AI Search Behaviour Statistics 2026: How People Query, Click and Verify

Most reporting on AI search has been about counting heads. How many people use ChatGPT. How many searches show an AI Overview. Those numbers answer the wrong question for anyone producing content. The more useful question is what people do differently inside an AI tool. Do they type the same clipped keywords? Do they ask follow-ups? Do they click anything? Do they check the answer? By mid-2026 there is finally enough clickstream and panel data to answer that. Queries have roughly doubled in length, sessions have started to lengthen after eighteen months flat, almost nobody clicks, and hardly anyone verifies. There is a UK section below, because British behaviour diverges from the US data in ways that matter.

Key AI search behaviour statistics for 2026

• Google AI Mode queries average 7.22 words against 4.0 words for traditional Google queries (Semrush) [3]
• Google says the average AI Mode search is triple the length of a traditional one, and follow-up queries have grown more than 40% month on month [4]
• Search-enabled ChatGPT prompts grew from 4.7 to 8.7 words between early 2025 and early 2026 (Semrush) [1]
• Queries per ChatGPT session jumped roughly 50% to 1.75 between November 2025 and February 2026, after a year flat at 1.16 to 1.21 (Semrush) [1]
• AI Mode sessions contain 2 to 3 queries versus about 5 for a traditional Google session (Semrush) [3]
• Just 6% to 8% of AI Mode sessions produced an external click (Semrush) [3]
• The share of ChatGPT prompts recognisable as traditional search keywords rose from 18.9% in October 2025 to 34.9% by February 2026 (Semrush) [1]
• Only 17.4% of AI users always verify what an AI tool tells them (Clear Spark Digital, 2,127 US adults) [12]
• Ofcom found 45% of UK AI users use them to find factual information, and 42% of online adults read AI summaries often or always [14]
• 95% of ChatGPT users also used Google in the same window (Similarweb, March to May 2026) [10]

How long are AI search queries compared with Google searches?

Roughly double. Traditional Google queries average about four words. AI Mode queries average 7.22 words on Semrush's measurement of 69 million US desktop sessions [3], and "triple the length" on Google's own, published by its VP of Data Science and User Experience Research, Shivani Mohan, in May 2026 [4].

ChatGPT tells a stranger story. Semrush found search-enabled prompts grew from 4.7 words to 8.7 words, whilst prompts that did not trigger a search shrank from 24.9 words to 13.5 words [1].

That convergence is the interesting bit. Search prompts got longer whilst everything else got shorter, meeting in the middle at around eight or nine words. People are learning both that they do not need to write an essay, and that four keywords is not enough.

Where this misleads is if you read "longer query" as "long-tail keyword". A seven-word AI query is one intent stated fully, not a narrower slice of a topic. Building 400 long-tail pages to catch them is the wrong reflex.

Do people still type keywords, or have they switched to prompts?

Both, and the mix is shifting back towards recognisable search language. Semrush found 65% to 85% of ChatGPT prompts could not be matched to any of the 27 billion keywords in its database for most of the study period. By February 2026, 34.9% matched traditional search language, up from 18.9% in October 2025 [1].

That reversal is worth sitting with. The early adopter phase produced genuinely novel language. As the user base broadened, ordinary search habits came with it.

Google's data points the same way. The most common opening words in AI Mode are "What", "How", "I", "Is" and "Can" [5]. Pew's US panel found 60% of Google queries beginning with a question word returned an AI summary, against 8% of one and two-word searches and 53% of searches of ten words or more [6].

So question-form headings are not a stylistic choice any more. They are the shape of the input. But do not overcorrect: a third of AI prompts still look exactly like a 2019 keyword.

How many turns does a typical AI search session last?

Fewer than you would expect, though it is finally rising. ChatGPT sat at 1.16 to 1.21 queries per session for a solid year from October 2024, then jumped roughly 50% to 1.75 by February 2026 [1]. Google AI Mode runs at 2 to 3 queries per session against about 5 for a traditional Google session [3].

The "conversational refinement" story has been oversold. For most of 2025 the median AI search interaction was one prompt, one answer, done. That is not a conversation, it is a vending machine.

What changed in late 2025 looks like a genuine shift rather than noise. Google reports follow-up queries in AI Mode growing more than 40% month on month across May 2025 to April 2026 [4]. Multi-turn behaviour is real, it just arrived a year later than the marketing narrative claimed. Note the direction of the AI Mode comparison. Those sessions are shorter in query count than traditional Google ones. People run five queries on Google because the first four did not work. They run two in AI Mode because the first one mostly did.

What are people actually using AI search for?

Advice and information, overwhelmingly. OpenAI's analysis of 1.5 million conversations classified 49% of messages as "Asking", 40% as "Doing" and 11% as "Expressing" [7]. Around 70% of consumer usage was non-work.

Google's breakdown shows which task types are growing fastest rather than which are largest [4][5].

The comparison growth is the number commercial teams should care about. "Which of these two" is a bottom-of-funnel query answered by a model that may never link to either option. The top retail attributes people specify in AI shopping queries are price, location, colour, brand and availability [5], the exact structured facts most product pages bury in a tab.

Search intent has not changed. The interface for expressing it has.

How often do people click through from an AI answer?

Rarely, and the measurements agree unusually well for once. Semrush put external clicks at 6% to 8% of AI Mode sessions [3]. Pew found people clicked a traditional result on 8% of searches showing an AI summary versus 15% without, and a source link inside the summary on just 1% [6].

Pew also measured what happens instead. Sessions ended after 26% of visits to a results page carrying an AI summary, against 16% without [6]. The AI answer is not just diverting the click, it is ending the task. SparkToro, using Similarweb clickstream, found US Google searches ended without any click 68.01% of the time across January to April 2026, up from 60.45% in 2024 [9].

There is a genuine counterweight. Adobe's Q3 2026 analysis found AI-referred retail visitors were 36% less likely to bounce, viewed 23% more pages per visit, spent 53% more time on site and converted 54% better than non-AI traffic [8]. Travel showed 41% lower bounce and sessions 70% longer.

So the click is rarer but heavier, a pattern I unpack in the AI search conversion rate statistics. That breaks any model that values all sessions equally.

How do people verify what AI search tells them?

Barely. Only 17.4% of AI users say they always verify what an AI tool tells them, according to Clear Spark Digital's April 2026 survey of 2,127 US adults, whilst 78.5% said they had made a real-life decision based on AI output [12].

Among those who do verify, 72% use Google, 46% manually check sources or citations, 41% ask a follow-up question of the same AI tool and 15% consult a different one [12]. The last two are not verification in any meaningful sense. Asking the model whether it was right is asking the defendant to review the verdict.

A separate study of 500 AI users run in November 2025 by Eight Oh Two put the figure far higher, at 85% who "always, often or sometimes" double-check answers [13]. That shows how much a survey scale determines its own answer. "Sometimes" is doing enormous work in that 85%.

Underneath sits a structural problem. Similarweb found only 6.8% of ChatGPT answers in the US included a link to an external source as of May 2026, up more than fivefold in a year [10]. You cannot check a citation that is not there.

What device are people on, and does it change behaviour?

Mobile, mostly, though device split is the weakest area of the public data. Ofcom found smartphones accounted for 77% of UK adults' daily online time in May 2025, up from 75% the year before [15]. Similarweb recorded 4.4 billion generative AI app downloads in the year to May 2026, up 58% [11].

The behavioural consequence is multimodal input. Google reports more than one in six AI Mode searches in the US now use voice or images rather than typed text, with image searches growing more than 40% month on month since launch [4]. That is a phone behaviour, and it is invisible in every keyword tool on the market.

Be careful with device claims generally. Most rigorous clickstream studies quoted here, including Semrush's AI Mode work, are US desktop only [3]. If mobile AI search behaves differently, and the multimodal data suggests it does, a chunk of what the industry believes is desktop behaviour wearing a universal label.

How does AI search behaviour differ in the UK?

British users see AI answers more often than they seek them, and they trust them less. Ofcom's Online Nation report, published December 2025, found around 30% of keyword searches now deliver an AI-supported response and 53% of UK adults often see AI summaries, whilst noting most are "not seeking these but finding them now included by their search services" [15][17].

Ofcom's Adults' Media Use and Attitudes report, published April 2026 on fieldwork with 7,533 UK adults aged 16 and over, found 47% of AI users turn to them for work or study, 45% for finding factual information and 43% out of curiosity. Three-quarters of online adults read AI search summaries at least sometimes and 42% read them often or always, rising to around 60% among 16 to 34s [14].

Trust is the UK's distinguishing feature. 57% of UK adults would trust AI-generated news less than human-written news, and even among AI users 50% say the same [14]. Only 10% feel very confident they can recognise AI-generated content.

That scepticism shows up in how narrowly Britons will go to AI first. YouGov's July 2026 survey of 2,100 UK adults found just 8% turn first to a standalone AI tool for health information, rising to 13% of Millennials and falling to 3% of Baby Boomers [16]. Among those who had used AI for health, 39% saw no advantage in it at all.

This cuts both ways for UK marketers. Exposure to AI answers is near-universal and passive, so you cannot ignore it. But willingness to act on one without a second opinion is markedly lower than US data implies, so a findable, human-authored source page is still doing real work at the point of decision. The wider adoption picture behind this sits in the AI search adoption statistics.

Is AI search replacing traditional search or sitting alongside it?

Sitting alongside, decisively. Similarweb found 461 million of ChatGPT's 494 million users, or 95%, also used Google in the same window across March to May 2026 [10]. Traditional search averaged 3.3 billion monthly unique visitors globally against 655 million for AI chatbots combined [10]. Ofcom reports Google Search is still used by 82% of UK adults [15].

The behaviour underneath the overlap is what makes the substitution argument weak. People are not migrating, they are triaging. Ambiguous, exploratory, comparison-shaped tasks go to the AI tool. Navigational and verification tasks stay on Google, which is why 72% of people who verify AI answers do it there [12].

One demographic note: Similarweb found under-35s fell from 61% of generative AI users in May 2024 to 50% by May 2026, with growth coming from the over-45s [11]. As the user base ages into the mainstream, expect query language to keep drifting back towards conventional search habits [1].

How to read these numbers

Almost every figure above comes from one of three measurement types, and they are not interchangeable.

Clickstream panels (Semrush, Similarweb, SparkToro) observe real behaviour, but only from panellists who installed tracking software. They skew towards desktop and towards people willing to be monitored. Semrush's AI Mode study was US desktop only [3].

Survey data (Ofcom, YouGov, Clear Spark Digital) captures self-perception, not action. The 17.4% versus 85% verification gap between two 2026 studies [12][13] is a scale artefact more than a real disagreement.

First-party data (Adobe, OpenAI, Google) is the most accurate about what happened but the least representative and least neutral. Adobe sees only its own customers' sites. Google's "triple the length" figure has no published methodology.

Two further cautions. Several figures here are growth rates on tiny bases: AI Mode accounted for 0.34% of Google searches in SparkToro's January to April 2026 window [9]. And "AI search" is not one thing. AI Overviews, AI Mode, ChatGPT and Perplexity produce genuinely different behaviour, so any statistic lumping them together is directional at best.

What this means going into 2027

Query language will keep normalising. The keyword-matching share of ChatGPT prompts nearly doubled in four months [1]. As the mainstream arrives, AI prompts will look more like searches, not less.

Multi-turn is the metric to watch. Sessions moved from 1.2 to 1.75 queries in four months after a year flat [1]. If that holds, the unit of optimisation becomes the thread rather than the answer, which means being present in the second and third response.

The verification gap is a commercial opening. Fewer than one in five users always check AI output [12], and only 6.8% of ChatGPT answers carry a source link [10]. Brands that become the thing people check with will capture the clicks that remain.

Judge AI traffic on value, not volume. Adobe's engagement multipliers are large and consistent across retail, travel and financial services [8]. Reporting that treats an AI-referred session like any other will keep understating it.

The UK needs its own read. Higher passive exposure to AI summaries, lower trust, and a smaller share willing to act on AI answers unverified [14][16]. US benchmarks will overstate UK adoption of AI-first decision making for some years yet.

If you want to translate this behaviour data into a plan for how your own pages get retrieved, cited and clicked inside AI answers, you can book a call with me and we will map it out together.