If retrieval and trained knowledge disagree, which one wins?
It depends on the assistant. Some weight the freshly retrieved page over the trained answer; others blend both into one response, and the mix isn't consistent across sessions.
ANSWERS
Training data vs live retrieval: where does AI get brand facts?
AI models get brand facts from trained knowledge and from retrieved sources, and a fact can appear in both. Trained knowledge changes with new model versions, on the provider's own cycle. Retrieval may draw on indexed, cached, or live sources, so a page change does not update answers instantly. ye5's GEO Brand audit compares the answers and citations observed across Claude, ChatGPT, and Gemini.
Where brand facts come from, and how each route updates
This distinction runs through GEO measurement generally, and the full timing mechanics live on a dedicated page (see: How long does GEO take to show results?). Applied to brand facts specifically: trained knowledge is built into the model's weights and changes with new model versions, on the provider's own cycle. Retrieval may draw on indexed, cached, or live sources, so a published change does not appear instantly. Blocking one crawler can reduce access on that route, but it does not cut off every route to your information (see: Does blocking AI training bots hurt my visibility in AI search?).
Which route carries which brand facts
A brand fact can appear in trained knowledge and in retrieved sources. Stable facts — company category, founding year — often sit in trained knowledge and rarely need a refresh. Facts that change — pricing, a new product, a recent review — depend on retrieved sources to stay current.
ye5's GEO Brand audit runs customer questions across repeated sessions and compares the answers and citations it observes. It cannot prove hidden training provenance (see: What is GEO (Generative Engine Optimization)?).
A new model version does not guarantee that a wrong fact changes. A corrected page can change answers once it is retrieved, on no promised timeline (see: Does content freshness matter to AI models?).
Why you can't diagnose this from a single answer
A single wrong answer doesn't reveal which source produced it. The same question run a second time might return a different result, if retrieval pulled a different page. It might also return the identical wrong fact, if the error sits in trained knowledge. Personalization compounds the problem: account memory, phrasing, and session history each shift what a lone check returns.
FAQ
Does a wrong fact in trained knowledge fix itself over time?
No. A new model version is trained on the provider's own cycle and does not guarantee the fact changes. Published corrections can reach retrieved sources rather than the trained layer.
Do Claude, ChatGPT, and Gemini retrieve live pages by default?
No. Retrieval behavior differs by product and by query. Some trigger a live fetch automatically; retrieval may also use indexed or cached sources rather than a live fetch.