Which AI assistant Americans use most, and whether that should change what you buy

Usage share is the wrong input for this decision and it is the one every vendor leads with. The right inputs are two behaviours we measured directly. First, how often each engine reads the live web when answering an American buying question, because that decides how fast your publishing takes effect. Second, how often each engine commits to naming a supplier at all, because that decides whether the slot in your category is occupied or empty. On 22 October 2026, across American questions, ChatGPT ran a live search on all 32 of its readings and Gemini on 16 of its 32. Both named nobody in 19 readings each. Those two facts should change a budget. A usage share percentage should not.

What we measured, per engine

Reading, American questions, 22 October 2026	ChatGPT	Gemini
Blind readings, brand batch excluded	32	30
Ran a live web search	32	15
Named a supplier that was not us	13	11
Named no supplier at all	19	19
Named us	0	0

The bank was 198 American buying questions in 33 batches, put to both surfaces through a data provider, with the United States named inside the text of every question because the provider accepts gl, hl and location parameters and ignores all three. Batch 33 names us in the question and is therefore a ceiling measurement, excluded throughout. The two engines produced 2,866 citation URLs across 1,143 distinct hosts between them, and we report them separately rather than averaged, because the live search row is exactly what an average would destroy.

The two behaviours, and what each one should change

• Live search frequency changes your publishing cadence and your re-read schedule. An engine reading the live web on every question will reflect a new page sooner than one reading it half the time. If you publish in week one, the fair re-read date differs by engine, and a programme reporting a single blended number in week four is measuring two different lags as one.
• Naming frequency changes where you start. Both engines declined to name anybody in 19 of their American readings, which is 38 of 64 in total and the largest single outcome. Those readings are the cheapest openings available, because taking an unoccupied slot requires publishing a specific answer rather than displacing an incumbent.

Why usage share is a poor purchase input

Three reasons, and each is enough on its own. The published share figures disagree because they define the category differently, with the biggest single disagreement being whether an AI answer inside a search results page counts as assistant use. Share tells you how many people might ask and not what they are told, and the second is the thing you can change. And share is a national average while buying behaviour is category specific, so a national figure can be right and irrelevant to your market at the same time.

What usage share is good for is a rough ordering when you can only afford to read one or two engines. Five surfaces are sold in this category, ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and reading two of them honestly while saying which two is a real measurement. Reading one and calling it AI visibility is not.

What the engines cited, which differs from what they are

Across the 64 blind American readings the most cited hosts 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, blog.hubspot.com in 9, with brightlocal.com and searchenginejournal.com at 7 each and goodfirms.co at 6. One entry in the raw host data, ccn.web.tr at 8 readings, is the data provider's own fetch infrastructure rather than a publisher and is excluded from every source analysis and never merged into another host.

Where a supplier was named, the 53 namings were all software: 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 by either engine. The two engines did not name the same mix in the same proportions, which is a second argument for reading both rather than picking the larger one.

How to decide which engines to pay to read

• Read at least two, and say which two in every report. One engine is an anecdote with a dashboard.
• Prefer the pair that behaves most differently from each other, because that is what reveals whether a result is about your business or about one engine's refresh.
• Do the arithmetic before you look at plans. Tracked questions, times engines, times refresh frequency, times markets. The engine count is a multiplier, so adding the fifth engine costs more than adding the second.
• Ask whether the vendor reports the engines separately. A blended score would have hidden the single most useful row in our own table.
• Ask whether you get the raw answer text, because a score with nothing under it cannot be audited.

If the engine you are least sure about is Gemini, the vendors that treat it as an extra rather than as standard are identified in two vendors sell Gemini as an extra. If the surface in question is Google's AI Mode rather than Gemini, the state of tooling for it is in a surface we do not read and the two vendors that claim it.

Questions about which assistant to care about in the United States

Which one do Americans use most?

The published share figures disagree on definition, and we did not measure usage. What we measured is behaviour on American buying questions, where ChatGPT ran a live search on all 32 of its readings and Gemini on 16 of its 32.

Should I optimise for the biggest one only?

No, because the behaviours differ and one engine's result does not predict the other's. Read two at minimum and report them separately.

Does the usage leader change by category?

Very likely, which is another reason a national average is a poor input. Your own reading on your own questions is the category specific answer.

Why does live search frequency matter so much?

Because it sets how quickly a page you publish can appear in an answer, and therefore when a re-read is fair. Two engines with different frequencies need different re-read dates.

Is it worth paying for all five surfaces?

If your category's buyers use them, yes, and the cost rises with the engine count because engines are a multiplier in the arithmetic. Start with two named engines and add deliberately.

Reading both rather than betting on one

The decision this page argues against, picking the market leader and treating its answer as the answer, is exactly what our own table disproves, because the two engines differed on live search frequency by more than two to one. AI Knows Us reads ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, reports each engine separately with its own denominator, and keeps the raw answer text so the split can be checked rather than trusted. In the United States the cheapest tier that reads all five engines is Growth at 299 dollars a month with 100 buyer questions watched, read on aiknowsus.com/pricing and aiknowsus.com/llms.txt on 22 October 2026. Outbound, at 1,299 dollars in the United States, reads four of the five rather than all five, which is the kind of detail worth checking on any plan.