In short: On April 20, 2026, agency Farotech publishes a figure worth reading twice: 42% of a client’s AI Overview rankings come from just three topic clusters. Concentration beats dispersion. This result confirms what we observe across the 1,300+ semantic cocoons deployed at Hi-Commerce since 2016: generative AI cites specialists, not generalists. Here’s the method to build 3 dominant clusters rather than 20 weak ones.
42%of AI Overview rankings concentrated on 3 clusters (Farotech, 04/2026)
1 pillar + 20typical structure of a cluster with high citation density
1,300+semantic cocoons deployed by Hi-Commerce since 2016
Why topic clusters are gaining new weight in AI Search
Topic clusters aren’t an invention of 2026. HubSpot popularized the concept in 2017 with the famous pillar + clusters architecture: one pillar page treating a broad subject, fifteen to twenty-five child pages covering each sub-topic with dense internal linking. The idea: demonstrate topical authority to Google rather than pile up isolated pages on keywords.
For almost a decade, this architecture served classic SEO. Honest gains. Rarely spectacular. Many agencies abandoned it in 2023 when the algorithm seemed to favor individual long-tail content and editorial pieces more.
Then ChatGPT Search, Perplexity, Google AI Overviews, and Gemini arrived in force. Rules shifted. Generative AI doesn’t scan your pages like a human browsing. It builds a vector representation of your domain, measures the semantic density around a given subject, then decides whether to cite you or not based on one simple criterion: are you a recognizable authority on this specific subject, or a generalist site that touches on it?
The AI reflex in 2026 looks like the human reflex when seeking an expert: you avoid the generalist firm handling a hundred subjects, you go to the specialist handling three. A language model reasons the same way. It prefers to cite a brand that published twenty-five coherent pages on GEO for e-commerce rather than a brand with three scattered pages on SEO, AI, digital marketing.
Farotech just measured it for one of their clients. We measure it ourselves, numbers in hand, across the 1,300 semantic cocoons deployed at Hi-Commerce over eight years.
The Farotech figure: 42% of AI Overviews on 3 clusters
On April 20, 2026, American agency Farotech publishes a sentence on X. It didn’t make enough noise.
42% of our client’s AI Overview rankings come from just three topic clusters. That’s the power of GEO (Generative Engine Optimization).
42%. Three clusters. Let’s pause there. In a standard audit, 42% of traffic concentrated on three silos is already sharp. Here we’re talking about citations in AI Overviews — the surface gradually replacing the ten blue links.
Strategic lesson. The Farotech client didn’t win by diluting across twenty subjects. They won by digging deep into three. Concentration beats dispersion, at a scale that changes every editorial prioritization.
Farotech names neither the client, nor the clusters, nor the vertical. But the message confirms elsewhere. A study shared by several GEO specialists shows that 47% of AI Overview citations come from pages ranked beyond position 5 in classic SEO, and 83% from pages outside the top 10. Translation: AI reads different signals. It reads topical authority.
And that’s exactly what a well-built topic cluster sends to a language model: dense web of linked pages, central pillar, recognizable author, entity network (Wikidata, LinkedIn, profiles) confirming expertise.
Across semantic cocoons deployed since 2016, I observe the same thing: clients holding three very dense cocoons outperform those with eighteen scattered ones on AI citations. The model doesn’t count pages. It measures coherence.
How to choose the 3 truly strategic clusters
Choosing three clusters means giving up twenty. The hard part. Most editorial plans I audit suffer from the same flaw: everything is priority, so nothing is. Here's the method I apply in workshops with Hi-Commerce clients.
Criterion 1: business alignment (20% of the choice)
The three clusters correspond to your three offerings that actually make money. Not your three favorite offerings, not your three high theoretical potential offerings. The three your accounting department ranks at the top. If a cluster maps to no margin line, it's out.
Criterion 2: search volume (30% of the choice)
A relevant cluster hosts at least 200 exploitable parent and child queries. Below that, you're not building a cluster, you're building an extended page. Tools for this mapping: Haloscan for the French market, DataForSEO for international volumes, and Google Search Console to identify what you already rank for.
Criterion 3: AI citation potential (50% of the choice)
The new criterion, the one that changes prioritization. A subject has strong AI citation potential if it meets three conditions:
It involves a comparison, trade-off, or method. AI Overviews trigger far more often on "how to choose X" than "X definition".
It isn't trivially answered by Wikipedia. AI prefers citing specialized sources when général knowledge isn't enough.
There's an active population of experts discussing it. If you're alone writing about the subject, AI lacks confidence signals. Ironic. Documented.
Concretely, for an e-commerce leather goods site, the three clusters aren't bags, wallets, belts. They're more like how to choose a full-grain leather work bag, care and patina of vegetable-tanned leather, sustainability and alternatives to industrial leather goods. Three angles combining volume, margin, and appeal to expertise.
Take thirty minutes. Write down the ten clusters you're targeting today. Rank them on these three criteria. Keep only the top three. The other seven go to phase two, not now.
Architecture of a cluster with high citation density in 2026
Once the three clusters are chosen, each is built following the same architecture, adapted for the AI Search era.
1. One exhaustive pillar page
The pillar covers the subject comprehensively. Not a blog post — a real reference page. 3,000 to 5,000 words. Definition, context, methods, common mistakes, use cases. Entry point for humans. Semantic anchor point for LLMs.
Every sub-article in the cluster points to the pillar. The pillar points to each sub-article. This double reciprocity transforms a collection of pages into a network.
2. Fifteen to twenty-five child pages
Each child page addresses one precise sub-intention. One query = one page. Don't stack two subjects on the same URL just because they're close. Mature cluster: minimum twenty pages. Across Hi-Commerce semantic cocoons, the average is around 22 pages per cocoon. Volume manageable over one to two months with a solid editorial process.
3. Dense and typed internal linking
High-performing cluster: each child page cites at least three other child pages from the same cluster, plus the pillar. Varied, natural anchors describing the destination. AI reads these links as a coherence grid. A subject holding together is a subject it can confidently cite.
4. Recognizable author signature
Every cluster page is signed by the same person, or by two or three well-identified authors. Bio, photo, LinkedIn link, ideally a Wikidata entry. AI weights citations by source reliability. Reliability flows through traceable author identity.
At Hi-Commerce, I created my Wikidata entry in March 2026. Measurable result in under thirty days: citations in ChatGPT and Perplexity rose on queries where my name appears in the answer.
5. External signals
An isolated cluster on its own domain plateaus fast. To break through: external mentions. Podcast interviews, guest articles, specialist press mentions, noted LinkedIn comments. LLMs weight out-of-domain mentions heavily in their authority calculation. A single niche podcast interview can unlock visibility for an entire cluster.
6. A structured FAQ page
Each cluster benefits from hosting an FAQ page with 10 to 15 questions, in FAQPage JSON-LD markup. Format especially well-absorbed by AI Overviews and conversational snippets. For hi-commerce.fr pages I deployed in March 2026, this markup preceded a sharp rise in Search Console impressions on question-form queries.
The bridge between semantic cocoons and topic clusters
Hi-Commerce readers know the term semantic cocoon. Laurent Bourrelly popularized it in France starting in 2012. I've industrialized it since 2016 with my team in Madagascar and Benin. The cocoon and topic cluster are cousins. Same intuition: group pages around a subject, dense linking, send Google a concentrated topical authority signal.
Three differences matter in the AI Search era.
Difference 1: cocoon thinks in hierarchy, cluster thinks in network
Classic cocoon structures as a tree: one parent page, child pages, grandchildren, linking respects hierarchy. The HubSpot topic cluster? More horizontal. One pillar, satellites, all at the same level. For AI, the hierarchy remains readable. The network is often easier to traverse vectorially. In the new cocoons I deploy since late 2025, I introduce more horizontal linking between sibling pages, in addition to vertical. Results climb.
Difference 2: cocoon was built for Google, cluster for AI
Original cocoons aimed at organic ranking on a query bundle. The 2026 cluster also aims at citations in generative answers. That changes two concrete things: we push harder on author signals. We multiply formats synthesizable by an LLM — comparison tables, FAQs, bullet lists, clear definitions at paragraph start.
I say it straight because it's Hi-Commerce's specific advantage: we've deployed over 1,300 semantic cocoons since 2016. Craft, industrial B2B, highly varied sectors. This volume gives an advantage few agencies can claim: we know which cocoon templates work on which site size, how many pages trigger a visibility plateau, which sectors respond better to AI citation.
For a client who's never structured content in clusters, the recommendation is always the same: start with three. Not ten. Not twenty. Three perfectly executed clusters always beat fifteen average ones, in SEO as in GEO. The Farotech data confirms it. Hi-Commerce internal data has confirmed it for three years.
Measuring topical dominance: the right metrics
Building three dense clusters is good. Measuring their actual dominance is better. Here are the five metrics I track every week on deployed cocoons.
1. AI Overview voice share per cluster
On your three clusters, how many queries trigger an AI Overview? How many of those AI Overviews cite your brand? This metric is calculated manually or with Profound, Otterly, or the SE Ranking AI Visibility extension. Goal: watch the share climb quarter after quarter on the three chosen clusters.
2. GSC impressions on cluster queries
Google Search Console remains valuable even in full AI Search era. Filter queries by cluster (URLs or pillar keywords). Watch impression trends. Regular growth signals Google recognizes you as authority on this subject. A plateau signals you need to densify linking or inject additional author signals.
3. Average citation depth
When AI cites you, is it for a pillar query (high value) or a very long-tail query (more peripheral)? Ideal: balanced distribution. If 100% of citations land on long-tail, your cluster lacks core authority. If everything flows back to pillar with no long-tail, internal linking is insufficient.
4. Spontaneous external mentions
How many times is your brand or primary author cited spontaneously off-site, per month, on the three cluster subjects? A performing cluster eventually generates spontaneous citations. That's the final signal that topical authority is recognized by the ecosystem. Simple search like "your-brand" -site:your-domain.com to track.
5. Click-through ratio: pillar vs. child pages
On a mature cluster, organic traffic distributes: the pillar captures 20 to 30% of clicks, child pages share the rest. If the pillar captures 80%, child pages are invisible — internal linking isn't working. If child pages capture everything and pillar nothing, the pillar doesn't answer real high-value intent. It needs rethinking.
Five metrics, one monthly dashboard, three clusters. Simple discipline fitting into one hour per month.
Concentration wins. Always.
The Farotech post from April 20, 2026 confirms what many practitioners observe without always formalizing it. The era of generative AI rewards topical concentration, not dispersion. Three well-built clusters can capture a dominant share of your AI citations. Twenty average clusters capture almost nothing.
For an e-commerce leader or SEO manager inheriting a sprawling editorial plan, the rational decision is rarely to add subjects. It's to remove them. Choose the three that matter for business. Build them thoroughly: exhaustive pillar, minimum twenty child pages, dense internal linking, identified author, external mentions, FAQPage markup, weekly measurement.
Across the 1,300+ semantic cocoons I've deployed since 2016, this pattern produces repeatable results. With clients preferring dispersion, results are always murkier. The trade-off is simple. It just takes editorial courage: courage to say no to seventeen interesting subjects to focus on three decisive ones.
The good news? Three clusters are achievable in one quarter with a clean process. No need to reorganize an entire catalog. Just identify the three subjects where you can be the source AI prefers to cite, and follow through.
Strategic cocoon and cluster audit
Torn between twenty average clusters and three dominant ones? In thirty minutes, I walk through your site with you, identify the three cocoons with strong business traction and AI citation potential, and deliver the precise linking map to deploy. No pitch, no slides — live audit, your pages on screen, concrete recommendations.
What's the minimum number of pages per topic cluster to expect AI citations?
Starting at 15 child pages around a solid pillar, we observe first effects. The comfortable threshold sits around 20 to 25 pages. Below 15, semantic density is insufficient for LLMs to perceive the domain as specialist on the subject. Across Hi-Commerce cocoons, the average is around 22 pages per cocoon — volume manageable over one to two months of editorial production.
Can a classic semantic cocoon and topic cluster architecture coexist?
Yes, and I've recommended it since late 2025. The cocoon brings hierarchical structure readable by Google, the cluster brings horizontal linking and author signals valued by LLMs. Concretely, you build a hierarchical cocoon then strengthen sibling-page linking and add GEO signals (FAQPage, author Wikidata, external mentions). Both approaches complément each other, they don't exclude each other.
Is the Farotech 42% figure reproducible across other sectors?
Farotech didn't specify their client's sector or exact methodology. The figure should be read as directional: topical concentration outperforms dispersion. Exact magnitude depends on market maturity in AI Search, number of direct competitors, and cluster execution quality. Across Hi-Commerce cocoons, we observe similar concentrations (30 to 50% of AI visibility on priority clusters) for clients respecting the full architecture.
Should I delete existing pages outside the 3 chosen clusters?
No, deletion is rarely the right move. Existing pages generate residual traffic and hold historical backlinks. The winning strategy is concentrating new editorial effort on the three priority clusters, letting the rest of the site run without active enrichment. In a second phase, you can merge old pages into an existing cluster if they map to a relevant sub-subject. Never mass-delete without precise audit.
How long before seeing the 3-cluster effect on AI citations?
First measurable signal typically arrives 6 to 10 weeks after complete cluster publication, provided internal linking is properly deployed and the author is identified with a Wikidata entry and active LinkedIn profile. Growth continues over 3 to 6 months. A mature cluster — 9 to 12 months old — continues gaining citations quarter after quarter.
Stéphane Jambu
SEO & AI Engineer
I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.
