Your Category Pages Are Thin: A Hub Depth Diagnostic

Thin hub pages — category, collection, and topic-index URLs that offer a heading, a grid of links, and little else — are the most common structural weakness we find in site-architecture audits. They sit at the exact URLs a site nominates as canonical for its most valuable topics, yet many of them answer no question a buyer actually has.

Here is what makes this topic harder than most SEO writing admits: the evidence is genuinely mixed. A 300-page study of UK e-commerce category pages found that pages ranking #1 in classic search average just 310 words of unique content — thin hubs demonstrably win rankings. Meanwhile, two large AI-citation studies and one quality-signal framework suggest that generic, low-signal pages are cited by AI answer engines at lower rates. Neither body of research directly tests the question this post cares about, and we will say so plainly rather than pretend otherwise.

This guide covers what the research actually measures, why hubs go thin at scale, a six-signal diagnostic you can run on your own architecture this week, and a repair pattern — rebuilding hubs as decision documents — that improves the page for buyers regardless of how the AI-citation question eventually resolves.

Before any diagnostic, the claim inventory. A popular line in AEO and GEO content goes something like: AI engines cite your category page because it is the canonical URL for the topic, find nothing quotable there, and your brand loses the citation. It is a tidy mechanism. It is also, as far as we can determine, unproven — none of the citation studies we reviewed for this piece segments AI citations by site-architecture depth, hub versus leaf.

What the located research does support is narrower. AI answer engines cite certain page formats — product pages, comparison pages, articles, and category pages as a template class — at rates that vary by funnel stage and industry. Page quality signals showed the strongest associations with citation among the on-page pillars the one framework study we located audited. And faceted navigation is, per Google’s own crawling team, the most common source of overcrawl issues site owners report to them. Those three findings are adjacent to the hub question without answering it directly.

Why write the post anyway? Because the practical problem is real whether or not the citation mechanism is ever measured. A hub page that answers nothing is a weak asset in every channel it appears — classic SERPs where a competitor’s richer page wins the long-click, AI answers where quotable substance appears to matter, and direct traffic where a buyer bounces off a wall of tiles. The diagnostic and the repair earn their place on buyer value alone.

Hub thinness is rarely a content decision. It is an architectural byproduct. Most platforms generate category pages from a template: H1, optional intro field nobody fills in, product or article grid, pagination. Every new category inherits the template, and the template has no opinion about substance. Multiply by faceted navigation — size, color, price, brand filters — and each filter combination generally creates its own URL showing largely the same inventory, which is how large sites accumulate thousands of technically unique but sparse category-adjacent pages.

This is not a fringe concern. In a December 2024 Search Central post on faceted navigation, Google’s Gary Illyes wrote that faceted navigation is “by far the most common source of overcrawl issues site owners report to us” — and Google’s recommended fix is a binary choice: block the filter URLs outright, or optimize them for crawling with consistent parameter order, canonical tags, and 404s on empty result sets. Google is watching this URL class closely; that is documented. What it means for AI citation is not, and we return to that distinction in section 06.

One vocabulary note worth getting right: Google’s current spam policies do not use the phrase “thin content” at all — that is industry shorthand, not Google’s term. The closest defined violations are doorway abuse, described as pages “created to rank for specific, similar search queries” that “lead users to intermediate pages that are not as useful as the final destination,” and thin affiliation, where product descriptions are “copied directly from the original merchant without any original content or added value.” A zero-content category page usually violates neither policy. The problem with thin hubs is almost never a penalty — it is opportunity cost.

Any honest version of this argument has to start with data that cuts against it. In a study by Digitaloft, 300 UK e-commerce category pages ranking #1 for commercial-intent keywords (collected via Semrush from a 524-keyword list, each page manually verified) averaged just 310 words of unique content — excluding product-grid text, buttons, faceted navigation, and the H1. The distribution is even more striking than the average.

Two caveats keep this honest. First, the author’s own framing: “The aim wasn’t to prove that word count directly affects rankings, but to see what top-ranking pages are actually doing in practice.” It is a correlational snapshot of one country and vertical mix, not a causal claim. Second, the word-count benchmarks in this space disagree: a separate practitioner claim — which we could not verify against a disclosed methodology — puts top-ranking category pages at 500 to 700 words, roughly double the Digitaloft figure. We present the sourced number, note the disagreement, and deliberately settle on neither as a target.

The interpretive point matters more than either number: many hubs are thin by design and still perform in classic search — most of the 29 zero-content pages in the sample, 18 of them, were household-name brands whose authority, in the study author’s reading, likely explains the ranking. So the case for rebuilding a hub cannot rest on “thin pages lose rankings.” For plenty of strong brands, they demonstrably do not. The case has to rest on how thin hubs fail differently — for buyers who land and find nothing, and plausibly for AI answers that need something quotable. That distinction shapes everything that follows.

Two independent datasets dominate the AI-citation evidence base at the time of writing, and they must not be blended. The first, reported by Search Engine Journal from research by XFunnel, tracked 768,000 citations over 12 weeks across ChatGPT, Google AI Overviews, and Perplexity. Within that dataset, product content captured 46 to 70 percent of citations depending on funnel stage — rising past 70 percent at bottom-of-funnel queries — while blog content took only 3 to 6 percent and PR materials under 2 percent. B2B and B2C mixes differed sharply: product pages earned 56 percent of B2B citations in the sample versus 35 percent for B2C.

The second, a DeltaV Digital study published in July 2026, analyzed 25,337 citations from 21,075 AI-engine responses across five engines, narrowed to the 30 most-cited URLs per brand across 8 brands — a 240-URL sample. In that sample, articles took 23.7 percent of citations, listicles 19.6 percent, and product pages 16.3 percent — but comparison pages, at only a 4.1 percent share, posted the highest citation rate per retrieval at 1.87, about 45 percent above the portfolio average. Small footprint, high intensity: the page format built around helping someone choose gets cited hardest when it is retrieved at all.

The table below puts the two side by side — not to reconcile them, but to show why they cannot be reconciled, and what each one does and does not tell you about hubs.

DeltaV’s self-disclosed limitations deserve passing on: the top-30-per-brand cut captures the head of a power-law distribution but not the long tail, the prompt sets skew commercially relevant, page classifications are automated and imperfect at the margins, and 8 brands in 8 industries is a directional sample, not a census. Directional is still useful — both studies agree that generic, undifferentiated content under-cites relative to specific, decision-useful formats. One DeltaV row is the closest adjacent number this post has: pages its classifier labelled category pages took 5.2 percent of the 25,337 citations at 1.06 citations per retrieval, the lowest rate of the ten page types it reports. We use that as suggestive, not decisive — it is a template label applied by an automated classifier DeltaV itself calls imperfect at the margins, and it says nothing about where in your architecture that content lives.

The closest thing to a mechanism-level account comes from the GEO-16 framework paper on arXiv, which audited 1,702 citations across 1,100 unique URLs generated by 70 product-intent prompts, spanning three engines: Brave Summary, Google AI Overviews, and Perplexity. Its headline finding: among the 16 on-page pillars it audits, those covering metadata and freshness, semantic HTML, and structured data showed the strongest associations with citation. A logistic regression put the citation odds ratio for overall quality score at 4.2, with a 95 percent confidence interval of 3.1 to 5.7. The paper describes itself as observational, and every pillar in it is an on-page quality signal — page length, URL type, and hierarchy position are not among the variables it models.

Two more findings and one large hedge. Pages cited by more than one engine — 134 of the 1,100 audited URLs — scored 71 percent higher on quality than single-engine citations, suggesting that quality is what generalizes across engines while individual engines tolerate individual weaknesses. The hedge: GEO-16’s sample is narrow. English-language B2B SaaS pages only, 70 prompts, three engines — no ChatGPT, no Gemini — and it predates the 2026 model generations. Treat its odds ratio and thresholds as directional, not universal. What it contributes to the hub question is the direction of the gradient: pages weak on metadata, structure, and entity coverage under-cited even when topically relevant. A template-generated hub page is, almost by definition, weak on exactly those signals.

What follows is our inference, not a measured finding — we want that unmistakable. Built from the adjacent evidence above, the argument runs in four steps:

• Step 1 — hubs are the nominated topic URLs. Your internal linking, breadcrumbs, and canonical structure all tell crawlers the category page is the authoritative URL for the category-level topic. That is what site architecture is for, and Google’s crawling guidance confirms this URL class gets heavy crawler attention — heavy enough that faceted navigation, the URL family category pages spawn, is what Google calls by far the most common source of overcrawl issues site owners report to it.
• Step 2 — retrieval and citation are separate hurdles. The DeltaV data distinguishes citation share from citations per retrieval — comparison pages were retrieved rarely but cited at the highest rate (1.87 per retrieval in its 240-URL sample) when they were. Being fetched is not being quoted.
• Step 3 — quotable substance predicts the quote. GEO-16 found quality signals carried a 4.2× citation odds ratio within its sample, and both citation studies found generic content under-cited relative to specific, decision-useful formats. The nearest directly relevant number: the category-page class in DeltaV’s sample posted the lowest citations-per-retrieval of its ten page types, 1.06. A grid of product tiles offers an answer engine nothing to extract.
• Step 4 — the inference. If a hub is the URL your architecture nominates for a category query, and it contains nothing quotable, the plausible outcomes are that the engine cites a competitor’s richer page, cites one of your leaves out of context, or skips the citation entirely. We consider that mechanism likely. We cannot show you a study that measures it — and any competing article that claims one exists is, as of our review, ahead of the evidence.

The way to convert this from argument to observation on your own site is measurement: track which of your URLs AI engines already cite and watch how that changes after a hub rebuild. Our brand citation audit checklist covers exactly that instrumentation. One honest n=1 beats a fabricated n=768,000.

This table is ours. No located source combines these signals into one self-audit: existing coverage measures word count alone, or citation-quality signals alone, never a working diagnostic aimed at a site owner deciding which hubs to rebuild. The verdict thresholds are our recommended starting heuristics — explicitly not sourced numeric cutoffs, because as section 03 showed, the published word-count benchmarks disagree with each other by roughly 2×. One signal, text-to-HTML ratio, is included deliberately as a cautionary example of a metric to drop.

The entity-coverage signal deserves one expansion, because it is the best depth measure that is not word count. The heuristic — covered in entity-based SEO explainers and implemented by commercial tools — checks entities per thousand words and gap-analyzes a page against the entity set its topic implies. A running-shoes hub whose leaves collectively cover cushioning classes, drop heights, surface types, and brand lines, while the hub itself names none of them, has a measurable depth gap no word counter sees.

Running this across a large site by hand is tedious, which is why we fold it into automated audits — the deep-research audit prompt method we published alongside this piece surfaces thin hubs as a standard finding category. And once the diagnostic produces a list, resist fixing it top-to-bottom: our refresh prioritization matrix is the tool for deciding which hubs earn the rebuild first — commercial value times depth gap, not alphabetical order.

The repair is not “add 500 words.” Word-count targets are exactly the trap section 03 warned about. The repair is a structural template we call the decision document: a hub that answers, above or alongside its grid, the four questions a buyer at category level actually has — what the options are, how to choose between them, what each costs, and who each one suits. That template is informed by the one strong per-page signal in the citation research: comparison content’s 1.87 citations per retrieval in the DeltaV sample — and it improves the page for human buyers regardless of what any engine does with it. Not every hub needs the same treatment, so match the repair to the diagnostic verdict:

Three build notes. First, the cost question is the one most brands skip and the one with the clearest bottom-funnel case — we made the full argument in our piece on cost transparency pages, and a hub that includes honest price ranges inherits it. Second, a rebuilt hub only pays off inside a coherent structure: the pillar-and-cluster build method is the constructive counterpart to this post’s diagnostic — this piece finds the weak hubs, that one builds the architecture they should sit in. Third, do not neglect the usability layer: Nielsen Norman Group maintains a long-running body of paywalled e-commerce UX research on category and listing pages — two report volumes with 62 and 139 design guidelines respectively — which we cite here as evidence that category-page usability is a mature research field, not as a source of specific guidelines we have not read.

Looking forward, our expectation — clearly labeled as projection — is that someone will publish the missing study within the next research cycle: citation datasets already carry URL paths, and segmenting by architecture depth is one classifier away. If that study lands and shows hub URLs under-cited relative to their retrieval rates, the diagnostic above becomes more urgent. If it shows the opposite, you will have rebuilt your most valuable category pages into genuinely useful buyer documents — which is the rare SEO bet that pays out even when the thesis is wrong. This is the standard we hold agentic SEO engagements to: recommendations that survive the evidence changing.