When an AI engine mentions your brand, it is drawing from one of two completely different sources: real-time search (it fetched a live page during the session) or training memory (it learned about you during training and is reciting from weights). These are not the same channel and they do not respond to the same optimization. Search-path citations are won with crawlability and extractable structure. Memory-path mentions are won with entity strength and repetition across the corpus over time. Optimizing for the wrong one is why effort produces no movement. This is how to diagnose which path your brand is on, per engine, and what to do about each.
Most operators treat "getting mentioned by AI" as a single objective with a single set of tactics. It is two objectives. A model can talk about your brand because it just read your site thirty seconds ago, or because your brand was common enough in its training data that it carries a representation of you in its weights. Those are different mechanisms with different failure modes, different timelines, and different fixes. The reason a lot of AEO effort goes nowhere is that the operator is pulling the search-path lever while the engine is answering from memory, or the reverse.
The distinction is not academic. It changes what you build, how long you wait, and how you measure. Before you spend another cycle on AI visibility, you need to know which path each engine is using for your brand, because the answer is often different across engines for the same company.
The two paths, mechanically
The search path is retrieval-augmented generation. When you ask ChatGPT with browsing, Perplexity, or Claude with web search a question, the engine issues live queries, fetches pages, and grounds its answer in what it just retrieved. The brand it mentions is the brand it found in the live results. This sits at the top of the layered pipeline an LLM uses to pick sources: retrieval happens at session time, against the live web.
The memory path is the model answering from its parameters. No fetch happens. When you ask a base model a question without browsing enabled, or when the engine decides the question does not warrant a live search, it answers from what it learned during training. The brand it mentions is the brand that was represented strongly enough in the training corpus to survive into the weights. This is a fundamentally different selection process: it was decided months or years ago when the model was trained, and nothing you publish today changes it until the next training cycle ingests your content.
The critical fact most operators miss: the same engine uses different paths for different questions. A specific, current, or local question triggers search. A general or well-established question gets answered from memory. So your brand might be winning the search path for "best X tool 2026" while being completely absent from the memory path for "what companies do X," because the first triggers a live fetch and the second does not.
A concrete way to feel the difference: ask any browsing-capable engine "what is the current pricing for [your product]" and it will almost always fetch, because pricing is volatile and the model knows it should not trust stale memory for it. Then ask "who are the main players in [your category]" and watch it answer instantly with no fetch, listing brands from memory. The first question is a search-path question by nature. The second is a memory-path question by nature. If you only appear in answers to the first kind, you have built search presence with zero memory presence, and the moment someone asks the category question without triggering a fetch, you do not exist. That asymmetry is invisible until you test both question types deliberately, which is exactly why so many brands believe they have "AI visibility" when they have only half of it.
How to diagnose which path you are on
The diagnostic is simpler than it sounds. The presence or absence of live source citations is the primary tell. When an engine answers from the search path, it shows its sources: Perplexity lists them, ChatGPT shows the browse activity and links, Claude cites the pages it fetched. When an engine answers from memory, there are no live citations because nothing was fetched. The answer just appears, fluent and source-free.
Run the same brand question two ways. First, ask it in a context that forces search, with browsing explicitly on and a current-flavored phrasing. Note whether you appear and whether there are citations. Then ask the same factual question in a way that invites a memory answer: general, timeless phrasing, browsing off if the interface allows. If you appear in the first but vanish in the second, you are winning search and losing memory. If you appear in the second with no citations, you have genuine memory presence, which is rarer and more valuable.
There is a second tell worth checking. If the engine describes your brand with details that are outdated, that is a memory-path signal, because the search path would have fetched current information. Stale details mean the model is reciting an old representation from training, which is its own problem with its own fix.
The optimization paths do not overlap
This is where conflating the two paths costs you. The tactics are almost disjoint.
| Search path (RAG) | Memory path (training) |
|---|---|
| Crawlability: content in raw HTML, not client-rendered | Entity strength: a recognized, well-defined entity |
| Extractable passages that stand alone | Repetition across many independent sources |
| Fresh, current pages with valid dates | Consistency of facts across the corpus over time |
| Valid schema for clean parsing | High-authority external corroboration |
| Feedback in days | Feedback in training cycles (months) |
The search path rewards technical readiness. Ship content that a non-rendering crawler can read, structure it for clean extraction, keep it fresh, and you can move search-path citations in days. Serving clean content to the crawlers via content negotiation is a search-path tactic with a fast feedback loop.
The memory path rewards something you cannot rush: being a strong, consistent, widely-corroborated entity over time. You earn memory presence by being mentioned consistently across enough independent high-authority sources that the next training run encodes you. The contrarian point is that you largely cannot optimize the memory path on demand. You can set the conditions, then wait for a training cycle. Anyone promising fast memory-path results is selling search-path tactics under the wrong label.
What to do with the diagnosis
Once you know your path per engine, the strategy writes itself. If you are losing search, that is the fast fix and where to start, because the feedback loop is days, not months. Get crawlable, get extractable, get fresh. If you are losing memory but winning search, accept the timeline: set the entity and corroboration conditions now, and understand the payoff lands on the next training cycle. If a model is reciting stale facts from memory, that is the re-grounding problem, where you push fresh signals through trusted channels and wait for re-ingestion.
Track them separately. A single "AI visibility" number hides which path is moving. The full set of AEO signals splits cleanly along this line, and reading them per-path tells you whether your effort is landing where you aimed it. The brands that get this right run two programs in parallel: a fast search-path program measured weekly, and a slow memory-path program measured across model releases. Most operators run one program, point it at the wrong path, and conclude AEO does not work.
It works. You were just pulling the wrong lever. And the temporal dimension matters more than people expect, because the way training corpora handle temporal validity and deduplication determines whether your repeated signals actually compound into memory presence or get collapsed into a single deduplicated mention. Know your path, measure it on its own clock, and stop expecting memory-path results from search-path work.
Sources
- OpenAI, ChatGPT search: when and how ChatGPT decides to browse versus answer from training. help.openai.com
- Anthropic, Claude web search: how Claude grounds answers in live retrieval versus training knowledge. docs.claude.com
- OpenAI Cookbook: retrieval-augmented generation patterns and when retrieval is triggered. cookbook.openai.com
- Website AI Score, the Citation Stack: where search-path retrieval sits in the source-selection pipeline. View article
- Website AI Score, temporal validity: how training corpora handle freshness and deduplication. View article
