AI Overviews Out-Trust TikTok. The Headline Number Isn't in the Report.

What the study actually measured

The fieldwork, as reported by ppc.land, covered 1,840 respondents across the United States, the United Kingdom, France and Japan, run during June and July 2026 through the survey platform Cint.

The trust results split cleanly by geography.

Tool	US trust	Non-US trust
Google AI Overviews	43%	35%
ChatGPT	43%	41%
Claude	23%	not reported at this level
Grok	21%	not reported at this level

Note the reversal. Inside the US, Google AI Overviews and ChatGPT tie. Outside it, ChatGPT leads and AI Overviews drops eight points. If you optimise for one surface on the assumption that the two behave alike, that gap is the first thing worth knowing.

The generational split runs against the usual assumption too. Chain Store Age reports that US Gen X (61%) and Baby Boomers (62%) were among the most receptive to the idea that AI is good at surfacing brands they did not previously know. AI-assisted discovery is not, on this evidence, a young-shopper behaviour.

One channel still beats every AI tool: friends and family, at 59% in the US.

Where the report stops agreeing with itself

Here is the part that matters for anyone planning to quote this study in a deck.

ppc.land, which read the full document, documents that the report puts friends and family at 59% US trust “against 44% for the leading AI tool” — while its own trust chart records 43% for that same tool. A one-point gap, in the same document, left unreconciled.

That is small. It is also the kind of small that tells you how carefully the rest was assembled.

ppc.land documents two more. The report’s non-US narrative figures for the decision-delay finding sum to 32% while the press materials state 33%. And the stacked bar chart illustrating that same finding appears to have its US and non-US generational panels transposed, with the panel labelled United States showing the numbers that match the non-US narrative text.

We are attributing all three to ppc.land rather than claiming them, and the distinction is deliberate. The full 14-page report sits behind a form. We did not open it. ppc.land did, and we are reporting what it says it found there. If you intend to cite the discrepancies yourself, cite them the same way.

The number in the press release that isn’t in the report

This one is cleaner, because it does not depend on reading the gated PDF at all.

The Business Wire release leads its bullet points with: “46% of US Millennials have already purchased from a brand they discovered through AI this year.” We fetched that release and read the line ourselves.

ppc.land’s verdict on it, verbatim:

That figure appears in the release headline bullets but not in the published 14-page report, which contains no supporting chart, question wording or cross-tabulation for it.

A shopping statistic with no question wording behind it is not a finding. It is a sentence. And it is the sentence most likely to end up in a slide deck, because it is the one that says the thing everyone selling AI-visibility services wants said.

Who commissioned this, and what the finding sells

RTB House is not a research house. It is a performance-marketing company, and its own landing page for the study sits inside a commercial site.

Now look at what the study’s central non-trust finding says: 42% of US respondents report that AI tools increase the time it takes them to reach a purchase decision, by surfacing more alternatives.

A longer consideration window means more opportunities to reach a shopper between first interest and checkout. That is, mechanically, more retargeting inventory. RTB House sells retargeting. Its VP of Product Marketing and Analytics is quoted in the release advising marketers to “take advantage of that extended window”.

This does not make the 42% wrong. Commissioned research can be accurate, and a company is allowed to study its own market. It does mean the study’s most strategically convenient finding for its publisher is also the one carrying the least independent corroboration, and that is worth naming rather than passing over.

What survives

Strip out the press release and the commercial framing, and something real is left.

The trust ranking is the most robust part. Four independently-published accounts agree on the topline figures, the sample size, the fieldwork window and the fielding platform. Nobody has to take RTB House’s word for what its own respondents said, because the numbers reproduce across outlets that reported them separately.

So: as of mid-2026, in a 1,840-person sample, two AI surfaces out-rank every social platform on stated shopping trust, and they do it differently inside and outside the United States. That is a genuine data point on a beat that has had almost none.

What it is not is evidence that those surfaces send proportionate traffic, or convert, or that trust translates into a click. Stated trust in a survey and observed referral behaviour are different measurements, and this study only ran the first one.

What we could not check, and would want to

Three limits sit on top of everything above.

RTB House does not publish per-country sample sizes for its non-US group. The United Kingdom, France and Japan are reported as a single bloc throughout, which means three markets with quite different retail structures are averaged into one line and cannot be separated back out.

The agentic-commerce findings measure stated willingness in a hypothetical, not transactions. Chain Store Age reports that 42% of US Millennials say they would be comfortable giving an AI agent a budget of up to $250 when a seven-day return window is guaranteed, against 27% of non-US Millennials. Comfort in a survey has historically run well ahead of observed behaviour in this category.

And the trust question measures relative preference among people already using these tools. It does not measure how many people use them, how often, or whether trust survives a bad recommendation.

If RTB House published the report ungated, with per-country samples and question wording, most of this post would become unnecessary. That is the standard worth holding vendor research to, including the vendor research that happens to agree with you.