Two ways an AI can know something
An AI assistant can answer a question about your company in two fundamentally different ways. It can answer from memory: whatever it absorbed about you during training, frozen at the point its training data was collected. Or it can answer from search: look up the live web at the moment of the question, read a handful of pages, and summarise them — usually showing you which pages it used.
Perplexity answers from search by design. ChatGPT, Gemini and Claude can do either, depending on the product, the plan and whether the question seems to need fresh information. When a model's API is called without web search, as many monitoring tools do, you are seeing memory only.
A real example
In September 2026 we asked both kinds of assistant about AutoChase, a young invoice-chasing tool. Asked without search, the model said it did not recognise the company. Perplexity, searching the web, described it correctly — automatic payment reminders for freelancers and small businesses — and cited autochase.app as its source.
Neither answer is a bug. The memory answer reflects that AutoChase wasn't prominent in the data the model was trained on. The search answer reflects that its website is live, crawlable and clear.
Why this matters for your strategy
The two channels respond to different work on very different timescales.
Search-backed answers can change within days. If an assistant can crawl your site, understands what you do, and finds independent pages that agree, it can start citing you soon after those pages exist. This is where crawlability, clear direct-answer pages, schema, and third-party listings pay off quickly.
Memory-based answers change only when a model is retrained on newer data, on the vendor's schedule. You can't speed that up, but you can make sure that when it happens, the web contains a consistent, well-corroborated description of you — in the places training data tends to draw from heavily, such as encyclopedic sources, established review sites, and widely discussed pages.
What to measure
Because the channels behave differently, measure them separately. Ask a search-backed assistant and check whether it cites your own site. Ask a model without search and check whether it knows you, and whether what it says is accurate. A brand that search finds but memory doesn't is on the right path; the reverse — known from memory but never cited by search — usually means a crawlability or content problem on your own site.
Anovox's AI Awareness audit runs both checks side by side, quotes each assistant's answer, and judges the memory answer against your own website so a confident description of the wrong company is caught rather than counted as a win.
See this on your own brand
Run the free check — same scoring the dashboard uses, no account needed.
Run free check