Skip to content

Guide

How does AI search pick its sources?

Which crawlers feed ChatGPT, Perplexity, Claude and Google, and how much of an engine's choice of sources a small business can influence.

Published ยท 7 min read

In short

AI search picks its sources in two steps. A crawler first has to be allowed to fetch and index your page. Then, for each question, the engine runs its own searches, reads the pages that match and writes an answer that links the ones it drew on. You decide the first step in robots.txt; the second you can only influence, with plain, specific pages.

Definition

Grounding: Grounding is the step where an AI system looks up current web pages before it answers, instead of relying only on what it learned in training. One common way to do it is retrieval-augmented generation: the system runs searches, retrieves pages from an index, reads the passages that match the question and writes its reply from them, usually with links to the pages it used. Google describes its AI Overviews and AI Mode this way, and the search features in ChatGPT, Perplexity and Claude each run their own crawler for the same purpose.

The crawlers that feed each engine

In short

Each answer engine reads the web through its own crawler, and a site that blocks that crawler in robots.txt should expect to appear less, or not at all, in that engine's answers. Where an engine also collects pages for training, it uses a separate crawler, so a site can refuse training and still be found.

OpenAI says OAI-SearchBot is the crawler that surfaces websites in ChatGPT's search features, and that sites which opt out of it will not be shown in ChatGPT search answers, though they can still appear as navigational links. Its GPTBot collects pages that may be used to train models, and ChatGPT-User visits a page when a person asks ChatGPT to, so robots.txt rules may not apply to it. OpenAI adds that a robots.txt change can take about 24 hours to reach its systems.

Perplexity runs PerplexityBot to surface and link websites in its search results and says it is not used to crawl content for AI foundation models; its Perplexity-User fetcher follows a person's request and generally ignores robots.txt. Anthropic names ClaudeBot for collecting content for its models, Claude-User for pages a person asks Claude to read, and Claude-SearchBot for search, and says that blocking Claude-SearchBot may reduce a site's visibility in Claude's search results.

Google works differently. AI Overviews and AI Mode draw on the ordinary Search index that Googlebot builds, so there is no separate AI crawler to let in. Google-Extended is a control in robots.txt for training and grounding in Gemini models and apps, and Google says it does not affect a site's inclusion or ranking in Google Search.

Which crawler feeds search answers and which collects for training, by engine
Feeds its search answersCollects for training
ChatGPT (OpenAI)OAI-SearchBotGPTBot
PerplexityPerplexityBotNone named: Perplexity says PerplexityBot is not used to train models
Claude (Anthropic)Claude-SearchBotClaudeBot
Google AI Overviews and AI ModeGooglebot, the ordinary Search crawlerGoogle-Extended, a robots.txt control rather than a crawler

Swipe sideways to see every column

How an engine chooses among the pages it can read

In short

The engine turns a question into searches of its own, often several, and writes from the passages that answer them best. No engine publishes the rules it uses, and the same question can bring different sources on a different day.

Google's guide to its generative features says they are rooted in its core ranking and quality systems: a model retrieves pages from the Search index, reviews what they say and shows links to the pages that support its response. The same guide says it adds no special requirements of its own, so the pages it can cite are the pages Search already indexes.

Research points the same way. A study by researchers at Princeton and elsewhere, accepted to the KDD 2024 conference, rewrote pages in different ways and measured how much of each one a generative engine used. Adding citations to sources, quotations and statistics gave relative gains of 30 to 40 percent on one of its main measures, while keyword stuffing, common in classic SEO, gave little or no gain.

Those tests ran on a research setup built on a GPT-3.5 model and on Perplexity, not on every engine people use today, so treat them as a direction rather than a rule. A passage that states a specific fact, says where the fact comes from and answers the question without leaning on the paragraphs around it is easier for any engine to use.

What you control, and what you do not

In short

You control whether the crawlers may come in and how clearly your pages answer. You do not control which sources an engine picks for a given question, and no engine sells a place among the sources its answers cite.

The parts in your hands are practical. Keep your pages public and crawlable, with the words in the page itself rather than inside images, list them in a sitemap, and keep the facts about your business the same on your site and your listings.

Google's guide says there is no need to cut content into tiny chunks or to write in a special way for AI, and that Google Search ignores AI text files such as llms.txt. It warns that seeking inauthentic mentions across the web is not as helpful as it might seem, and that no third-party tool has access to its internal ranking or AI systems, so be wary of any tool or service that promises you a place in the answers.

Search Console shows part of the picture for Google. Since August 31, 2026 its generative AI performance report has been open to sites worldwide: it counts how often links to a site appeared in AI Overviews and AI Mode, page by page. A setting beside it lets an owner take a site out of those features, and every site stays in them unless its owner opts out.

FAQ markup and llms.txt: kept, not relied on

In short

Both cost little and neither is a lever. Google stopped showing FAQ rich results from May 7, 2026, and says Google Search ignores llms.txt files, so a site keeps them for other readers rather than for a place in Google's results.

FAQ markup labels a page's questions and answers as structured data. Google once showed it as expandable questions under a result, then only for a few government and health sites, and its documentation changelog says the feature no longer appears in Google Search starting May 7, 2026. The markup still describes the page truthfully, as long as it matches the questions a reader can see.

llms.txt is a proposed file that summarizes a site for AI tools. Ahrefs checked 137,000 domains and found that of the roughly 38,000 with a valid file, 97 percent saw no requests for it at all in May 2026. Google says keeping one will neither help nor harm a site in Google Search, so it is a cheap extra rather than a strategy.

What this means for a small business

In short

Let the search crawlers in, then spend your effort on pages that answer your customers' questions plainly. Measure it by asking the engines those questions yourself and noting which sources they name.

Open your robots.txt, or ask whoever runs your site, and check that OAI-SearchBot, PerplexityBot, Claude-SearchBot and Googlebot are allowed. If you would rather OpenAI, Anthropic and Google not train their models on your pages, you can block GPTBot, ClaudeBot and Google-Extended and still appear in ChatGPT search, Claude's search and Google Search, though Google-Extended also keeps your pages out of grounding in Gemini's own apps.

Then look at the questions a customer asks before calling you, such as what you charge, how long a job takes or whether you serve their part of town. A page that answers one of them in its first lines, in words a stranger understands, gives an engine something it can quote.

For a business that serves one city, such as a shop or a practice in Los Angeles, Google adds one thing: details in a Business Profile can feed its AI answers about local businesses. Keep the profile and the site in agreement, down to the hours and the services you list.

Common questions

Should I block AI crawlers from my site?

Block the ones that collect for training if you prefer, and allow the ones that feed search answers. Blocking OAI-SearchBot, PerplexityBot or Claude-SearchBot can keep your pages out of those engines' answers or make them rarer there, which a business seldom wants.

How long does a robots.txt change take to reach the engines?

OpenAI says its systems adjust about 24 hours after a robots.txt update. The other vendors' pages cited below give no figure, so allow a few days before you check.

Why do two engines cite different pages for the same question?

Each one runs its own crawler and index, turns the question into its own searches and makes its own choice. The same engine can also cite different pages on different days, so a single check proves little.

Can an engine cite a page that sits behind a login?

No. Crawlers read public pages, so anything behind a login or a form stays out of the answers, as it stays out of ordinary search. Put the facts customers ask about on public pages.

Is getting cited by ChatGPT different from ranking on Google?

The crawler is different and so is the index, but the same kind of page suits both: open to crawlers, specific and easy to quote. Google says its own AI features run on its ordinary Search ranking systems.

Can I pay to be one of the sources an answer cites?

No. No answer engine sells a place among the sources its answers cite, and none publishes the rules it uses to choose them. Anyone selling guaranteed placement is selling something the engines do not offer.

Sources

  1. Overview of OpenAI crawlers, OpenAI, read September 24, 2026
  2. Perplexity crawlers, Perplexity, read September 24, 2026
  3. Does Anthropic crawl data from the web, and how can site owners block the crawler?, Anthropic, April 7, 2026
  4. Google's common crawlers, Google Search Central, July 14, 2026
  5. Optimizing your website for generative AI features on Google Search, Google Search Central, July 10, 2026
  6. Generative AI performance report (Search), Google Search Console Help, August 31, 2026
  7. Search generative AI control, Google Search Console Help, August 31, 2026
  8. Latest Google Search documentation updates, Google Search Central, September 18, 2026
  9. We analyzed 137K sites: 97% of llms.txt files never get read, Ahrefs, June 15, 2026
  10. GEO: Generative Engine Optimization, arXiv (Aggarwal and others, KDD 2024), June 28, 2024

Start with a written plan

Tell us about the business and what you need. We reply with questions, then a written plan that lists the work and its price, and nothing starts until you have read it.