Do people trust AI recommendations? What the measured data shows

People act on AI recommendations more than they say they do. The clearest evidence is behavioural rather than attitudinal: Pew Research Center's browsing study found that when a Google AI summary appeared, users clicked a traditional result link on 8 per cent of visits against 15 per cent when no summary appeared, and ended their browsing session on 26 per cent of those pages against 16 per cent. Reading the summary and stopping is reliance. On adoption, Pew's survey of 5,123 US adults found 34 per cent say they have ever used ChatGPT, rising to 58 per cent of adults under 30. This page sets out what the measured figures are, where reliance is high and where it is low, and what a business should do with that.

Adoption: the measured numbers

Pew Research Center surveyed 5,123 US adults between 24 February and 2 March 2025 as part of its American Trends Panel, and published the results on 25 June 2025. The findings give a clean picture of how far one assistant has reached.

• 34 per cent of US adults say they have ever used ChatGPT, about double the share in 2023.
• Among adults under 30 the figure is 58 per cent, up from 43 per cent in 2024 and 33 per cent in 2023.
• 41 per cent of adults aged 30 to 49 have used it, 25 per cent of those 50 to 64, and 10 per cent of those 65 and over.
• Among employed adults, 28 per cent say they use it for work, a rise of 20 percentage points in two years.
• Education tracks closely: 51 per cent of those with a bachelor's degree and 52 per cent with a postgraduate degree have used it, against 18 per cent of those with a high school education or less.
• 79 per cent of Americans have heard at least a little about ChatGPT, including 34 per cent who have heard a lot.

Two readings of that matter commercially. The audience is large and growing, and it is concentrated exactly where professional buying decisions are made: working age, educated, using it at work.

Reliance: what people do rather than what they say

Attitude surveys about trust in AI move around a lot. Browsing data is steadier, because it records what somebody did when nobody was asking them about it. Pew's separate analysis of the browsing behaviour of 900 US adults covered 68,879 unique Google searches in March 2025, of which 12,593 produced an AI summary. The findings are the best available measure of reliance.

Behaviour	With an AI summary	Without one
Clicked a traditional result link	8 per cent of visits	15 per cent of visits
Clicked a link inside the summary	1 per cent of visits	Not applicable
Ended the browsing session	26 per cent of pages	16 per cent of pages

A person who reads a summary and stops has treated it as sufficient. That is the operational definition of trust, and 18 per cent of all searches in the study produced a summary, so this is already happening at scale.

Where reliance is high, and where people still check

The pattern across categories is consistent enough to plan around, and it maps onto how much a wrong answer costs the person asking.

• High reliance. Explanations, definitions, drafting, summarising, shortlisting options, and first pass research into an unfamiliar category. The assistant is used to narrow the field.
• Checked afterwards. Prices, availability, opening hours, regulated advice, and anything the person will hand to a boss or a client. Here the assistant produces the shortlist and the buyer verifies the name on the company's own site and on review platforms.
• Still a results page job. Buying on the spot, reading a primary document, and anything local and time sensitive.

The middle row is the commercially important one. For most businesses the assistant is not making the sale, it is deciding whose name the buyer types next. Being absent from the shortlist removes you before any comparison happens, which is the mechanism behind why your competitor is recommended and you are not.

What raises trust in a specific recommendation

Buyers do not treat every AI answer the same way. Four things visibly raise their confidence in one.

• Citations they recognise. An answer linking to sources the reader knows is checked less and believed more. Pew's count found Wikipedia, YouTube and Reddit the most frequently cited sources in both summaries and standard results, together making up 15 per cent of the sources in the summaries examined.
• Several sources agreeing. 88 per cent of the AI summaries in that study cited three or more sources. A name that appears across several of them reads as consensus.
• Specific, checkable claims. A count, a price with a date, an accreditation with an issuing body. A vague superlative invites a check; a verifiable fact survives one.
• Consistency when verified. The buyer's next move is often to open your site. If the site says something different from the answer, you lose the name you had just won.

What a business should do about it

The useful conclusion is not a trust percentage. It is that a measurable share of your buyers are now getting a shortlist from an assistant, and that you can find out whether you are on it today. Ask your five real buyer questions in two assistants, without naming your company, in a temporary chat with memory off. Record who was named and which sources were cited. If a competitor is named and you are not, that is your finding, and the citation list is your work plan.

Trust is the reason this matters commercially. If a buyer acts on the first answer, the shortlist is decided before your website is ever opened. So the measurement worth having is simply whether you are in that answer, for the questions your buyers actually type. AI Knows Us takes that reading on a fixed question set and keeps the sources each answer used, which also shows you where the facts are coming from.

This page covers adoption and reliance, with the measured figures. For what changes in your own search data, see zero click search. For how an assistant decides which names to give, see how AI assistants choose which businesses to recommend.

do people trust AI recommendations?: the questions that keep coming up

How many people use AI search?

For one assistant there is a clean figure: Pew's survey of 5,123 US adults found 34 per cent have ever used ChatGPT, and 58 per cent of those under 30. For AI summaries inside Google, Pew's browsing study found 18 per cent of searches produced one and 58 per cent of respondents saw at least one during a single month.

Do people click the sources in an AI answer?

Rarely, on the Google surface at least. Pew recorded clicks on a link inside the summary on 1 per cent of visits to pages that had one. That is why being named in the text matters more than being linked from it.

Do buyers verify what an assistant tells them?

For anything with a price or a commitment attached, most do, and the verification usually starts with your own website and a review platform. Making those two agree with each other is the cheapest trust work available.

Does an AI recommendation carry more weight than a search result?

It carries a different kind of weight. A results page offers ten options and leaves the choosing to the reader. An answer hands over two or three names with the choosing already done, so inclusion is worth more and absence costs more.

Is this only a consumer pattern?

No. Pew found 28 per cent of employed US adults use ChatGPT for work, rising to 45 per cent among those with a postgraduate degree, which is the profile of the people writing shortlists in a B2B purchase.

What to do first

• Ask your five real buyer questions in two assistants today, in temporary chats, without naming your company.
• Record who was named, in what order, and every source cited.
• Compare what the answer said about your category against what your own site says. Fix every disagreement on your side.
• Add one question to your enquiry form asking how the person first heard about you, and include an AI assistant as an option.
• Get listed properly on the two most cited third party sites you are missing from, then re-run the questions in a month.