Does the way a question is worded change which brand an AI names?
The evidence
One rewording flipped a result completely.
Asking which watch brand a named golfer is an ambassador for cited the brand's own domain zero times in three runs. Asking whether that golfer is that brand's testimonee cited it first, three times out of three. Same fact, same brand, same pages, same afternoon.
Naming the brand roughly doubles the chance its own site is cited.
Across 89 registered measurements covering 12 brands, questions that named the brand returned an owned citation in 82% of rows. Questions that left it unnamed returned one in 42%. Founding and origin questions returned zero owned citations across 17 rows.
A question can be worded so that it returns no brands at all.
Five questions about who does a particular kind of professional work were asked in the way a buyer would naturally put them — “who advises a new CMO in their first 90 days?” All five came back with job titles. Marketing director, analyst, chief marketing officer. No firms, and no web search: the model answered from its own parameters, at a few hundred input tokens. Reworded to name the thing being looked for — “which consultancies advise a new CMO in their first 90 days?” — all five searched the web and named firms. Thirty-six of them across the five questions, against none before.
That is a harder condition than an unnamed brand, and most audits do not test for it.
A question that leaves your brand out still returns a field you could be in. A question the model reads as asking which internal role does the job returns no field at all. The buyer who asks it is not comparing vendors and is not going to see one, however visible you are. Counting brand names in your question list does not catch this; you have to count how many of your questions a model would answer with a job title.
Question shape can change the answer, not just the sources.
In a separate study, one model asked a direct factual question answered correctly three times out of three, and asked an open-ended question about the same subject contradicted itself on the underlying fact. The wording was the only variable.
Most measurement runs one shape and reports it as a score.
If a report gives you a single AI visibility number, ask which question shapes produced it. Two shapes that disagree about you are not an error in the measurement — that disagreement is the result, and a single-shape report conceals it.
What would have to be true for this to be wrong
Then count a second column: how many of those ten a model would answer with a job title rather than a list of firms. Those are not a harder condition, they are no condition — there is no vendor set to be in. If most of your ten fall there, visibility work is the wrong spend and the question is whether buyers consult an assistant for this decision at all.
89 registered measurements across 10 passes, 30 questions, 12 brands, three runs each, October 18, 2026, published in full at cultmath.com/ip-association. Question-shape measurements from October 24, 2026. Predictions registered with probability bands before retrieval in both studies.
Role-versus-vendor wording added November 10, 2026: five questions in each shape, ChatGPT gpt-4.1-mini-2025-04-14, web search enabled, US-scoped, via DataForSEO. One generation per question in each shape, so the five-for-five flip is the finding and no rate is quoted from it. Platform-default budget tier, not the tier a consumer is served.
Every row behind this answer is published in the standing study at cultmath.com/research/games/gta, which is re-run on a schedule and corrected in place, dated.
Published 2026-10-24 by Ipalibo Da-Wariboko, CULT+MATH LLC. Where a number could not be sourced, this page says so and gives none.
This is one of six decisions covered in the ai visibility decision brief — each with the evidence behind it, who owns it, who it is contested with, and what would have to be true for the answer to flip.
$495 · PDF on payment · 14-day guarantee
“Does the way a question is worded change which brand an AI names?”
This answer, its date and the sources under it are available to ChatGPT and Claude through the CULT+MATH app. Connect it in either one, ask the question above, and the findings come back cited rather than summarized.

