Which AI assistant names your business? The same 25 questions, put to three assistants

The short version: in our published sample, the same 25 unbranded buyer questions went to three assistants. ChatGPT named the business in 13 of its 25 answers, Gemini in 9, Claude in none — 22 named answers out of 75 graded. If you have only ever checked one chat window, you have measured one of those columns.

What we measured

The run behind these numbers is the published sample: one real local business, pseudonymised, measured on 2026-09-20. Twenty-five questions a buyer would ask without naming the business, spread across four buying stages, each put separately to ChatGPT (searching the web), Gemini (searching the web) and Claude (from memory). Five further questions name the business directly and are graded apart, because a name in the prompt proves the assistant has heard of you, not that it would recommend you.

Why the columns differ

Two of the three assistants go and search before answering. The third does not: our method asks Claude from memory, so its answers carry whatever its training data already connected to the question. That is why the zero is not a verdict on the assistant. It is what one model's memory said about this business on one day, and model memory changes on retraining cycles rather than on the schedule of your fixes.

The other two columns differ for a more ordinary reason: they search, find different pages, and assemble different answers. One of them found enough about the business to name it in half its answers; the other in a little over a third.

Being named is also uneven across the buying journey, even inside the same run. The business appeared in 1 of 9 answers to questions about what exists, 3 of 24 when buyers compared approaches, 16 of 27 when they were choosing who to use, and 2 of 15 when they were about to book. The stage matters more than the total, and the total is the only number a single check tends to produce.

What a single chat window cannot tell you

A check in one assistant answers one question: does this assistant name me. It cannot tell you what the other assistants say, and the gap between columns in this run was the whole width of the result. It also cannot tell you whether the traffic you care about comes from the assistant you tested — that depends on your buyers, not on the market share of an app in general.

If you are buying a measurement rather than doing a free check, ask which surfaces it covers and whether the count is per surface. A single blended score hides exactly the split above. Several monitoring tools price per engine for the same reason: each surface is a separate measurement, not a setting.

The limits

This is one business in one market, measured once. The sample publishes a confidence range with its decision-stage rate, 28.5% to 58.4% on a re-run, and that range is the honest way to read any percentage here: a point, not a law. Two of the three columns come from live web search, and search results move; the third comes from model memory, and model memory moves on a slower clock. The numbers support one conclusion firmly — that the three columns are not interchangeable — and no conclusion about which assistant is best in general.

Questions we get asked

Which AI assistant names a local business most often?

In our one measured run, ChatGPT named the business in 13 of 25 answers, Gemini in 9 and Claude in none. That is one business in one market, not a ranking of the assistants.

Why did Claude name the business in none of its answers?

Because our method asks Claude from memory rather than with a live web search, so its answers carry whatever its training data connected to the question, and that data did not connect this business to these questions.

Can one assistant's answer stand for all of them?

No. The same question produced different answers on each surface, and a check on one chat window says nothing about the other two.

Does being named mean a buyer will choose you?

No. In the same run the business was the most-named in its market, rank 1 of 168 candidate businesses, and still scored 46.1 out of 100, because some of the facts were wrong and the sources the answers were built from barely mentioned it.

Where to go next

• Check all the surfaces you care about: the free check reads your site and shows you the exact questions a full measurement would ask, before any payment.
• Read the full measurement behind this article: the published sample prints every answer and every grade, including the per-assistant recall table.
• See what the report costs and contains: the pricing page — one report, $4.99 once, no subscription.
• Check the method: how the measurement works — surfaces, repeat counts, and the confidence range we publish with every percentage.

Sources

• SeeWAIS, published sample report, run of 2026-09-20 — https://seewhataisays.com/sample (the per-assistant recall table in Part 2 and the headline score in Part 1). Accessed 2026-11-04.
• SeeWAIS, method v1.3 — https://seewhataisays.com/method (the five surfaces, the repeat counts and the confidence note). Accessed 2026-11-04.
• Google Search Central, "AI features and your website" — https://developers.google.com/search/docs/appearance/ai-features (accessed 2026-10-27). Cited for the search-side surface behaviour; the access date is the one already recorded on this site's method page, not a fresh read this week.