How to Track AEO Mentions in ChatGPT and Gemini

To track AEO mentions in ChatGPT and Gemini you have to check them two different ways, because the two engines expose two different kinds of evidence. ChatGPT answers from a live web search and attaches the source URLs it used, so a mention can be tied directly to the page that earned it. Gemini answers from Google Search grounding and returns something ChatGPT never does — the list of search queries it actually issued to answer your question — but its source links come back as opaque Google redirect URLs, so your own domain will never match them unless those links are resolved first. A single manual spot-check misses both problems, and it misses a third: the model tier you ask changes the answer. The reliable method is the same question set, run on a schedule against both engines, recording per-answer whether the response was grounded, truncated, and which sources it cited.

The two engines retrieve differently, so "did it mention me" means two different things

ChatGPT's search path returns cited source URLs as annotations attached directly to the answer text. Gemini's grounded path returns retrieved source chunks plus the full list of search queries the model issued while answering. These are not cosmetic differences — they are different evidence types, and a tracking method built for one silently drops information from the other. Both are real, structured fields worth capturing on every stored check: citations, search queries, whether the response was grounded, and whether it was truncated.

What only Gemini tells you: the fan-out

Gemini hands back the searches it ran to answer your question — the clearest evidence available of how an engine decomposes a buyer's question, and which brand names it carries into its own searches along the way. Nothing equivalent exists on the ChatGPT side, so a Gemini-only check that ignores the query fan-out is throwing away the one signal that engine uniquely offers.

What only ChatGPT tells you: clean source URLs

ChatGPT's citations resolve to real, readable URLs. Gemini's citation URIs come back as opaque vertexaisearch…/grounding-api-redirect/… links; left unresolved, a brand's own domain can never be matched against them, and the citation signal is silently lost. That resolution step has to happen before a Gemini citation count means anything at all — treat any Gemini citation tally that skips it as unverified.

The trap that invalidates most manual checks: which model you asked

On the same question, the same day, a smaller search-preview model can put a brand at the top of its answer while the full search model does not name it at all. Checking whichever tier is cheapest to query flatters whoever happens to rank well in it. The number that matters is the tier your actual buyers talk to — usually the default, full-capability model — and a method that does not pin the model tier is measuring the wrong thing.

Truncation looks exactly like invisibility

A cut-off answer and an answer that genuinely never mentioned a brand are indistinguishable downstream unless the check records a truncation flag. This is not a hypothetical: a real production tracking pipeline carried a run of truncated Gemini rows for weeks before the response-length caps were fixed and truncation dropped to zero. Any historical "zero mentions" number that predates a fix like that should be treated as unverified until it is re-run.

A repeatable method, five steps