Who answers for your brand when somebody asks an AI?

Read this before the numbers

What this measures, and the thing it does not

This study measures citation: which website a language model points to when it answers a question about a brand. It is a direct read, with no Google in the path — the model was asked for real, with web search on, and its citations were recorded.

Interactive · the receipts

All 89 measurements

Every row, every cited domain, classified by who controls it. Input tokens are the tell: about 320 means the model answered from memory without searching, and 8,174 means it searched and read pages.

The finding

Three conditions have to hold together

One: the content has to exist — something about that specific association, on a domain the brand controls. Two: the question has to reach it. Three: that page has to be the one selected, out of everything else that also answers. No one of them is enough alone, and that is the whole result.

Rolex failed the second with the first intact, so rewording rescued it. TAG Heuer and Jacob Elordi failed the third, and that is the case worth your attention. TAG was cited zero times in three asked one way and zero in three asked the other. We first read that as missing content. It was not: magazine.tagheuer.com published an article in April 2023 naming Elordi, and it names Ryan Gosling in the same piece. Gosling got that host cited back 2 of 3 in the losing form. Elordi got it zero times out of six.

Retrieval fired on all six Elordi runs at 8,174 input tokens, so the model did search. It cited tatlerasia.com at position 1 in all six, while the brand’s own article naming him was never selected. Publishing the page is not sufficient, and neither is naming the brand in the question.

The limits

What this evidence cannot say

Every question here is one we wrote. We do not know which phrasings real buyers use, and the whole finding turns on phrasing. That is the biggest hole in it.

The 2×2 is one fact at one brand — twelve rows. It is clean and it is small. Rolex also maintains a per-testimonee page library and a separate newsroom host, which most brands do not.

We inferred missing content from missing citations, and we were wrong. This page first read TAG Heuer’s absence on Jacob Elordi as TAG having no Elordi page. The Business of Marketing’s fact checker found one: magazine.tagheuer.com, April 2023, naming him directly and naming Ryan Gosling alongside him. Corrected 19 October 2026. The finding is now three conditions rather than two, and it is stronger for it. Where content has still not been verified by hand — Cade Cunningham and Nike — this page now says so rather than inferring.

The origin dead zone was only ever asked in origin phrasings. By this study’s own logic, a brand-named version might retrieve and cite. Until that is run, treat the zero as provisional.

Subjects were chosen because they were publicly verifiable, which selects for well-documented associations, which selects for ones that have pages. The true owned share across all commercial associations is probably lower.

Questions and answers

What people ask about this study

Method: ADMADS Engine, the evidence system CULT+MATH built in-house. DataForSEO AI Optimization live endpoint, model gpt-4.1-mini-2025-04-14, web search enabled, United States and English, output capped at 1024 tokens, three runs per question. Every cited domain was classified as the brand’s own, the person’s own, press, reseller, encyclopedia, wire or platform, under rules written down before the data was pulled. One run returned an HTTP 500 and is excluded; a daily spend cap refused 17 rows of an earlier pass, and that pass is reported as incomplete. Models are not deterministic: a single answer is an observation, not a ranking, which is why nothing here rests on fewer than three runs. Figures last pulled October 18, 2026 and they will drift.

© 2026 CULT+MATH LLC · Ipalibo Da-Wariboko