Every AI answer falls into one of two worlds. In the first, the model consulted sources on the spot — went to the web, pulled pages, and grounded the answer in them. In the second, it answered from memory, drawing on what it absorbed in training, without checking anything live. The two worlds have different rules for who gets to appear. And confusing them is one of the most expensive strategic mistakes in GEO.
Two Mechanisms, Two Entry Points
The sourced answer is a retrieval problem. The model is, in that instant, choosing among real, available pages. Your brand appears if your page is one of the best answers to that question — clear, relevant, trustworthy, interpretable. The entry point here is the quality and presence of your specific page, right now.
The unsourced answer is a training problem. The model isn't looking at any page; it's accessing patterns it internalized months or years ago. Your brand appears if it was already consolidated in the data the model saw — mentioned enough, in good enough sources, associated with the right concepts. The entry point here is your historical, accumulated presence.
The same brand can be strong at one door and weak at the other. It's common to have excellent pages (and win the sourced answers) but thin consolidated presence (and vanish from memory answers). Or the reverse: be a known brand the model recalls from memory, but have a weak site that doesn't get pulled when there's live search.
Why the Sourced Answer Is a More Accessible Opportunity
There's an important timing asymmetry. Influencing the memory answer is slow: it depends on building presence that makes it into future training cycles, which takes time and runs on a calendar you don't control. Influencing the sourced answer is faster: if you improve a page today, it can be retrieved tomorrow in a live search.
That makes the retrieval layer the most practical leverage point for most brands. Not because memory doesn't matter — it does, a lot — but because it's where your action has the most immediate, measurable effect.
The Risk of Aiming at Only One World
Optimize only for retrieval — flawless pages, but no external authority building — and you're hostage to appearing only when the model decides to search. In the many answers that come from memory, you vanish.
Bet only on brand presence — lots of press and mentions, but a site that doesn't answer questions in a structured way — and you appear from memory while losing the sourced answers, where the page has to measure up.
Robust presence comes from covering both: pages that deserve to be retrieved AND a distributed reputation that makes it into training. They reinforce each other. A good page becomes a source others cite; those citations feed memory presence; memory presence makes the model trust your page more. The cycle is virtuous when both sides exist.
How to Tell Which World You're Losing In
Here's the operational catch: from the outside, the two answers look identical. The user sees text with the AI's take on your industry. You can't tell, just by looking, whether that answer came from live search or from memory — and so you don't know which of your two doors is stuck.
The clue is in the behavior. Answers with links, named sources, and very recent information probably came from retrieval. Generic answers, no links, stopping at a certain date, probably came from memory. Crossing that with where you appear and where you vanish starts to reveal the diagnosis.
From Distinction to Action
The resulting strategy is twofold and simple to state. To win the sourced answers: make your pages the best available answer to the real questions in your industry, and keep them current. To win the memory answers: build consistent presence in authority sources over time, so your name is consolidated when the next model is trained.
Measuring the two separately is what stops you from wasting effort reinforcing the door that's already open while the other stays locked. Genoma distinguishes these scenarios in its monitoring — showing not just whether you appear, but in what kind of answer — precisely so you invest where the gain is real.