ChatGPT Search: do answers with sources name the same brands?

We paired an answer with no recorded citations with the nearest answer to the same question that had at least one, up to fourteen days apart. Among 38,997 pairs where both answers named a brand, the first brand matched in 34.5%.

For comparison, two cited answers to the same question matched in 45.9% of 81,859 pairs. The practical point is that seeing a source list does not make a single answer a stable read of your brand’s position.

Mixed-citation pairs shared an average of 28.0% of the brands in their combined lists. Pairs where both answers had recorded citations shared 36.1%. This measures shared membership, rather than whether either answer made the better choice.

Track the brands an answer names and the role it gives each one. The first brand mentioned may be background context, an example or a recommendation; its position in the text alone does not settle that.

The study classifies answers by recorded citations. It does not observe a clean retrieval-on versus retrieval-off experiment. Search can occur without a visible citation, and collection differences can affect which citations are stored.

The no-recorded-citation share varied sharply during the study: 44.0% in May, 28.1% in June, 1.9% in July and 0.81% in August. That collection-era change is why these pairs cannot establish a causal effect of web search.

Repeat the buyer questions that matter. Record which brands appear, which are recommended and which pages support the answer. Look for a recurring change before treating one different first mention as a gain or loss.

When a cited page contains an outdated price or misses a capability, there is a concrete fact to investigate. When the brand list moves without a clear evidence change, collect more observations before choosing an explanation.

The source study contains 452,990 ChatGPT Search answers to 19,792 tracked questions, May 24–August 28, 2026. It produced 43,757 mixed-citation pairs and 90,858 two-cited-answer pairs; first-brand comparisons use only pairs where both answers named a brand. The comparison pairs are restricted to before July 6 to overlap the main period containing answers without recorded citations.

Canonical brand identities merge name variants. The first-named brand is the one appearing earliest in the answer text. A separate storage-path calculation gave 34.8% first-brand agreement versus 34.5% in the primary calculation.

A separate storage-path calculation gave 34.8% first-brand agreement versus 34.5% in the primary calculation. The study groups answers by recorded citations; it does not establish whether a search took place.