlap | Clustering by meaning |
The rule of thumb: questions that are close in meaning belong on one page. Only when the meaning clearly diverges is a separate URL worth it. That doesn't mean cramming everything into one monster page, but slicing by meaning instead of by keyword.
An example: the same knowledge, two structures
Take the topic of time tracking: five separate posts (What is time tracking, legal requirements in Austria, Excel template, app comparison, costs) mean five URLs, five half-answers and no shared structure. The alternative: one page, “Time tracking for companies in Austria”, with clear sections on legal requirements, methods, costs, common pitfalls and an FAQ, linking to the service page. Because a cluster isn't made up of guides alone: product and service pages are part of it too.
The blueprint in five steps
- Core topic: choose a field with real substance, close to your core service.
- Collect questions: gather sub-questions from real sources such as sales and support.
- Group: bundle related questions, separate distant ones.
- Build pages: one page per group, with sub-questions as visible subheadings.
- Connect: link to the service page using the exact product or service name as anchor text, not “click here to learn more”.
Step five is the one most often forgotten, and it has the best effort-to-impact ratio of the whole topic. It connects your own entity with the problem the text is about, and it helps decide what gets mentioned when someone asks a follow-up question in the chat. You'll find more on the content strategy behind it in our guide to answer-first content.
Every paragraph has to work on its own
AI systems pick up individual sections, not whole pages. A section that doesn't make sense without the context before it (“As described above, this makes sense”) won't be used, even if its content is good. This results in four rules:
- Subheadings as a question or a clear statement
- Answer first, details after
- One topic per section
- Spell out the topic instead of writing “it” or “this”
“A well-known brand is still a nobody to an AI in a new topic area. Authority doesn't carry over, you have to earn it anew for each topic.”
Diana Skof, Head of Marketing at Improove
Your own data beats good wording
Three things decide why a system should cite you in particular:
- Your own numbers: If you're the only source for a figure, you'll be cited for that figure.
- Visible accountability: authors with profiles, clear source citations.
- A clear position: well-reasoned disagreement with industry consensus gets cited, pure consensus doesn't.
Maintenance over new output: the most overlooked insight
75 percent of the pages cited by AI systems were updated in the past year, but only 42 percent were published in the past year. Freshness comes from maintenance, not from new production. For around 86.5 percent of prompts, there's a fixed core of one to five domains that get cited reliably, and one-off measurements are worthless: on average, two ChatGPT answers to the same question share only about 21 percent of the same sources.
That's why a cluster needs ongoing care, not just a launch: one person in charge, a review cycle based on citation data instead of the calendar, and new content that gets built into the existing cluster instead of springing up next to it.
Anchor cluster logic in your own editorial team
In the AI Editorial Workshop, you practice cluster building and question research on your own topic and set up governance that lasts beyond the launch.
More about the workshopHow you measure success: at the cluster level, not the individual URL
Because two answers to the same question often share only a fraction of their sources, you measure whether the cluster as a whole is visible, along four stages: is the brand mentioned, is the source cited, is it actively recommended and, looking ahead, does a system take an action.
An honest note on the fourth stage: agentic action as a metric isn't a standard yet. Even Amazon and Shopify are still working on the technical foundations for it. This stage belongs in your reporting as an outlook, not as a metric you can already rely on today.
Prompt tracking instead of rank tracking
20 to 50 core prompts, tested across several models, evaluated as a trend over weeks instead of a snapshot. Prompt tracking doesn't measure positions, it measures whether and how a model understands your brand. We explain the specific metrics for this, citation frequency, share of voice and prompt coverage, in detail in our article on GEO KPIs. The free way to get started without your own tool: since August 31, 2026, Google Search Console shows AI impressions directly in the interface. More on this in our analysis of Google's guide to AI search.
Where clusters fail in practice
Six mistakes that keep coming up in projects:
- Orphaned pages that nobody links to
- Shifting terms for the same thing
- Contradictions between pages
- Generic mass without substance
- No one in charge
- Technical blockers that crawlers can't read
Five of these six causes are organizational problems, not technical ones. You don't need a new tool for them, you need ownership and discipline.
Frequently asked questions about content clusters and GEO
How long does it take to build a cluster?
The build itself (research, structure, copy) realistically takes six to ten weeks for eight to twelve pages. Until you see an impact in AI answers, it's more like three to six months. Anyone who promises faster results is selling you something.
How long does a cluster page need to be? Is there a word count?
No. Neither the old 300-word rule nor the ten-thousand-word skyscraper has a solid foundation. The only useful test: is the question answered without filler? Then the page is long enough.
Does this also work for a small website with little content?
Yes, even better than people often think. Only 38 percent of AI citations come from the top 10, so a top 10 ranking isn't a prerequisite. A cleanly built cluster in a narrow topic area is more effective than two hundred scattered pages.
Is top-of-funnel content now redundant because AI answers it away?
Only half. The click on “What actually is XY” goes away, but the mention still has an effect, and that's exactly where vendor selection takes shape. You can cut pure definition pages without substance of their own, but keep content with its own position and its own numbers.
Does the cluster approach also apply to e-commerce and product pages?
In principle yes, with a shift. For product decisions, AI systems rely heavily on comparison and marketplace content, which usually isn't on your own domain. Part of the work is therefore helping to influence content on other sites. Your own clusters then mainly cover the advisory and decision questions around it, plus the connection to category and product pages.
How Improove helps with GEO content strategy
For us, content clusters for AI search aren't an isolated project but part of an integrated system: we first check how a company is currently perceived in ChatGPT, Perplexity, Gemini, Copilot and AI Overviews (SEO/GEO Check), and based on that, we develop a prioritized action list rather than a collection of observations. If you want to get started right away, the AI Editorial Workshop is the right place. Both paths are part of our SEO/GEO Management, which doesn't treat SEO and GEO as separate disciplines but as one system that builds on itself. We keep collecting more answers about SEO and GEO in our big FAQ.
