AI Search in the US and UK: What the Data Shows About Clicks and Citations

When each market got it

Start with the calendar, because “AI search” arrived in the US and the UK on different dates, and any before-and-after comparison has to respect that. Every date below comes from Google’s own announcements.

Scale has moved since. At I/O in May 2026, Google said that AI Mode “has surpassed one billion monthly users” one year after its debut. That is a global figure, not a US or UK one, and it is Google’s number, not an independent one. The practical reading is narrower: in both countries, the AI layer is no longer a test a few users might see. By the dates above, AI Overviews have been rolling out for more than two years in both the US and the UK, and AI Mode for more than a year in both.

What Pew measured, and what it didn’t

The study everyone cites is from Pew Research Center, published on July 22, 2025. Pew analyzed the browsing of 900 U.S. adults in March 2025: 68,879 unique Google searches, of which 12,593 produced an AI summary. That is real behavior captured by a tracking app, not a survey of what people say they do. The headline numbers:

Pew’s own wording: users who encountered an AI summary “clicked on a traditional search result link in 8% of all visits,” and those who did not clicked “nearly twice as often (15% of visits).” Pew also found that 18% of all the Google searches in the study generated an AI summary, so most searches in that sample were still traditional results.

Four limits are worth keeping in mind before quoting it:

• It is US-only. We have not found an equivalent click-level study of UK users from a primary source, and this article does not borrow the US figures for the UK.
• It is Google-only. Pew cites technical limits in identifying AI-generated summaries on other search engines, so the analysis includes only Google searches.
• It is a snapshot. The browsing period was March 2025, before AI Mode reached the general US audience in May 2025. It says nothing about behavior inside AI Mode.
• It measures clicks per visit, not total traffic. It compares visits that had a summary with visits that did not. These are different searches, not a controlled experiment on the same ones.

Exposure, at least, is not in doubt. A later Pew survey (field dates February 17–23, 2026) found that 60% of U.S. adults say they have read AI summaries at the top of a search result, with 30% saying they have not and 10% unsure.

What Google says about traffic

Google’s position has been consistent in direction and thin on numbers. In May 2024, at the US launch, it said “the links included in AI Overviews get more clicks than if the page had appeared as a traditional web listing for that query.” In August 2024, announcing the UK expansion, it said that showing links to supporting web pages directly within AI Overviews is “driving higher traffic to publisher sites.”

The most direct statement on total traffic came on August 6, 2025, from Liz Reid, who heads Google Search: “Overall, total organic click volume from Google Search to websites has been relatively stable year-over-year.” The same post says Google is sending slightly more “quality clicks” than a year earlier, and defines them as clicks “where users don’t quickly click back”. Earlier, at I/O 2025, Google also said AI Overviews were “driving over 10% increase in usage of Google for the types of queries that show AI Overviews” in major markets such as the U.S. and India.

These are statements from the company that runs the product, about the whole web, in aggregate. They are not figures you can audit, and they say nothing about your site.

Why both can hold

Pew and Google look like opposites. They are not quite measuring the same thing. Pew compares the click rate per results-page visit with and without a summary. Google describes the total volume of clicks across all of Search. Volume is rate times the number of visits: if summaries make individual visits less likely to end in a click while people also run more searches, including longer and more complex ones, which Google says is happening, the total can hold while the click-through on the affected queries falls.

The useful consequence is about distribution, not about who is right. An aggregate that is flat can hide a large shift underneath: clicks moving from some queries, and some sites, to others. A small company does not live in the aggregate. It lives in a specific set of buyer questions, and for those the honest answer is that no published study tells you which way your traffic will go.

What you can observe directly is whether your brand is in the answer. That is a different question from whether the answer sends a visit, and it is the one this piece argues you should start measuring. The underlying mechanics are in SEO vs GEO: what actually changes.

Where the searching happens

Before choosing which engines to watch, look at where the activity is. StatCounter Global Stats publishes monthly shares by country. For August 2026, rounded to whole percentages:

Two cautions on that second table. First, it is not a count of chats. StatCounter introduced the series as a measure of “the number of referrals from individual AI chatbots to websites”, taken from the traffic of the sites that carry its tracking code, and it noted that Grok cannot be included because it does not provide referral data. So the table shows the share of clicks that chatbots send out, not the share of people who use each one. Second, the figures move slightly with the device filter and as StatCounter updates, so treat them as orders of magnitude and re-check the live page.

The reading that survives both cautions is simple. In both countries Google is still where most search happens, and the UK is the more concentrated of the two. The Google-owned AI surfaces, AI Overviews and AI Mode, are therefore where most of the exposure sits. ChatGPT dominates the chatbot referral picture in both, with Gemini, Copilot, Claude and Perplexity a long way behind. A panel that covers those five plus Google’s two features covers most of the ground.

There is a second lesson in how that table is built. A referral-based measure only counts an answer when it sends somebody to a website. An answer that names your brand, or a competitor’s, and sends no one anywhere is invisible to it. That blind spot is exactly the one your analytics share.

Measure answer share, not visits

Definition

Answer sharethe percentage of relevant answers, from a fixed panel of buyer questions, in which an AI engine names your brand. It is the GEO counterpart of share of voice.

Traffic has three problems as the lead indicator in an AI-search market. It is a lagging measure: visits arrive after the decision has been shaped. It is a partial one: a buyer can be told to shortlist you and type your name directly, which shows up as direct or branded traffic, if at all. And, as the Pew numbers show, it can fall on the queries where you are doing well. A brand that is named in the answer and a brand that is absent can have identical traffic this month and very different positions.

Answer share has its own weaknesses. Answers vary between runs, so one result proves little. It says nothing about revenue by itself. It works as a leading indicator, read alongside the lagging ones, and only if the method stays fixed. The case for it, and how it differs from rank tracking, is laid out in SEO vs GEO; the wider method it belongs to is a five-layer GEO audit.

A fixed panel, run separately for the US and UK

The method is deliberately unglamorous. Write the questions a real buyer asks before choosing a product like yours, freeze the list, and run it on a schedule.

1. Write about thirty questions per country. Comparison questions, “best for” questions, problem questions, and pricing and compliance questions. Write them in the vocabulary that buyers in that country use: US and UK buyers do not always use the same words, units, currency or legal terms for the same thing, so a panel written once and reused will drift from at least one of them. Keep the wording fixed after the first run.
2. Choose engines by the tables above. Google AI Overviews, Google AI Mode and ChatGPT first; then Gemini, Copilot, Claude and Perplexity as capacity allows.
3. Run from the country you are measuring, monthly. Log the date, engine, country, whether your brand was named, whether it was linked, and which competitors and sources the answer used. Run each question more than once, since individual answers vary.
4. Report each country on its own line. Answer share for the US and answer share for the UK, per engine. Do not average them. The features reached the two markets on different dates, the audiences differ, and a blended figure cannot tell you where to act.
5. Read the sources, not just the score. Who does the engine cite when it names a competitor? That list is your real to-do list, and the page-level fix for it is covered in answer-shaped content.

The same discipline applies on a language axis if you also sell in Spanish; there the gap is different, as argued in why Spanish is the most underpriced language in AI search.

What Search Console now adds, and leaves out

Until recently there was no first-party number for this. In June 2026 Google introduced Search generative AI performance reports in Search Console. The help page defines the unit precisely: “Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search.” It covers AI Overviews and AI Mode, can be grouped by country, and does not include clicks or queries.

That makes it useful for exactly one job: confirming, for each country, whether your pages are being surfaced as supporting links at all, and whether that is trending up. Its limits define the job of the panel. It covers only Google. A link shown is not the same as your brand being named in the text. And with no queries, it cannot tell you which buyer questions you are winning.

Google’s own guidance on what to do is plain. Its optimization guide says that “optimizing for generative AI search is optimizing for the search experience, and thus still SEO”, and that you do not need new machine-readable files, AI text files or markup to appear in its generative AI features. The foundations still apply. What changes is the scoreboard.

What this data can’t tell you

• Nothing here is a UK click study. The one rigorous behavioral source above is American, from March 2025.
• No source here measures the effect on small, specific sites. Pew’s numbers are an average over searches; Google’s are an average over the web.
• Nothing here measures AI Mode clicks. Pew’s sample predates the general US rollout of AI Mode.
• Market-share tables measure referrals on a sample of sites, not the population of searches or chats.
• This article reports no result of ours. We cite published data and propose a method; we make no claim here about any company’s rankings, traffic or citations.

A first month

If you have a product that earns and a small team, this is a proportionate first month:

1. Week one: write thirty buyer questions for each country you sell in, and freeze them.
2. Week two: run them on Google AI Overviews, AI Mode and ChatGPT, three times each, and log the results. That is your baseline answer share for each country.
3. Week three: open the generative AI report in Search Console and note the impressions by country. Read the sources the engines credit for your competitors.
4. Week four: pick the five questions where a competitor is named and you are not, and fix the page that should answer each one. Then re-run the panel next month and compare.

A baseline like this takes a few days of work and no software, and it tells you more about your position than another month of traffic charts. If you want to know which layer of your business is holding back distribution before you spend on any of it, the Tier scale is where we start.