How AI assistants choose their sources
AI assistants choose sources in two stages: first a search over an index (their own, Google's, Bing's or a partner's) returns candidate pages; then the model reads them and cites the passages it uses. So a page is cited only if it is crawlable, ranks for the searches the assistant runs, and contains a clear answer.
Updated · Citable
Stage 1: retrieval. Stage 2: synthesis
Most assistants follow the same pattern, often called retrieval-augmented generation:
- The assistant decides the question needs fresh or specific information.
- It rewrites the question into one or more search queries, sometimes many (Google calls this "query fan-out").
- A search system returns candidate pages, ranked by its usual relevance and quality signals.
- The model reads the top candidates (or extracted passages) and writes an answer, attributing claims to the sources it used.
The consequence for brands: stage 1 is search engine ranking under another name. Stage 2 rewards pages whose passages answer the question clearly and agree with other sources.
Engine by engine
| Assistant | Where it searches | Crawler to allow |
|---|---|---|
| ChatGPT | OpenAI's own crawling plus third-party search providers (Bing is widely reported) | OAI-SearchBot |
| Perplexity | Its own index, built by its crawler, plus other sources | PerplexityBot |
| Google AI Overviews and AI Mode | Google Search index | Googlebot |
| Gemini | Google Search grounding | Googlebot (Google-Extended governs training and some grounding) |
| Claude | Its own search crawler plus a provider (Brave is widely reported) | Claude-SearchBot, Claude-User |
| Microsoft Copilot | Bing | Bingbot |
What gets cited
Observed across engines and in published studies:
- Pages that rank. The retrieval step uses search ranking, so authority and relevance still decide who is a candidate.
- Third-party sources for brand questions. "Is X good?" and "best X for Y" are answered from reviews, lists, forums and news more than from brand sites.
- Clear, factual passages. Definitions, numbers, steps, comparison tables. The GEO research found that adding statistics, quotations and source citations increased how much engines used a page.
- Freshness for anything time-sensitive: prices, "2026", recent launches.
- Agreement. Facts repeated across independent sources are stated confidently; outliers are dropped or hedged.
When there is no search: the model's memory
If the assistant answers without searching, it relies on training data. Which brands it knows, and how it describes them, reflects how often and where they were written about before the training cutoff. Training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot from Common Crawl) collect that data. A broad footprint of mentions across news, directories, documentation and forums is what gets a brand into future models.
What this means for you
- Allow the search crawlers: AI crawlers and robots.txt.
- Rank in both Google and Bing for your buyers' questions.
- Get named on the third-party pages each engine cites: brand mentions vs backlinks.
- Make your pages quotable: how to write content AI assistants cite.
- Measure per engine: check what AI says about your brand.
Questions people ask
Do AI assistants cite the top Google result?
Often, but not always. They read several results and cite the passages they use, so a page ranking fifth with a clearer answer can be cited over the first result.
Why do different assistants cite different sources for the same question?
They search different indexes, rewrite the question differently, and read different numbers of pages. Even the same assistant varies between runs.