When someone types "best CRM for a small dental practice" into ChatGPT, nothing gets crawled or ranked in real time the way a Google search works. The model either already "knows" an answer from what it was trained on, or it reaches for a live web search tool and reads a handful of pages before answering. Which of those two things happens — and what each one rewards — is the entire game for GEO.
Two different mechanisms, two different games
Training-data recall is what happens when a model answers from memory: it read about a brand enough times, in enough independent places, during training, that the association stuck. This rewards frequency and consistency — the same name, the same core claims, repeated across many sources the model actually trained on, not just a brand's own site.
Retrieval (a live web search the model runs before answering) is a completely different mechanism, and it's become the more common one as engines increasingly ground answers in fresh sources. Here the model is reading actual pages in the moment — which means it rewards whatever it can fetch and quote right now: structured, factual pages with clear, quotable claims, on domains that show up in the search results it trusts.
What training-data recall rewards
- Being mentioned by many independent sources, not just once anywhere
- Consistent facts — the same name, category, and claims everywhere a brand appears
- Being discussed in the kind of text models train on heavily: reviews, forum threads, comparison articles, documentation
What retrieval rewards
- Pages that state facts plainly and once — a price, a feature, a differentiator — not buried in marketing copy
- Presence on domains the engine's search actually surfaces for that query (review sites, directories, comparison content)
- Content that's genuinely current — retrieval has no reason to prefer a stale page over a fresher competitor
The tell that separates real visibility from a fluke
Here's the part that matters most for anyone actually trying to measure this: one engine confidently naming a brand that no other engine has ever heard of isn't a win, it's usually a hallucination. Real visibility shows up as multiple engines independently landing on the same name for the same kind of question. That's why cross-engine corroboration — not a single ChatGPT answer — is the actual signal worth tracking.
This is exactly the distinction GeoSurfaced's scoring is built around: a brand mention only counts as corroborated visibility once it holds up across engines, not just once, somewhere. If you want the full breakdown of what to actually do about it, the GEO strategy guide walks through a concrete checklist. Or run a free GEO audit to see where your own brand stands today.