Blog · Diagnosis

The small-site authority problem

Published 31 August 2026 · Updated 12 October 2026 · 8 min read

The small-site authority problem
Photograph: Ryan Stone / unsplash

The short answer: Small sites rarely lose AI answers to a hidden authority score. They lose because an answer has room for roughly three sources, because almost nothing outside your own domain repeats your claim, and because sparse crawling can leave an engine reading an older copy of a page you have since corrected. Each of those three has a different fix, and you can work out which one is biting you in an afternoon.

The feeling is familiar. You ask ChatGPT or Perplexity who the good vendors are in your category, and the answer names four companies, none of them you. Then you ask about your own product by name and the engine describes it fluently and mostly correctly. That gap between "knows us when asked" and "never volunteers us" is the whole problem, and it is diagnostic.

Google documents no authority dial

Start by deleting the mental model, at least for Google. Google's documentation on AI features states that there are no additional technical requirements for appearing in AI Overviews or AI Mode, and that you do not need new machine-readable files, AI text files or special schema.org markup. A page is eligible if it is indexed and can be shown in Google Search with a snippet. That is a low bar, and nothing in it scales with the size of a site.

That is Google's statement about Google. OpenAI, Perplexity and Anthropic publish crawler and access documentation rather than eligibility criteria, so for those engines we cannot show you an equivalent statement either way. Read the rest of this post as a description of retrieval behaviour we can observe, not as a disclosure from an engine.

So when a nine-page site loses to a nine-thousand-page site, the cause is downstream of eligibility. It sits in retrieval: which documents the engine pulls for that question, and what it finds when it reads them. Why AI never mentions your brand walks the full chain. This post is about the part of the chain that punishes being small.

Three slots and sixty-seven words

Pew Research Center analysed 68,879 Google searches from the browsing data of 900 US adults in March 2025 and published the results on 22 July 2025. Two numbers from it matter here. Eighty-eight per cent of the AI summaries in that dataset cited three or more sources, and the median summary ran 67 words. Both figures describe Google AI summaries in that one month and that one panel. We have not found an equivalent published dataset for a chat assistant's answer, so we cannot tell you whether the shape carries across.

Think about what that implies for a category question. The answer has space for a handful of citations and about four sentences of prose. The engine is not ranking fifty candidates and showing you the top ten. It is picking two or three documents that each cover a lot of ground, and a well-maintained industry roundup covers more ground per document than your product page ever will.

This is the first failure mode, and it is structural rather than punitive. Roundups, comparison pages and directories win the broad query because they answer the broad query. Your site answers one narrow version of it.

Query width is the lever you actually control

The same Pew analysis found that 8 per cent of one-word and two-word searches in that March 2025 sample produced an AI summary, against 53 per cent of searches of ten words or more. On Google, in that dataset, longer and more specific questions were far more likely to produce a summary at all.

They also pull from a thinner candidate pool. When somebody asks a fifteen-word question with a constraint in it, an industry, a company size, a regulation, a country, the set of documents that genuinely address it may be very small. Small sites can own that set. The same site has no realistic path to the two-word version.

Here is the check. Take one buyer question and write it at three widths. Three words, eight words, fifteen words with a real constraint. Run all three on two engines, twice each, and record whether you were named. In the small-site audits we run, there is usually a width above which the site appears and below which it does not. We have not counted that split across a defined sample, so take it as a pattern worth testing on your own site rather than a measured rate. Wherever your own threshold sits, it is your current reach, and moving it down by one width band is a quarter of work, not a week.

The named-versus-unnamed test

Run your category question with no brand names in it, then run a question that names you. Compare what comes back.

If the engine describes you accurately when named and omits you when not, your pages are fine and your corroboration is thin. Nothing the engine already trusts repeats your claim, so there is no reason for your name to surface unprompted. If the engine gets facts wrong even when you are named, the problem is on your own site, and Can AI crawlers actually read your site is the place to start.

In the audits we run, the first test fails more often than the second. We have not kept a count, so that is an impression from our own case load and not a published rate. Where it holds, the pages are clean and the site is simply the only place we can find that says what it says.

Crawl frequency quietly makes it worse

Cloudflare published crawler data on 1 July 2025 covering May 2024 to May 2025 across its own network. GPTBot requests grew 305 per cent over that year and its share of AI and search crawler traffic rose from 2.2 per cent to 7.7 per cent. ChatGPT-User, the agent that fetches a page in response to a live question, grew 2,825 per cent.

Those are aggregate numbers from one network, and they say nothing about how often any particular site is crawled. Cloudflare does not publish a breakdown by site size, and we have not measured one. What we can say is narrower: on the small sites whose logs we have read during audits, revisits from AI crawlers are sparse enough that a page corrected in one month can plausibly still be answered from an older copy weeks later. Your server logs show which bot fetched which URL and when. They do not show which copy an answer was built from, so this stays an inference.

Two things help. Keep the answer to each important question in the first eighty words of a page, so a shallow fetch still gets it. And when you change a material fact, pricing, a product name, a supported region, change it everywhere on your own domain the same day. When we find two live pages on one domain contradicting each other, the answers we get back tend to hedge or pick the older figure, which is the outcome you are trying to avoid.

What to do with the diagnosis

The three failure modes need different money.

Slot scarcity is fixed by getting inside the documents that already win, which means being listed, reviewed and quoted on the pages the engine cites for your category. You do not own those pages. That is the slow work.

Query width is fixed on your own site, this month, by writing the specific pages that answer the constrained questions nobody else has bothered with. It is the only one of the three a small team fully controls.

Crawl staleness is fixed by consistency and by shipping fewer, more durable pages rather than a weekly post nobody fetches. A self-audit checklist for AI visibility covers the mechanical part of this, and What is answer engine optimisation sets out the wider frame. If you want to see how we write findings like these up with the evidence attached, the sample audit report shows the shape.

Honest limits

This post does not establish that the three mechanisms are the causes. We do not know how an engine weights a domain internally, and none of them publish it. The mechanisms are inferred from Google's own documentation, from the Pew and Cloudflare datasets cited below, and from repeated prompt runs in our own audits, which are small samples of client sites rather than a census of anything.

Two specific soft spots. The Pew figures describe Google AI summaries in a single month on a single US panel, and we have extended that picture to chat assistants without a dataset that covers them. The crawl-staleness argument rests on Cloudflare's network aggregates plus the logs of the small sites we happen to have read, and we cannot see which copy of a page any answer actually used, so the link between sparse crawling and a stale answer is reasoning rather than measurement.

We also cannot give you a reliable time to effect. Corroboration arrives when a third party decides to publish, and that is outside anyone's control.

FAQ

Does domain authority matter for AI citations? The engine documentation we have been able to open names no third-party authority score as an input, and Google's AI features page lists no technical requirement beyond indexing and snippet eligibility. That is an absence in the public documentation, not proof that nothing similar exists inside a ranking system. The thing those metrics approximate, how many sources an engine already trusts repeat your claim, is the part worth working on.

Can a seven-page site get cited at all? Yes, on narrow questions. Google's AI features documentation sets out no size requirement, and specific multi-constraint questions have small candidate pools. Broad category questions are a different matter and in our audits usually need third-party coverage.

Should I publish more pages to look bigger? The documentation we have read names page count nowhere, and we have not seen volume on its own move an answer in the audits we run. A page that answers a real question completely gives an engine something to quote. Ten that circle it do not.

How do I know whether it is my site or my reputation? Run the named-versus-unnamed test above. Accurate when named and absent when not points at corroboration. Wrong when named points at your own pages.

Sources

Related reading

Why AI never mentions your brandA self-audit checklist for AI visibilityCan AI crawlers actually read your siteWhat is answer engine optimisationHow to check what engines say about you, properlyContradictions across your own pages