Most sites AI assistants cite for UK searches miss Google's top 10

Google's own AI Overviews mostly look past page one

Start with Google itself. If any AI should echo Google's rankings, it is the summary Google places above them. About seven in ten of the sites cited in Google's AI Overviews did not appear in the top 10 ordinary results for the same search, and roughly one overview in five cited no top-10 site at all.

Google shuffles its results by person, place and moment. Each search here was checked once, on a phone, through google.co.uk, so the figures are a single snapshot of what one UK searcher could have seen.

Even so, the comforting belief has a crack in it before any chatbot enters the picture. The links in Google's own AI box and the blue links underneath it are, for the most part, different lists.

Perplexity and Claude stray furthest of the four from page one

The four assistants go further still. Perplexity and Claude sit together at about nine in ten of their cited sites falling outside Google's top 10, with only a fraction of a percentage point between them. Gemini and ChatGPT are neck and neck, each a shade under eight in ten.

So even the two assistants that overlap most with Google's first page, of the four checked, are pointing readers somewhere else most of the time. For a firm sitting proudly in position three, that is a different shelf in a different shop.

The four were not all asked in the same way. ChatGPT and Gemini were read from their free chat apps. Perplexity and Claude were asked through their developer versions with web search switched on, which can behave differently from the apps people download. Claude's share also rests on far fewer cited sites than Perplexity's, because it linked out much less often.

Data table: assistant citations against Google's top 10, October 2026

Claude links to a website in about one answer in four

A fair objection: perhaps the assistants barely cite anything, and the gap is built on a handful of stray links. The answer depends on which assistant you ask.

Perplexity put at least one link in every answer, around eight sites each time. Claude linked out in only about a quarter of its answers, though when it did, it offered around four sites. ChatGPT and Gemini sat between the two.

AI answers change from one asking to the next, and each search went to each assistant once, so these rates describe one round of questions. The point for a business is plain enough. An assistant that rarely links offers fewer chances to be cited; when links do appear, they still lean heavily away from Google's first page.

ChatGPT and Claude rarely land on the same site

If the assistants are not copying Google, perhaps they copy each other, and winning a citation in one wins it everywhere. The evidence points the other way.

When ChatGPT and Claude both answered the same search with sources, they shared at least one site in only about one search in five. That rests on a small group of searches, since Claude linked out so rarely. The closest-matching of the six pairings, Claude and Perplexity, did better, sharing a site on a little over half the searches both answered.

Picture four shop assistants asked the same question at the same counter, each pointing to a different aisle. That is closer to what a UK searcher meets than the idea of one web, ranked once, read aloud by every machine.

What this means for UK businesses

• Treat a Google ranking as a head start. Page one still matters for Google itself, but most AI citations for these searches came from elsewhere.
• Check each assistant on its own. They disagree with each other, so being cited by one tells you little about the rest.
• Expect some assistants to link sparingly. Where an assistant rarely cites, being named in the answer text may matter as much as the link.
• Answer the question outright. A page that gives a clear, complete answer to an everyday question gives any assistant something worth linking to.
• Recheck regularly. A single round of answers is a snapshot; results change between askings.

The row of links under a chatbot's answer looks familiar, and that familiarity is misleading.

How we built it

This analysis used one fixed list of 360 everyday UK searches across twelve topics: shopping, home and DIY, health, money, travel, food, tech, motoring, jobs and education, legal and government, fashion and beauty, and home services and property. In October 2026 each search was typed once, word for word, from the UK into four AI assistants, giving 1,440 answers. ChatGPT and Gemini were read from their free chat apps; Perplexity (Sonar) and Claude (Haiku 4.5) were asked through their developer interfaces with web search on, which can differ from their apps.

The same searches were checked once that month on Google, on mobile, through google.co.uk. "Google's top 10" means the first ten organic results: unpaid, and excluding features such as maps, videos or AI Overviews. A site counts as "cited" when an assistant or AI Overview linked it as a source; links to Google Maps or a Google search were not counted as cited sites. The outside-the-top-10 shares are counted across all sites cited, with the counts shown. The pairing figures are counted across searches that both assistants in the pair answered with at least one source.

All figures are direct counts from a single check, so they are a snapshot: Google varies its results by person, place and moment, and AI answers vary between askings. The search list was chosen to cover common kinds of everyday search; it is not a random sample of all UK searches. Claude's figures rest on fewer answers with sources than the other assistants', and the ChatGPT and Claude pairing rests on 96 searches.