Where American buyers start looking for a supplier in 2026, measured on 64 readings rather than asserted

American buyers start in four places, and only one of them is a search box. They start with somebody they already trust, they start on a platform where suppliers are listed and compared, they start on a search engine, and from 2025 onward a growing share start by asking an assistant. The fourth is the one nobody has evidence about, so here is ours. On 22 October 2026 we put a 198 question American buyer bank to the ChatGPT and Gemini product surfaces and counted what came back: in 64 blind readings, 38 returned no supplier name at all, 26 named a supplier, and every one of those 53 namings went to a software product rather than to an agency or a local firm.

What we measured, exactly

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

Reading	ChatGPT	Gemini	Both
Blind readings	32	30	62
Ran a live search	32	15	47
Named a supplier that was not us	13	11	24
Named no supplier at all	19	19	38
Named us	0	0	0

The two engines are reported separately and never averaged, because the one row that would be destroyed by averaging is the most useful one. ChatGPT ran a live search on every American reading. Gemini ran one on half of them. A supplier publishing something today is therefore reaching those two surfaces on different schedules.

The hosts the engines leaned on, which is where American buyers are actually sent

Across the 64 blind readings the engines produced 2,866 citation URLs on 1,143 distinct hosts. The concentration at the top is the finding.

• developers.google.com, cited in 17 readings, and support.google.com in 15
• searchengineland.com in 13 and searchenginejournal.com in 7
• blog.google in 11 and business.google.com in 7
• clutch.co in 11 and goodfirms.co in 6, the two agency directories
• semrush.com in 10, blog.hubspot.com in 9 and brightlocal.com in 7, all vendor properties
• ftc.gov in 9 and consumer.ftc.gov in 6, the regulator

One entry in the raw data, ccn.web.tr at 8 readings, is the data provider's own fetch infrastructure rather than a publisher. It is excluded from this analysis and should never be presented as a source.

Read that list as a map of where an American buyer gets routed. Documentation and the regulator at the top, then trade press, then two directories, then vendor blogs. A supplier's own site appears nowhere near the top. The practical consequence is that being a good page on your own domain is a necessary condition and not a sufficient one.

The four starting points, and what to do about each

Starting point	What the buyer does	What a supplier does about it
A referral	Asks somebody in their network, then checks the name online	Make the check easy. A current site, a verifiable price position, and a real address survive a referral check.
A platform or directory	Compares listed suppliers side by side	Claim and complete the profiles on the directories the engines read, clutch.co and goodfirms.co among them, and keep the price position current
A search engine	Types a commercial query and reads the first page	The familiar work, with one addition. Google's own documentation and support properties dominate the citation list, so writing against what those documents actually say pays twice
An assistant	Asks a buying question in plain language	Publish a specific, dated, checkable answer to that question, and get cited on the third party hosts the engines already read

Why the 38 readings that named nobody are the opportunity

The largest single outcome in the American set is an answer with no supplier in it. An engine that answers in generalities has not committed to anybody, which means the slot is open and the cost of taking it is publishing a concrete answer rather than displacing an incumbent. Compare that with a reading that has already named three vendors, where the work is harder and slower. Sort the buying questions in your category by that distinction and work the empty ones first. It is the cheapest available order of operations in this channel.

What to do in the next thirty days

• Write down the twelve questions an American buyer in your category actually types or says. Not keywords, questions.
• Put each to ChatGPT and Gemini with the United States stated in the question text, and save the answer and its sources. That is your baseline and it takes an afternoon.
• Mark each reading as named a rival, named nobody, or named you. Work the named nobody set first.
• Claim and complete your profiles on the directories that appeared in the citation list, then check what they say against your own site.
• Re-read the same twelve questions in thirty days, same engines, same denominator, and report the two engines separately.

The city level version of this measurement behaves differently again, and the four metros we checked are set out in every reading returned a source and no source led. If you want the source mix behind a larger single engine run before you build your own baseline, it is counted in the source mix behind 306 readings.

Questions about where American buyers begin

Is search still the main starting point?

For most American categories it remains the largest single one, and the citation evidence above shows the assistants themselves leaning on Google's documentation and support properties. The change is that a second queue now exists alongside it with a different gatekeeper and no reporting in most marketing packs.

How many American buyers use an assistant first?

We do not have a share figure and will not publish one we did not measure. What we measured is what the assistants answer when they are asked, which is a different question with a solid number behind it, 64 readings on 22 October 2026.

Why did the engines name software and not agencies?

Because the checkable, dated, published evidence the engines could read in the American agency market was thinner than the equivalent for software products. Vendors publish specifications, prices and documentation. Most agencies publish claims.

Do directories matter if nobody clicks them?

They matter because the engines read them. clutch.co appeared in 11 of 64 readings and goodfirms.co in 6, which makes a current, accurate profile a cheap input into an answer a buyer sees.

How often should this be re-measured?

Monthly is enough for most categories, and the discipline that matters more than frequency is holding the questions, the engines and the denominator constant so two readings are comparable.

What a second queue with no reporting line costs an American supplier

Every American marketing report covers search and media. On 22 October 2026 ours found 64 American readings in which the assistants named us zero times, named a rival 24 times and named nobody 38 times, which is the shape of a queue most suppliers are not measuring at all. 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 figure can be audited. In the United States that is the Growth plan at 299 dollars a month with 100 buyer questions watched, the cheapest tier that reads all five engines, read on aiknowsus.com/pricing and aiknowsus.com/llms.txt on 22 October 2026.