The State of AI Search

UK Edition, Summer 2026

A working reference for anyone in the UK trying to understand how answer engines actually decide what to say, and who they say it about.

It is deliberately not a trend piece. It has three jobs: give you the UK numbers with their sources attached, explain what each major engine is built on and what it rewards, and be honest about which parts of the evidence are solid and which are not. Quote any of it, with attribution.

Published July 2026Updated quarterlyBy Tilio

In this post

At a glance

The short version

Six things we think matter this quarter.

AI tool use in the UK reached 26.4 million people in April 2026,
up 2.4 million in three months (UKOM / Ipsos iris). That is roughly half the adult online population, and it is still climbing.
ChatGPT’s grip on UK attention is loosening.
Its share of total time spent on AI tools in the UK fell from 84% in October 2025 to 76% in January 2026 (UKOM / Ipsos iris). Claude.ai posted the fastest UK audience growth of any tool between January and April 2026, up 232%.
Ranking and citation have formally decoupled.
The share of Google AI Overview citations coming from pages ranking in the organic top 10 fell from 76% in July 2025 to 38% by March 2026 (Ahrefs). Being cited and ranking are now two different games with two different scoreboards.
Google’s two AI surfaces disagree with each other.
Across 540,000 query pairs, AI Mode and AI Overviews cited the same URLs only 13.7% of the time, despite reaching broadly the same conclusions around 86% of the time (Ahrefs). One company, one index, two different answers about who to trust.
Retrieval is not selection.
ChatGPT cites roughly 15% of the pages it retrieves; the other 85% are pulled in, assessed and discarded (AirOps, analysing 548,534 pages across 15,000 prompts). Getting found is now the easy half.
The engines run on different indexes.
ChatGPT reads Bing. Gemini and AI Overviews read Google. Claude appears to read Brave. Perplexity mostly reads its own crawl. “Ranking in AI search” is not a thing you can do once.

Part one

The UK picture in numbers

Most AI search statistics you will read are American, global, or vendor-generated. These are the UK-specific ones we would actually put in front of a board.

Metric	Figure	Source and date
UK people using AI tools monthly	26.4m	UKOM / Ipsos iris, April 2026
Growth in UK AI users, Jan to Apr 2026	+2.4m	UKOM / Ipsos iris, April 2026
ChatGPT share of UK time spent on AI tools	76% (was 84% in Oct 2025)	UKOM / Ipsos iris, January 2026
Fastest-growing AI tool in the UK, Jan to Apr 2026	Claude.ai, +232% monthly audience	UKOM / Ipsos iris, April 2026
UK visits to ChatGPT, Jan to Aug 2025	1.8bn (up from 368m a year earlier)	Ofcom, Online Nation 2025
UK ChatGPT visits, August 2025, year on year	+156%	Ofcom, Online Nation 2025
UK year-on-year growth to Aug 2025, other engines	Gemini +146%, Claude +138%, Perplexity +100%	Ofcom, Online Nation 2025
Share of keyword searches returning an AI-supported response	~30%	Ofcom, Online Nation 2025
Daily AI tool use among UK 15 to 24s	~24% use one every day	UKOM / Ipsos iris, Sept 2025
Share of UK AI tool time occurring in work hours	44%	Ipsos iris, August 2025

Reading the numbers

Read the second row again. The story of 2025 was adoption. The story of 2026 is fragmentation. ChatGPT losing eight points of UK time share in a single quarter, while Claude grows 232%, is the number that should change your planning. A visibility strategy built around one engine is now a strategy with a measurable and growing blind spot.

The work-hours figure is the commercially interesting one. 44% of UK time on AI tools happens between 9am and 5.30pm. These are not teenagers writing essays. This is professional research behaviour: the exact top-of-funnel activity that used to belong to Google, and the reason B2B and professional services firms cannot treat this as a consumer story.

One caveat worth stating plainly: the Ipsos iris AI category excludes AI built into other products, notably Google Search’s own AI Overviews. So the real UK exposure to AI-generated answers is materially higher than 26.4 million. Most people encountering AI answers in the UK are not opening an app to do it. It is happening inside a Google search they were doing anyway.

Part two

The engines, one by one

There is no such thing as “AI search”. There are several engines, built on different indexes, with different retrieval logic, different citation conventions and different audiences. They frequently disagree about who to recommend. This is the section to bookmark.

ChatGPT

Reads Bing. The most under-done technical work in UK marketing sits here.

Built on: Microsoft’s Bing index, accessed through Web IQ, Microsoft’s AI-native grounding layer, which returns passage-level evidence rather than whole pages. OpenAI supplements this with its own crawlers: OAI-SearchBot for search indexing and ChatGPT-User for live fetches during a conversation.
What it looks for: Direct, extractable answers near the top of a section. Clean HTML it can parse without executing JavaScript. Question-shaped headings that match how people actually phrase things. Broad third-party corroboration of your brand, rather than repeated claims on your own site. Microsoft evaluates its grounding results against a metric covering completeness, freshness and authority.
What gets you cited: Being in Bing’s index, first and foremost. This is the single most under-done piece of technical work in UK marketing right now. If Bing cannot see you, ChatGPT cannot cite you, and your Google rankings will not save you. After that: server-side rendering, fast response times, specific factual claims, and a presence in the reference and media layer (Wikipedia, editorial press, established directories).
What gets you ignored: Client-side-rendered content. Blocking OAI-SearchBot. Padding. Pages that make claims nothing else on the web corroborates.
The number that matters: ChatGPT cites only about 15% of the pages it retrieves (AirOps, March 2026). And roughly 46% of ChatGPT queries trigger a live web search at all; the rest are answered from training data (Semrush analysis of 80 million queries). So more than half the time, your visibility is a function of what the model absorbed during training, not what it can find today. That is a slow-moving asset built by being written about, and it is not fixable in a sprint.

Google AI Overviews

Extracts passages, not pages. A page can rank and still lose the citation.

Built on: Google’s own index, Knowledge Graph and Shopping Graph, with Gemini generating the summary.
What it looks for: Passages, not pages. AI Overviews extract a specific block of text that answers a specific sub-question, which is why a page can rank well and still lose the citation to a competitor with a cleaner two-sentence answer.
What gets you cited: Answer-first structure. Self-contained sections of roughly 150 to 400 words that resolve one question completely, without needing the paragraph above for context. Schema that matches your visible content. Ranking still helps as an entry ticket, but it is no longer the decider.
The number that matters: Top-10 organic pages accounted for 76% of AI Overview citations in July 2025. By March 2026 that was 38% (Ahrefs). In eight months, the link between ranking and being quoted halved. Meanwhile the average AI Overview cites around 13 sources, against one for a classic featured snippet, so the number of slots went up even as the qualification criteria changed.

Google AI Mode

One question becomes a dozen retrievals. Query fan-out changes what a good page is.

Built on: The same Google infrastructure, but a different retrieval path. Google confirmed at I/O 2026 that AI Mode runs on a custom Gemini 3.5 Flash build, tuned for retrieval and citation rather than open-ended conversation. It passed one billion monthly users by May 2026, roughly a year after launch, with query volume reportedly doubling every quarter.
The mechanism: query fan-out: AI Mode does not run your query. It decomposes it into multiple sub-queries and retrieves sources for each one independently, then synthesises. One question becomes a dozen retrievals. This is the most important architectural fact in AI search, because it changes what a “good page” is: you are no longer competing to answer the question, you are competing to answer the questions behind the question.
What gets you cited: Coverage breadth on the sub-questions. Analysis of 10,000 keywords covered by Search Engine Land found pages ranking for fan-out queries were 161% more likely to earn citations, and pages ranking for both the head term and at least one fan-out variant accounted for 51% of all citations, against under 20% for pages covering only the main query. The correlation between fan-out coverage and citation likelihood came in around 0.77, which is unusually strong for search data.
What to watch: AI Mode overlaps with the organic top 10 on roughly 12% of citations (Moz, 40,000 keywords). And Google cites itself heavily: google.com accounted for 17.42% of AI Mode citations by February 2026, up from 5.7% in under nine months (SE Ranking).

Perplexity

The one engine where citation is still a traffic strategy, not only a visibility play.

Built on: The most architecturally independent of the majors. Its own crawler, PerplexityBot, builds its own index, supplemented by partner search APIs. Its default answers run on in-house Sonar models. Retrieval is a multi-stage pipeline: semantic retrieval pulls candidates, an LLM reranks them, and a quality classifier can discard the lot. Perplexity-User handles live, request-driven fetches separately.
What it looks for: Freshness, harder than any other engine. Fact density. Entity clarity. Pages that state a specific, bounded answer rather than hedging.
What gets you cited: Being in the index at all, then being new. Content begins losing ground within months. Visible “last updated” dates, refreshed statistics and current-year references do real work here. Original data does the most work of all, because there is no substitute source for it.
Why it earns disproportionate effort: Perplexity cites more sources per answer than any other surface and surfaces them prominently, so its citations convert to actual clicks at rates well above AI Overview citations. It is the one engine where citation is still a traffic strategy rather than only a visibility play. Its UK audience is small but skews to researchers, analysts and professional buyers.
A trap to avoid: Blocking PerplexityBot does not protect you the way blocking Googlebot would. Per Perplexity’s own documentation, a blocked page can still leave the domain and a brief factual summary in the index. You lose the citation and keep the exposure.

Claude

Passes Brave’s rankings through nearly untouched. The most directly testable lever in AI search.

Built on: Almost certainly Brave Search. Anthropic has not confirmed this on the record, and we flag that honestly, but the evidence chain is strong: Brave appears on Anthropic’s subprocessor list, the API internals reference Brave search parameters, and independent analysis found 86.7% overlap between Claude’s cited sources and Brave’s top organic results, with a p-value below 0.0001 (Profound). Image search is separately documented as Bing-powered.
Why 86.7% is the most actionable number in this report: Claude passes Brave’s rankings through nearly untouched. By comparison, ChatGPT reorders Bing’s results heavily, showing only around 26.7% alignment. So Claude is the one major engine where the lever is direct and testable: rank in Brave, get cited in Claude. Brave runs an independent index of roughly 40 billion pages, refreshed at over 100 million pages a day, built without Google’s or Bing’s results. It is smaller and more selective, which acts as a quality filter.
The UK angle: Claude.ai was the fastest-growing AI tool in the UK between January and April 2026, up 232% (UKOM / Ipsos iris). Its audience over-indexes hard on technical, enterprise and professional users. If you sell to those people and you have never checked whether you exist in Brave’s index, that is a five-minute job with an outsized payoff.
What gets you cited: Presence in Brave. Allowing Claude-SearchBot (Anthropic runs separate crawlers for search, training and user-initiated fetches, and honours robots.txt across all of them). Fact-dense writing where claims carry sources. Claude applies a visibly heavier credibility filter than the other engines and deprioritises unsupported assertions.

Microsoft Copilot

Runs on Bing, skews to the workplace, and almost nobody is contesting it.

Built on: Bing. Which means your ChatGPT technical work largely carries over, with one addition: IndexNow, which pushes updates to Bing immediately rather than waiting for a crawl.
Why it is underrated in the UK: Copilot’s usage skews to workplace and enterprise contexts more than any other AI surface, and Bing’s SEO competition is a fraction of Google’s. For B2B, it is the cheapest visibility available. Almost nobody is contesting it.

Gemini (the app)

The one engine where classic SEO strength transfers most directly.

Built on: Google’s index. This is the one engine where classic SEO strength transfers most directly, and where schema on your own domain does the most work. If your organic foundation is strong, Gemini is largely inherited rather than earned.

Part three

Five things that are true across every engine

Strip out the platform differences and this is what remains.

Ranking is now an entry ticket, not a result.
Across every study we trust, position acts as an upstream filter into the retrieval pool and then stops mattering. What decides the citation is the shape of your content, not its rank. This is why sites can lose rankings and gain visibility, and why the reverse also happens.
Getting retrieved is the easy half.
ChatGPT discards 85% of what it pulls. Perplexity’s quality classifier can bin an entire result set. There are two gates, and most optimisation advice only addresses the first.
The engines disagree, and the disagreement is the strategy.
Different indexes produce different answers to the same question. Two Google surfaces agree on the URL only 13.7% of the time. Optimising for one engine and assuming the rest follow is the single most common and most expensive error we see.
Freshness became a ranking factor.
Every engine weights recency, Perplexity most aggressively. A page that earned citations last year can quietly stop, without anything about it changing except the date. The fix is not backdating metadata; it is updating the evidence.
What other people say about you outweighs what you say about yourself.
Corroboration across independent sources is the shared currency. This is the mechanism behind the observation that AI engines lean on media, community and reference sites more heavily than on brand-owned domains, and it is why publishing volume on your own site has such poor returns in AI search compared to what it used to earn in SEO.

Part four

What we would actually do about it

In priority order, assuming a UK business starting from a normal position.

Check Bing.
Verify in Bing Webmaster Tools, submit your sitemap, enable IndexNow. This unlocks ChatGPT and Copilot in one move. It takes an afternoon and it is the highest-leverage hour in AI search.
Check Brave.
Run site:yourdomain.com on search.brave.com. If it returns nothing, you are invisible to Claude regardless of your Google position. Submit and re-test.
Audit your crawlers.
Confirm OAI-SearchBot, PerplexityBot, Perplexity-User, Claude-SearchBot and Claude-User are not blocked in robots.txt, at the CDN, or by a WAF rule nobody remembers writing. Blocking is not protection; it is just absence.
Kill client-side rendering on anything you want cited.
AI crawlers largely do not execute JavaScript. If your content only exists after hydration, it does not exist.
Restructure your top 20 pages for passage extraction.
Answer-first openings. Self-contained sections. Question-shaped headings. Visible last-updated dates. This is the work that converts retrieval into citation.
Map the fan-out.
Take your head commercial query, list the ten sub-questions behind it, and check which ones you answer with an extractable passage. The gaps are your losses.
Invest off-site.
Earned media, credible directories, community presence and original data. The engines reward corroboration, and you cannot corroborate yourself.
Measure per engine.
Not “AI visibility” as one number. Six numbers, because the engines disagree, and an average hides the fact that you are winning one and absent from another.

On the evidence

Methodology, and an honest note

We think this section matters more than any single statistic above.

What we did: We compiled UK-specific measurement from UKOM / Ipsos iris and Ofcom, the two sources with the most defensible UK methodology, and combined it with published 2026 research on engine architecture and citation behaviour from Ahrefs, Moz, SE Ranking, Semrush, AirOps, Search Engine Land, Profound and platform documentation from Google, OpenAI, Microsoft, Anthropic and Perplexity. Every figure in this report carries its source and date inline.
What we are confident about: The UK adoption and share-of-time figures. UKOM-endorsed panel measurement is the closest thing the UK has to ground truth, and it is independent of anyone selling AI visibility services. The engine architecture is well documented, mostly by the platforms themselves.
What we hold more loosely, and why you should too: A large share of the citation-behaviour research in this field is produced by companies selling AI visibility tools, including some cited here. That does not make it wrong, and much of it is the only data that exists. But the incentive runs one way: every study in this category has a commercial interest in the finding that AI search is important and hard. Where studies disagree, we have said so rather than picking the more dramatic number. Where a claim rests on inference rather than confirmation, notably Claude’s use of Brave, we have flagged it.
What nobody knows: None of the engines publish their ranking formula. Everything in Part two about “what gets you cited” is pattern-reading, not documentation. Treat it as a well-evidenced hypothesis, test it against your own data, and be suspicious of anyone who states it with more certainty than that, including us.
What we will not claim: This is a picture of the UK market as measured by the sources named. It is not a census.

Reuse

How to cite this report

Tilio, The State of AI Search: UK Edition, Summer 2026, July 2026. https://www.tilio.co.uk/resources/state-of-ai-search/uk-summer-2026

Journalists, analysts and anyone else are welcome to quote any figure or finding with attribution. If you want the underlying working or a comment for a piece, contact us.

Quick answers

Common questions

What is AI search, and how is it different from SEO?

AI search is when an answer engine such as ChatGPT, Google AI Overviews, Claude or Perplexity generates a direct answer and cites a handful of sources, rather than returning a list of links. SEO improves where you rank. Answer engine optimisation (AEO) improves whether you are selected and cited inside the answer itself. Ranking now acts as an entry ticket into the retrieval pool, not the thing that decides the citation.

Which AI engines matter most in the UK right now?

ChatGPT still leads UK time spent on AI tools at 76%, but its share is falling while Claude, Gemini and Perplexity grow. Because each engine reads a different index (Bing, Google or Brave), you cannot optimise once and appear everywhere. This report breaks the major surfaces down one by one.

Why do different AI engines cite different sources for the same question?

They run on different indexes and different retrieval logic. Even Google’s own two AI surfaces, AI Mode and AI Overviews, cite the same URL only 13.7% of the time. That is why measuring visibility per engine matters more than a single blended score.

How do I find out if my business appears in AI answers?

Start with a free AI visibility check to see how your domain looks to the major answer engines, or book an AI visibility audit for a written assessment across ChatGPT, Google AI Overviews and related signals.

This report explains the mechanism. These are the practical next steps.

Where you stand

Everything above describes the mechanism. The question it raises is a specific one: when a UK buyer asks ChatGPT, Claude, Perplexity or Google to recommend a business like yours, does your name appear?

That is measurable, and it takes about a minute to find out.

Check your AI visibility score

The next edition, Autumn 2026, publishes in October, with the same UK measurement tracked forward and sector-level visibility data added. Tilio is a UK answer engine optimisation agency. We help brands earn visibility in AI search.