AI assistants answer from two different systems, and companies routinely score well in one while being invisible in the other. Model memory is what an assistant recalls about your market from training, with no internet access; live AI search is what it composes after browsing the current web and citing sources. ChatGPT, Claude, Gemini, Grok and Perplexity all operate both modes, and because the two are built differently, they fail differently - which is why serious AI visibility measurement always reports them separately.
How model memory decides who gets recommended
Model memory is a compressed picture of the public web as it looked during training. When a buyer asks "best providers of X" with search off, the model answers from that picture: companies that were consistently described, frequently mentioned and clearly classified before the training cutoff get recalled; everyone else blurs into the background. Two consequences follow. Recall lags reality - a repositioning or product launch takes time to enter memory. And recall rewards consistency - a company described the same way across hundreds of pages beats one with five conflicting descriptions, regardless of which is the better vendor.
How live AI search decides who gets cited
Live AI search flips the logic: the assistant runs web queries, reads a handful of current sources and composes an answer with citations. Here the question is not "what does the model remember about you" but "do the sources the engine trusts for this query mention you today". Even a company absent from model memory can dominate live answers by being present in the right comparison articles, directories and review sites - and a famous company can vanish from live answers when the cited sources skip it. Live results also move fast: publish better evidence, earn a citation, and answers can change within weeks.
Why model memory and live AI search diverge
The divergence is diagnostic. Strong memory, weak live search usually means legacy fame with a thin current citation footprint. Weak memory, strong live search is the profile of newer companies doing citations right - memory catches up with training cycles. Weak in both means foundational problems: unclassifiable category language, unverifiable claims or a machine-unreadable website. Strong in both is the goal, and it is measurable. This is also why one blended "AI visibility score" misleads: it hides which system is failing, and the fixes are different. The methodology EntityRise uses reports both systems separately for that reason.
The $49 audit measures the answers of five assistants, so you can see which system leaves you out →
What to do about memory and about live search
For model memory: entity consistency everywhere (one category sentence, identical naming), broad accurate presence across profiles and press, patience across training cycles. For live AI search: a citation map of the sources engines actually use in your category, verifiable evidence those sources can reference, and technical accessibility so engines can read you directly. The work overlaps but is not identical - and only remeasurement shows which side moved.
Frequently asked questions
Which matters more for B2B?
Both reach buyers, but live AI search usually moves first and moves faster, so most programmes start there while building the consistency that memory rewards.
Can I tell which mode produced an answer?
Often yes: cited sources and browsing indicators signal live search; instant uncited answers usually come from memory. Measured audits record the mode explicitly instead of guessing.
Why do I get different answers on different days?
Model outputs vary between runs, and live search depends on which sources the engine fetched that day. Single answers are anecdotes; repeated runs are measurement.
Kamil Kępiński · Founder of EntityRise and Smophy Labs Inc.
Kamil Kępiński is the founder of EntityRise and Smophy Labs Inc. He writes about how models choose the companies they recommend and how to measure a company's visibility in AI answers.
About the authorXLinkedInThe methodologycontact@entityrise.ai
Two systems, two failure modes, one programme.
Measure both, let your own team or agency fix both, and repeat the audit to prove the change.