Retrieval vs memory: the two ways AI knows about your business
AI assistants answer questions about your business in one of two modes. Memory mode recites what the model absorbed during training, months ago. Lookup mode (retrieval, in industry terms) searches the live web and builds the answer from what it finds right now. You can tell the two apart in most cases, and the difference decides whether an answer about you is fixable in weeks or frozen until the next model ships.
Memory answers and lookup answers
A memory answer comes from the model’s training data, so it reflects the web as it looked months or years ago and can’t be updated by anything you do today. A lookup answer is assembled live from search results, listings, and pages, which means the sources behind it are yours to improve. For local business questions, lookup answers are the norm and the opportunity.
Both modes produce the same confident tone. The model doesn’t announce which one it used. You have to read the signals.
How to tell which one you’re seeing
The strongest signal is citations. When an answer carries linked sources, the assistant looked something up; every major platform attaches links or citations to retrieved material, per their own documentation1234.
Honesty about the “mostly”: the reverse isn’t a guarantee. An answer without visible citations usually leans on memory, but platforms sometimes search without surfacing every source, and some interfaces hide citations behind a tap. Signals that push toward memory: no links, hedgy dating (“as of my last update”), and details that match your business two years ago. Signals that push toward lookup: links, today’s phrasing from your own pages, mention of recent reviews.
A three-question self-test you can run in five minutes, from our own routine:
- Ask the assistant a question about your trade in your city and check whether the answer shows sources.
- Ask it something that changed about your business recently (new hours, new location). A correct answer almost certainly came from a lookup.
- Ask the same question with search or web access turned off, where the platform allows it. The difference between the two answers is the difference between its memory of you and your live footprint.
That third answer can be unsettling, and it’s worth making this check a deliberate routine rather than a one-off scare: one asking tells you almost nothing on its own.
Why lookup answers are winnable in months, not years
Memory only changes when a new model trains, on the vendor’s schedule, not yours. The training pile is closed to you: nothing you publish today reaches it for months, and no one outside the AI companies controls that timeline.
Lookup sources move at the speed of the web. Reviews accumulate into a readable track record in weeks. A Google Business Profile edit is live in minutes. A rewritten service page, built so an AI can actually quote it, gets recrawled in days. When an assistant builds its answer from those sources, your improvements are in the pipeline the next time someone asks.
Where each platform sits on this spectrum
Checked against each vendor’s published documentation on July 17, 2026. Platform behavior changes; this table carries a changelog and gets rechecked.
| Platform | Lookup behavior, per its own documentation |
|---|---|
| Perplexity | Lookup-first by design: “searches the internet in real-time” with numbered citations on answers3 |
| Google AI Overviews and AI Mode | Built on top of Google’s live index; issues multiple related searches per question2 |
| ChatGPT | Searches the web when it judges the question needs it, or when the user asks, with linked sources1 |
| Claude | Web search available with direct citations; conversations can also run without it4 |
How often each platform chooses to search for a given local question is behavior the vendors don’t publish. We label that plainly: the table above is what’s documented; the trigger rates are not, and anyone quoting exact ones is guessing.
What this means for where you spend effort
Retrieval-facing work first. Reviews, profile completeness, consistent listings, pages that answer real customer questions: these feed the pipeline every major assistant reads before it answers, and that pipeline refreshes constantly. Memory improves as a slow side effect of the same work, as durable mentions accumulate toward future training runs.
The practical shape of this is a rate, not a rank: how often you get named across many askings, tracked over time so movement is visible. A business that starts fixing lookup sources this month is competing again in weeks. A business waiting to be remembered is waiting on someone else’s training schedule.