# AI answer traffic > AI answer traffic is a fetch made while an assistant is answering somebody's question — the model decided a page was worth reading and went and read it, with a person waiting. It is distinct from a training crawl, which collects material for a model that ships months later, and from search indexing, which builds a list for later retrieval. **Category:** AI visibility **Also known as:** AI assistant fetch, answer engine traffic, live retrieval **Updated:** 2026-09-09 **Source:** https://truestat.io/glossary/ai-answers --- ## Why this is the category worth watching Of the three kinds of AI crawler traffic, this is the one with a person attached. Somebody asked a question, the assistant judged your page relevant enough to open, and in most products the answer that follows shows the source and links to it. That makes the shape of the traffic informative in a way a training crawl is not. Training crawls sweep everything once; answer fetches concentrate on the pages people are actually asking about. Reading them by page tells you which of your pages the market considers the reference on a subject. ## What it does not tell you Whether the answer was read, whether your page was the source that got cited, and whether the person clicked through. Nothing observable at your server distinguishes a fetch that produced a citation from one the model discarded, and any product claiming otherwise is inferring. What is observable: which pages were fetched, how often, by which assistant, and whether the request succeeded. A page that answer-fetchers keep requesting and keeps returning 404 is telling you something concrete. ## The tokens in this category `ChatGPT-User` and `ChatGPT-Agent` from OpenAI, `Claude-User` and `Claude-Code` from Anthropic, `Perplexity-User`, `Copilot` and `BingPreview` from Microsoft, several Google fetchers including `Google-NotebookLM` and `Gemini-Deep-Research`, plus `DuckAssistBot`, `MistralAI-User`, `Kimi-User`, `Amzn-User`, `meta-externalfetcher`, `Qwen-User` and xAI's `Grok-DeepSearch`. Each is listed separately in the crawler directory, because each carries its own token and can be allowed or refused on its own. ## Reading the pattern rather than the total The total is the least interesting number this category produces. Two sites with the same count can be in completely different positions depending on how it is distributed. A flat distribution across many pages usually means the assistants are crawling rather than answering — the fetches are following links, not questions. A concentrated one, where a handful of pages take most of the requests, means those pages are being treated as the reference on their subject. That is the signal worth having, and it is invisible in a total. Two secondary patterns are worth watching. Repeated fetches of the same URL over days suggest the page is being returned for a recurring question rather than a one-off. And fetches that return 404 are the most actionable thing in the whole category: an assistant asking repeatedly for a URL you removed is telling you, precisely, what it expected to find and where it expected to find it. ## What goes wrong - Counting search-index crawlers here. `OAI-SearchBot` and `Claude-SearchBot` are indexing, not answering, and both operators document them that way — an easy mistake that inflates the headline figure. - Counting link-preview bots as AI. `facebookexternalhit`, `Twitterbot` and `Slackbot` fetch to draw a card, and every share makes one, so folding them in inflates the number visibly. - Reading a burst as a runaway crawl. A deep-research session fetches ten to fifty sources for one question; that is the normal shape, not an incident. ## Worth doing - Read this figure by page rather than in total — the distribution is the finding. - Watch the 404s specifically. An assistant repeatedly asking for a URL you removed is describing what it expected to find. ## Questions ### Does AI answer traffic bring visitors? Sometimes. Assistants that cite sources link them, and some readers follow. But the fetch itself is not a visit, and a report that counts it as one is describing a different thing from a person arriving on your site. ### Why does my analytics not show any of this? Because browser analytics runs JavaScript and these crawlers do not. They request the HTML and leave, so the tracking tag is never executed. Seeing them at all requires reading requests server-side. ### How is this different from a referral? A referral is a person arriving on your site with a source attached. This is a machine fetching your page on a person's behalf, and the person may never arrive at all. The two can be related — an assistant that cites you sometimes sends the reader through — but counting a fetch as a visit describes something that did not happen. ## Related - https://truestat.io/glossary/ai-training-crawler - https://truestat.io/glossary/zero-click - https://truestat.io/glossary/server-side-tracking - https://truestat.io/glossary/geo-generative-engine-optimization --- From the TrueStat glossary — https://truestat.io/glossary. Privacy-first web analytics that also shows you which AI assistants are reading your site.