Spoke
How each engine actually behaves
One row per engine, each with its own last-verified date and a link to the platform's own documentation. Where a row could not be verified, it says so instead of being filled in from memory.
Every filled row is taken from the named platform's own documentation and carries the date it was retrieved. The one row that could not be verified is published empty rather than guessed.
Generative platforms do not behave alike, and advice that treats "AI search" as one system will be wrong about at least one of them. Some retrieve first and show inline citations on nearly every answer. Some answer from the model unless retrieval is invoked. The controls differ, and so does what each control actually does.
This table is published as crawlable HTML, not as an image or a PDF, for the reason given in the comparative content spoke: a table a machine cannot read is a table that cannot be cited.
Which AI engines cite sources?
All of the engines verified here surface links alongside generated answers when they have retrieved live pages, though how prominently varies. The more useful distinction is not whether an engine cites but whether it retrieves at all for a given question — an answer drawn from training data has nothing to link to.
The table below records what each platform documents about its own crawlers, which is the part you can act on.
| Engine | How it gets content | Live web access | Citation display | Documented user agents | What blocking actually controls | Last verified |
|---|---|---|---|---|---|---|
| ChatGPT (OpenAI) | Both. Model-native knowledge from training, plus live retrieval for search answers and user-triggered fetches. | Yes, through OAI-SearchBot for search features and ChatGPT-User for user-triggered visits. | Inline links in search answers. | OAI-SearchBot, GPTBot, OAI-AdsBot, ChatGPT-User | The four are independent. Blocking GPTBot signals only that content should not train foundation models. Only OAI-SearchBot governs whether a site can appear in ChatGPT search answers; opted-out sites may still appear as navigational links. ChatGPT-User is user-triggered and robots.txt rules may not apply to it. | 2026-11-15 — source |
| Google AI Overviews and AI Mode | Retrieval over the Google Search index. | Yes. The same index that serves Search. | Supporting links shown alongside the generated answer. | Googlebot (Search). Google-Extended is a separate token and is not a Search crawler. | Google states there are no additional requirements to appear: a page qualifies as a supporting link if it is indexed and eligible to be shown in Search with a snippet. The preview controls that limit what is shown are nosnippet, data-nosnippet, max-snippet and noindex. Google-Extended is NOT a Search control -- it governs training and grounding in other Google systems, and Google states it does not impact a site's inclusion in Google Search nor is it used as a ranking signal. | 2026-11-15 — source |
| Perplexity | Retrieval-first. | Yes. | Inline numbered citations on essentially every answer. | PerplexityBot, Perplexity-User | PerplexityBot surfaces and links sites in Perplexity results and is documented as not crawling content for AI foundation models. Perplexity-User handles user-initiated requests and generally ignores robots.txt, because a person triggered the fetch. | 2026-11-15 — source |
| Claude (Anthropic) | Both. Model-native knowledge, plus web search and user-triggered fetches. | Yes, through Claude-SearchBot and Claude-User. | Inline links when answering from web results. | ClaudeBot, Claude-SearchBot, Claude-User | ClaudeBot collects content that may contribute to training. Claude-SearchBot crawls to improve search quality, and disabling it may reduce a site's visibility and accuracy in search results. Claude-User accesses sites when a user's question requires it, and disabling it may reduce visibility for user-directed web search. Blocking is per user agent in robots.txt, and must be repeated for each subdomain. Blocking by IP is specifically discouraged, because the bots then cannot read robots.txt at all. | 2026-11-15 — source |
| Microsoft Copilot | Retrieval over Microsoft's own index. | Yes. | Citations shown in AI answers, and reported back to publishers. | Bingbot. PARTIALLY VERIFIED -- see control. | Microsoft reports publisher visibility in its AI experiences through an AI Performance report in Bing Webmaster Tools, covering citations, cited pages and the queries associated with retrieval. Microsoft states that citation activity is not a measure of ranking, authority or page importance. The crawler-control details for Microsoft's AI experiences were still not readable from Microsoft's own documentation when this row was updated, so that cell remains partially verified rather than filled in from secondary coverage. | 2026-11-15 — source |
How to read the "what blocking actually controls" column
That column is the one worth slowing down for, because it is where most published advice is wrong.
Every platform in this table that documents more than one crawler documents them as independent. There is no single "AI opt-out." There is a training control, a search control, and usually a user-triggered fetcher that the other two do not govern.
The user-triggered fetchers deserve particular attention. On more than one platform they are documented as generally ignoring robots.txt, on the reasoning that a person explicitly asked for that page. If your mental model is "robots.txt is how I control AI access", that model is incomplete.
The stale-row problem
This table is the fastest-decaying page on the site. Crawler names change, controls get added, and a row that was accurate in November can mislead by March.
So the verification date is part of the data rather than a footnote, and the build reports any row older than the review interval. When you read a row, read its date first. A confident table with no dates on it is telling you nothing about whether it is still true.
More in GEO & entities
How AI engines retrieve
Retrieval is what lets a model quote something you published this morning. Understanding it tells you which levers exist — and which widely repeated tactics are aimed at nothing at all.
What Google documents
Google published a guide to showing up in its AI features, and a big chunk of it is a list of popular tactics it says you don't need. That makes it the most useful source in this field, and the most uncomfortable one.
Training vs retrieval
One of these you can change this afternoon. The other changes when somebody else trains a model. Knowing which is which tells you where effort is worth spending and where it is wasted.
Entity consistency
The highest-value work in getting quoted by AI is also the dullest. Describe your company the same way everywhere, give it one stable identifier, and stop handing machines reasons to hedge about who you are.
Comparative content
Comparison questions are among the most valuable a buyer asks and among the most common an AI engine is given. The format decides whether you can be quoted — a table in HTML can be, the same table as an image cannot.