Topic Clustering Is What LLMs Actually Read
Classical topic clustering optimised for Google’s pre-2023 ranking system. LLM retrieval clusters topics differently, weights them differently, and rewards a structure that most SEO teams have spent a decade unlearning. Where the old hub-and-spoke pattern packed 30 supporting URLs around a money page, retrieval-grounded models prefer a denser, flatter cluster where each leaf is independently citable and the hub is a discoverable index rather than a thin overview. Across thirteen anonymised audits, the brands gaining LLM citation share were not the ones with the most URLs per topic. They were the ones whose clusters resolved cleanly to entities the model could read in a single chunk. This piece sets out the new clustering rules, the diagnostic that surfaces a broken cluster, and the rebuild sequence we have run on BFSI, manufacturing, and marketplace properties.
Three Things About Clustering That Stopped Being True
First, that the hub should be a long-form pillar. Retrieval layers do not chunk a 4,500-word pillar into one piece. They chunk it into roughly 30 passages of 150 to 300 tokens each, score each passage independently, and select the best-scoring one for the query. A pillar that tries to answer 30 questions usually loses to a leaf that answers one question well. The hub should index, not exhaust.
Second, that internal linking from hub to spoke is the engine of cluster authority. The retrieval layer does not read internal links the way Google’s PageRank-derived signal reads them. Anchor text on internal links is a weak signal compared with passage-level proximity and entity co-occurrence. Linking matters for crawl efficiency and for human navigation. It is not what concentrates citation share.
Third, that semantic relatedness equals cluster cohesion. Two URLs can sit in the same semantic space and still belong in different clusters because the retrieval intent behind their queries differs. A page on “loan eligibility” and a page on “loan EMI calculator” are semantically adjacent and retrieval-wise far apart. Treating them as one cluster produces a hub that answers neither query well.
The New Definition of a Cluster
A cluster, in 2026, is a set of URLs that share an entity anchor, cover an exhaustive question space around that anchor, and resolve consistently across a chunking pipeline. The shared anchor is the entity, not the keyword. The exhaustive question space is observed, not theorised, from real retrieval prompts. And consistent resolution means that whichever leaf a retrieval layer extracts from, the answer it grounds is consistent with every other leaf in the cluster.
Consider an example from our manufacturing work. An industrial-materials exporter had 648 pages, 6,470 ranked keywords, and 77% of all organic traffic flowing through about 20 pages. Inside that 20-page set were three pages targeting Wall Cladding, all written at different times by different writers, all making subtly different claims about installation timelines and warranty terms. A retrieval pipeline reading any one of them landed on a citable answer. Reading two of them surfaced contradiction. The cluster was a cluster on paper and a contradiction in practice. The fix was not consolidation. It was a single canonical anchor page with a verifiable claim block, and two satellite pages restructured to cite the anchor rather than restate the same claims.
How to Tell a Cluster Is Broken
The Six Failure Signals
- Signal 1: Three or more URLs surface for the same query across a 100-prompt panel, and the citation engine quotes different ones across runs. The cluster has no canonical answer URL.
- Signal 2: Two leaf pages contradict each other on a factual claim (price, eligibility, timeline, specification). Retrieval grounding fails, citation rate drops.
- Signal 3: The hub URL outranks its leaves but is never quoted by an LLM. The hub is summary; the leaves carry the citation-grade detail.
- Signal 4: A leaf has high impressions but zero conversions and zero LLM citations. The leaf is targeting a phantom intent. Reassign or kill.
- Signal 5: The internal link map shows fewer than three links into a leaf from cluster siblings. The leaf is isolated and crawl-orphaned.
- Signal 6: Schema across the cluster uses three or more different types for the same content class. Markup contradiction tells the model the cluster is not a cluster.
Two signals in the same cluster is a rewrite. Four signals is a rebuild.
The Rebuild Sequence
Once a broken cluster has been identified, the rebuild follows a fixed sequence. Skipping order is the most common failure mode.
Pull a real retrieval prompt set first. Not keywords from a research tool. Actual prompts a buyer would type, gathered from a representative panel or from the brand’s own support tickets and sales conversations. The 100-prompt minimum is a working floor. Below that, the cluster is rebuilt against a theoretical question space rather than the observed one.
Cluster prompts before clustering pages. Group the prompts by intent class. Categorise each group as informational, transactional, comparative, definitional, or troubleshooting. Within each intent class, identify the entity anchor. This produces the target shape of the cluster, independent of what URLs currently exist.
Map current URLs onto target intent slots. A spreadsheet with one row per current URL and one column per target intent slot. Mark each URL as primary fit, partial fit, or no fit. Pages with no fit are candidates for deprecation. Pages with partial fit are candidates for rewrite. Pages with primary fit get to keep their slot but may still need restructuring for citation-grade passage extraction.
Specify the anchor page. One URL per cluster that holds the canonical answer to the cluster’s defining question. Restructured to lead with the answer in the first 60 to 100 words. All sibling pages cite the anchor rather than restate the same claim. The anchor inherits the strongest internal link signal and the cleanest schema markup.
Rewrite the leaves around independent citability. Each leaf answers one question, in plain HTML, in a passage a chunking pipeline can lift. Leaves do not depend on the hub for context. The hub depends on the leaves for evidence. This inverts the classical pyramid.
Deprecate aggressively. Pages with no fit, low impressions, and no citation surface should be 301-redirected to the most appropriate anchor and removed from the sitemap. The 25K-page NBFC audit surfaced 3,620 sitemap waste URLs that were diluting cluster signal across every category. Removing them improved retrieval recall on the surviving URLs measurably within six weeks.
What a Working Cluster Looks Like
The shape is flat. One anchor page, four to eight leaves, each leaf covering a distinct intent within the same entity. The anchor links out to all leaves. Each leaf links to the anchor and to two or three sibling leaves where the cross-reference is genuinely useful. No leaf links to every other leaf. The internal link map should be readable as a hub-and-rim, not as a complete graph.
The depth is single-click. Every leaf is one click from the anchor. No nested sub-clusters. If a topic genuinely splits, it becomes its own cluster with its own anchor. The marketplace BRD we built mapped 21 distinct URL axes across a coworking property and produced a target footprint of 12,000 to 18,000 URLs. The discipline that kept those URLs navigable was strict single-click depth at the leaf level, with the entity hierarchy carrying the breadth.
The freshness is uneven, by design. The anchor is revisited quarterly. Time-sensitive leaves are revisited monthly. Definitional leaves are revisited yearly. The dateModified on each URL reflects the actual review event, not a CMS-wide republish. Perplexity and ChatGPT Search both read this signal directly. AI Overview reads it implicitly through Google’s freshness layer.
Measuring the Lift
Cluster work is measurable only against a real baseline. Three numbers matter and should be captured before any rebuild begins.
First, citation rate per engine on a fixed 100-prompt panel. Run the panel against ChatGPT, Claude, AI Overview, AI Mode, and Perplexity. Record which URL is cited and how often. The 25K-page NBFC case landed at 8% on ChatGPT, 15.6% on AI Overview, and 19% on AI Mode at baseline. The engine spread is part of the diagnostic.
Second, internal coherence rate. Pick ten prompts the cluster targets, run each through three engines, and check whether the answer the model gives is consistent across engines. Inconsistency is a sign of cluster-internal contradiction, not engine variance.
Third, conversion-adjusted impression share. Impressions alone do not tell you the cluster is working. Conversions per cited impression is the honest signal. The manufacturing work surfaced one product line with 74.7% conversion that was being out-trafficked by a 21.1%-conversion line. The cluster fix prioritised protecting the high-converting line first, traffic-shifting later.
Our AI visibility audit establishes these baselines and the rebuild plan against them. Sites with crawl or canonical issues that are blocking cluster signal benefit from a parallel technical SEO audit. Cluster patterns by industry are documented in BFSI growth engineering and manufacturing growth engineering.
Practitioner Takeaway
- Run a 100-prompt panel across five engines on your top cluster this month. Record citation rate and the specific URLs cited. The variance reveals the cluster’s coherence.
- List your URLs against intent slots, not against keywords. Use real prompts grouped by intent class. Mark each URL as primary fit, partial fit, or no fit.
- Pick the anchor page and rewrite its opening 60 to 100 words. Named entity, claim, number or date. Plain HTML. This is the passage retrieval will lift.
- Deprecate the no-fit URLs. 301 to the anchor. Remove from sitemap. Audit the cluster’s signal six weeks later.
- Set differential refresh cadences. Anchor quarterly. Time-sensitive leaves monthly. Definitional leaves yearly. Schema dateModified reflects real review.
Frequently Asked Questions
How is LLM clustering different from Google’s topical authority model?
Topical authority is a property-level signal that Google’s ranking system uses to weight an entire domain on a subject. LLM clustering is a passage-level retrieval decision that happens at query time. A site can carry strong topical authority and still produce a cluster where the model cannot extract a clean citation, because the two systems measure different things. Both matter. They require different treatments.
Should I delete pages that have impressions but no LLM citations?
Not automatically. Impressions still produce branded recall, classical click traffic, and ranking signal. The deprecation candidates are pages with low impressions, no conversions, and no LLM citations. That combination is where the editorial cost outweighs the surface value.
How many leaves should a cluster have?
Four to eight is the working range for B2B brands. Below four, the anchor is doing too much work alone. Above eight, the cluster has usually subdivided and would be cleaner as two clusters with separate anchors. The number is a guideline, not a rule.
Does this approach work for ecommerce category pages?
Yes, with one adaptation. Ecommerce anchor pages double as transactional surfaces, so the citation-grade answer block sits alongside product listings rather than replacing them. The first 60 to 100 words of the category description carry the answer. Product listings carry the conversion job. Both surfaces work in parallel.
Can a cluster work without a Wikidata-anchored entity?
It can work, but it underperforms. Entity anchoring concentrates the model’s confidence on a single canonical reference. Without it, the cluster competes against generic semantic space and earns lower citation share even when the content quality is high.
If your category cluster has stalled despite a strong content programme, the diagnostic is usually cluster-internal contradiction, not content scarcity. Request the audit that surfaces the six failure signals on your top three clusters and specifies the rebuild order.