Has AI search changed how Americans buy, and the 64 readings that say what has actually moved

Something has changed and it is narrower than the headlines and wider than the sceptics allow. Here is what we can show. On 22 October 2026, across 64 blind readings of the ChatGPT and Gemini product surfaces with American buying questions, the engines produced 2,866 citation URLs across 1,143 distinct hosts, named a supplier in 24 readings, and named no supplier at all in 38. Every one of the 53 namings went to a software product. No American agency was named once. The change an American business should act on is not that buyers have abandoned Google. It is that a second answer now exists, it is generic in most categories, and the slot for a specific answer is open.

What we measured, and the three limits of it

The bank was 198 American buying questions in 33 batches, put to both engines through a data provider. The United States was stated inside the text of every question, because the provider accepts gl, hl and location parameters and ignores all three, so the prompt is the only market lock that holds. Batch 33 names us in the question, which makes it a ceiling measurement rather than a market measurement, so it is excluded from every figure here. That leaves 64 blind readings, 32 from ChatGPT and 32 from Gemini.

Three limits, stated plainly because they change how the numbers should be used. This measures what the engines answer, not how many Americans ask. It covers two of the five commercial surfaces, so Perplexity, Claude and Google AI Overviews are outside it. And it is one day, which makes it a baseline rather than a trend.

The engine split, which is the finding that averaging would destroy

Reading	ChatGPT	Gemini
Blind American readings	32	30
Ran a live web search	32 of 32	15 of 30
Named a supplier that was not us	13	11
Named no supplier at all	19	19

One surface read the live web on every American question. The other read it on half. That single row has a direct operational consequence: a page published today is available to one engine immediately and to the other on a slower cycle, so a programme reporting a single blended AI visibility number is averaging two instruments with different refresh behaviour.

What has moved, in three sentences a board will accept

• A second answer exists and it is currently generic. In 38 of 64 American readings, no supplier was named. The engines answered the buying question with a framework.
• Where a name appears, software wins and services lose. All 53 namings were products: Semrush 14, HubSpot 12, Ahrefs 9, BrightLocal 5, Peec 3, Profound 2, SE Ranking 2, Otterly 2, and Scrunch and Conductor once each and Surfer twice. Products publish specifications, prices and documentation. Most service firms publish claims.
• The sources are not vendor marketing. The hosts the engines leaned on were developers.google.com in 17 readings, support.google.com in 15, searchengineland.com in 13, blog.google in 11, clutch.co in 11, semrush.com in 10, ftc.gov in 9 and blog.hubspot.com in 9. A host in the raw data, ccn.web.tr at 8, is the provider's own fetch infrastructure and is excluded from any source analysis.

The three decisions to make on this evidence

• Stop treating it as one channel. Report the engines separately. On this reading they differ on live search behaviour and on how often they commit to a name, and a blended figure hides both.
• Work the empty answers first. A reading that named nobody has no incumbent to displace, so the cost of taking it is publishing a specific, dated, checkable answer. A reading that already named three vendors is a harder and slower job. Sort your own question list by that distinction before you spend anything.
• Publish what the engines actually cite. On this evidence that means documentation style specificity, trade press, directory profiles and regulator aligned accuracy, not more opinion pieces. The shape that gets quoted carries a figure, a date and a denominator.

What this does not show, said plainly

We have no American share figure for how many buyers ask an assistant before contacting a supplier, and we will not publish one we did not measure. The published surveys in circulation differ on definition, sample and whether the question asked about shopping, research or supplier selection, so quoting one as a fact about your category would be borrowing somebody else's denominator. What is honestly available to an American business today is its own reading, which costs an afternoon to take and is specific to the questions its own buyers ask.

How to take your own baseline this week

• Write twelve buying questions in the words your American customers actually use, including the United States or your state in the text of each one.
• Put each to ChatGPT and Gemini, save the answer text and the source list, and record the date.
• Mark each reading as named you, named a rival, or named nobody. That three way split is the whole scorecard.
• Count the hosts. The repeat hosts are your source target list and they are usually not your link building list.
• Re-read in thirty days with the same twelve questions and the same denominator, and report the two engines in separate columns.

If you want the comparison between the two queues before you commit, the strengths of each are set out in what each is better at and what the data shows. For a worked example of a baseline that started at zero in this market and what the second reading looked like, see measuring from a United States baseline of zero.

Questions about whether AI search has changed American buying

Are Americans using assistants instead of Google?

Our evidence speaks to what the engines answer rather than to how many people ask, so the honest answer is that a second answer now exists alongside Google's and it is generic in most American categories. Measure your own category rather than adopting a market wide claim.

Which engine matters more for an American business?

They matter differently. On our reading ChatGPT ran a live search on all 32 of its American readings and Gemini on 16 of its 32, so recency reaches them at different rates. Report both.

Why did no American agency get named?

Because what the engines could read about American agencies was thinner than what they could read about software. Vendors publish checkable specifications and prices with dates. That is the gap, and it is closeable.

Is 64 readings enough to act on?

It is enough to act on the three decisions above, because each of them is about method rather than about a market share estimate. It is not enough to publish a national trend, which is why we have not.

How often does this need re-measuring?

Monthly, holding the questions, engines and denominator constant. The constancy matters more than the frequency, because two readings taken on different banks are two different measurements.

Why the second answer is worth a reporting line of its own

An American marketing report that covers search and media is covering one queue well and the other not at all. Our own reading of 22 October 2026 found 64 American readings in which we were named zero times, a rival was named 24 times, and nobody was named 38 times. AI Knows Us reads ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews for the buying questions in a category, reports each engine separately with its denominator and keeps the raw answer text, so the line in the report has something behind it. In the United States that is the Growth plan at 299 dollars a month with 100 buyer questions watched, read on aiknowsus.com/pricing and aiknowsus.com/llms.txt on 22 October 2026.