AI assistant

Two places an assistant gets its facts

The first is the model itself. A language model is trained on a large body of text and carries what it absorbed as weights; an answer given from memory alone reflects that text as it stood when training stopped. The second is retrieval: the assistant runs a search, fetches pages, and writes from them. The two are separate operations, and the companies running assistants separate them in their crawler tokens. OpenAI’s crawler documentation lists GPTBot for training data and OAI-SearchBot for ChatGPT search, and says the two are controlled independently. Perplexity’s crawler documentation states that PerplexityBot indexes for its search results and that neither of its bots collects training data.

For a business, the retrieval path is the one that matters. A shop that opened this year is not in anyone’s training data; it can still be named if its site is fetched and says something worth quoting.

What an assistant hands back

One answer, written for one person, containing a few names. When this site put one buyer question per business to Claude with web retrieval for 9,240 small businesses in September 2026, 14.9% of them were named, and of every name the assistant returned, 99.4% were businesses and 0.6% were directories or aggregators. A Claude number, one question each, one day; the full account is in what 9,483 small business websites serve to an AI crawler.