The American data on assistant use before a purchase, and what each kind of source actually measured

American data on assistant use before a purchase exists, and it is weaker than its headlines. The figures in circulation come from four different kinds of source and each measures something different: panel based traffic estimates, self reported consumer surveys, platform disclosures, and vendor commissioned research. Those four disagree by large margins and they are all answering slightly different questions. So the useful posture for an American business is to read any figure with four questions attached, and to take one measurement of its own, because the one thing nobody else can measure for you is what the engines say when somebody asks about your category.

The four kinds of source, and what each actually measures

Source type	What it measures	What it cannot tell you
Panel traffic estimates	Visits or sessions on assistant properties, modelled from a sample	Whether any of it was commercial intent, or what was asked
Consumer surveys	What people say they do	What they actually did, and people under report routine use and over report novel use
Platform disclosures	Usage the platform chooses to publish, on its own definition	Comparability with anything else, because the definitions are not shared
Vendor commissioned research	Something real, with a question set designed by an interested party	How the result would look under a neutral question set

The four questions to attach to any figure

• What was the exact question wording? Shopping research, supplier selection and general use are three different populations and the headline usually collapses them.
• Who was sampled, and how recruited? Consumers and business buyers behave differently, and an online opt in panel is not a random sample.
• Did an AI answer inside a search results page count? Including Google's AI answers raises the figure substantially over standalone assistant use, and this single choice explains much of the disagreement.
• What were the field dates? Not the publication date. This behaviour is moving fast enough that two quarters is a different market.

A figure with all four attached is worth putting in a board paper, labelled as what it measured. A figure with none of them is the one that gets challenged in the meeting, and rightly.

The measurement we took, and its denominator

On 22 October 2026 we put a 198 question American buyer bank in 33 batches to the ChatGPT and Gemini product surfaces through a data provider. The United States was named 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, so it is excluded from every figure.

• 64 blind American readings, 32 from ChatGPT and 32 from Gemini.
• 2,866 citation URLs across 1,143 distinct hosts.
• Named us in 0 readings, cited our domain in 0 readings.
• Named a supplier that was not us in 24 readings, 53 namings in total.
• Named no supplier at all in 38 readings, the largest single outcome.
• All 53 namings were software products. Semrush 14, HubSpot 12, Ahrefs 9, BrightLocal 5, Peec 3, Profound 2, SE Ranking 2, Otterly 2, Scrunch 1, Surfer 2, Conductor 1. No American services firm was named.
• ChatGPT ran a live web search on 32 of 32 readings. Gemini on 15 of 30.

What this measures is what the engines answer, with a denominator and a date. What it does not measure is how many Americans ask, and we do not publish a share figure because we did not take one. Those two are different instruments and a page that mixed them would be doing exactly what the four source types above do.

Why the answer matters less than the content of the answer

Here is the argument for spending ten minutes on your own reading rather than a week on the literature. Suppose the true share of American buyers consulting an assistant before contacting a supplier is 15 per cent. Suppose it is 50. In both cases, if the answer those buyers receive names nobody in your category, the action is identical: publish the specific, dated, checkable answer and get cited on the hosts the engines read. The share figure changes the urgency and not the action, and the action is knowable with a denominator you control.

How to take your own reading

• Write twelve buying questions in the words your American customers use, with the United States or your state named in each.
• Put each to both engines and save the answer text, the named suppliers and the source list, with the date and the engine.
• Score three ways. Named you, named a rival, named nobody.
• Count the hosts. The recurring ones are your target list and they will differ from a link building list.
• Repeat monthly with the same twelve questions and the same denominator, reporting the engines separately.

If you want a single engine source mix at larger scale before you build your own, the counted version is in the source mix behind 306 readings. If the thing you actually want to compare is the two queues rather than the share, that comparison is in what each is better at and what the data shows.

Questions about American assistant usage data

Is there a reliable American figure for assistant use before purchase?

There are several figures, from four different kinds of source, and they disagree because they measure different things. Read any of them with the four questions above attached rather than as a single fact.

Why will you not publish a share figure?

Because we did not measure one. We measured what the engines answered, 64 blind American readings on 22 October 2026, and mixing a measured figure with an estimated one in a single sentence would make both less useful.

Which source type is most trustworthy?

Whichever one publishes its question wording, its sample, its definition and its field dates. The type matters less than the disclosure.

Does AI Overviews count?

It is a choice rather than a fact, and it is the choice that most moves the published numbers. Decide for your own reporting, write it down, and keep it constant.

How small can my own reading be and still be useful?

Twelve questions is enough to decide what to publish. It is a census of the questions you chose rather than a sample of a population, so it supports your decisions and not a claim about the market.

A denominator you own, instead of a headline you borrow

The reason to take your own reading is that every published share figure answers a question somebody else framed, and the reading answers yours. AI Knows Us reads ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews for the buying questions in a category, reports each engine separately with the denominator stated, and keeps the raw answer text so a figure can be audited rather than quoted. In the United States that is the Growth plan at 299 dollars a month with 100 buyer questions watched and 1,600 credits, the cheapest tier reading all five engines, read on aiknowsus.com/pricing and aiknowsus.com/llms.txt on 22 October 2026.