For the complete documentation index, see llms.txt. Every page on this site is also served as Markdown: append `.md` to any URL, or send `Accept: text/markdown`.

ChatGPT vs Claude for brand visibility

OpenAI and Anthropic both split crawling across several user agents, but they split it along different lines. What each one publishes, and what follows for your robots.txt.

Last updated: 2026-09-03By Karl-Gustav Kallasmaa
openai.com logo

ChatGPT

by OpenAI

OpenAI's assistant, whose crawler documentation names four distinct user agents covering search, training, ads validation and user-initiated fetches.

Checked 2026-09-03
claude.com logo

Claude

by Anthropic

Anthropic's assistant, whose crawler documentation names three user agents covering training, user-initiated fetches and search-quality work, with a stated commitment to honour robots.txt.

Checked 2026-09-03

Which one should you choose?

The two vendors have converged on the same idea — one agent per job, so a site can refuse training without refusing to be findable — but they document it with different levels of precision. OpenAI attaches a consequence to each opt-out and publishes IP ranges per agent. Anthropic states a general commitment to honour robots.txt directives and warns that IP-based blocking may not work reliably.

Choose ChatGPT when

Prioritise the OpenAI side when you need to reason about consequences precisely. OpenAI states what opting out of each agent means, publishes a per-agent IP list, and documents roughly how long a robots.txt change takes to take effect, which makes the whole thing auditable.

Choose Claude when

Prioritise the Anthropic side when the decision you care about is training versus everything else. Anthropic states it will respect industry-standard robots.txt directives, supports the non-standard Crawl-delay extension, and separates its training crawler from its search-quality and user-initiated agents by name.

When neither is the right answer

If you are choosing between them as marketing channels rather than as crawl decisions, neither answers the question. Both publish what their crawlers do and neither publishes how a brand comes to be named in an answer, so channel priority has to come from where your own audience actually asks, not from either vendor's documentation.

What is specific to this comparison

  • OpenAI attaches a consequence to the opt-out — a site opted out of OAI-SearchBot will not be shown in ChatGPT search answers — while Anthropic's crawler article names Claude-SearchBot's purpose without stating what blocking it costs a site.
  • Anthropic actively discourages the IP-blocking approach, warning it may not persistently guarantee an opt-out, whereas OpenAI publishes a separate IP list per agent precisely so that operators can verify traffic.
  • Anthropic documents support for the non-standard Crawl-delay directive; OpenAI's crawler page documents no equivalent rate control, offering only allow or disallow per agent.
  • OpenAI documents a fourth agent, OAI-AdsBot, for validating pages submitted as ads on ChatGPT, and states its data is not used for training; Anthropic documents no advertising-related crawler at all.
  • The user-initiated agents diverge in their robots.txt status: OpenAI hedges that rules may not apply to ChatGPT-User, while Anthropic places Claude-User under its general commitment to honour robots.txt directives.

ChatGPT vs Claude, criterion by criterion

Crawling
Number of user agents documented
YesFour — search, training, ads validation and user actionsSource, checked 2026-09-03
Crawling
Agent governing appearance in the assistant's search
YesOAI-SearchBot, which surfaces sites in ChatGPT search featuresSource, checked 2026-09-03
Control
Stated consequence of blocking the search agent
YesThe site will not be shown in ChatGPT search answersSource, checked 2026-09-03
Control
Separate agent for model training
YesGPTBot, for making foundation models more useful and safeSource, checked 2026-09-03
Control
Does the user-initiated agent obey robots.txt
PartialChatGPT-User — robots.txt rules may not applySource, checked 2026-09-03
Control
Stated position on robots.txt overall
PartialPer-agent directives, with a documented exception for user actionsSource, checked 2026-09-03
Operations
Guidance on blocking by IP address
YesPer-agent IP lists published for verificationSource, checked 2026-09-03
Operations
Support for Crawl-delay
Not documentedNot documented on the crawler pageSource, checked 2026-09-03
Operations
Documented propagation time for a robots.txt change
YesAround 24 hours, as of 3 September 2026Source, checked 2026-09-03
Citations
Citation obligation on products built with the vendor's search
YesInline citations must be visible and clickableSource, checked 2026-09-03
Crawling
Separate agent for advertising safety checks
YesOAI-AdsBot, whose data is not used for model trainingSource, checked 2026-09-03

The same idea, documented differently

OpenAI and Anthropic have arrived at the same architecture: one crawler per job, so that a site owner can say yes to one use of their content and no to another. Both publish the list. That is genuinely useful and it was not the default a few years ago.

Where they differ is in what the documentation commits to. OpenAI writes down consequences and timings. Anthropic writes down a principle and a warning. Both approaches leave you able to act; they leave you able to verify to different degrees.

OpenAI's four agents

The OpenAI crawler documentation names four.

OAI-SearchBot is "used to surface websites in search results in ChatGPT's search features",[^oai-searchbot-purpose] and the consequence of opting out is stated plainly: those sites "will not be shown in ChatGPT search answers."[^oai-searchbot-optout]

GPTBot is described as "used to make our generative AI foundation models more useful and safe",[^oai-gptbot-purpose] with disallowing it indicating that content should not be used in training those models.

OAI-AdsBot is "used to validate the safety of web pages submitted as ads on ChatGPT", and OpenAI states that the data it collects is not used to train foundation models.[^oai-adsbot]

ChatGPT-User covers certain user actions in ChatGPT and Custom GPTs, with the caveat that "because these actions are initiated by a user, robots.txt rules may not apply."[^oai-chatgpt-user-robots]

Around all four sits an operational layer: a separate IP list per agent[^oai-ip-ranges] and a stated propagation time of roughly 24 hours for a robots.txt change to be picked up.[^oai-robots-lag]

Anthropic's three agents

Anthropic's crawler support article names three.

ClaudeBot "helps enhance the utility and safety of our generative AI models by collecting web content that could potentially contribute to their training."[^anthropic-claudebot-purpose] That is the training crawler, and the phrasing is careful: content that could potentially contribute, rather than content that will.

Claude-User "supports Claude AI users. When individuals ask questions to Claude, it may access websites using a Claude-User agent."[^anthropic-claude-user] This is the user-initiated fetch, and it is where the two vendors' documentation most clearly parts company: Anthropic places it under the same robots.txt commitment as everything else, while OpenAI's equivalent carries an explicit hedge.

Claude-SearchBot "navigates the web to improve search result quality for users", analysing online content to enhance the relevance and accuracy of search responses.[^anthropic-searchbot]

The commitment underneath them is a single sentence: Anthropic says it will "respect 'do not crawl' signals by honoring industry standard directives in robots.txt", and gives the User-agent: ClaudeBot / Disallow: / form as the example.[^anthropic-robots-commitment] It also supports the non-standard Crawl-delay extension,[^anthropic-crawl-delay] which is a rate control OpenAI does not document.

And it adds a warning that most vendors leave implicit: alternate methods such as blocking IP addresses "may not work correctly or persistently guarantee an opt-out."[^anthropic-ip-blocking-warning]

The asymmetry that actually matters

Read the two documents together and one gap stands out.

OpenAI tells you what you lose. Block OAI-SearchBot and your site will not appear in ChatGPT search answers — a sentence you can quote to a security team that wants to block everything, and a sentence you can plan around.

Anthropic tells you what each agent does but not what blocking each one costs. Claude-SearchBot exists "to improve search result quality"; the article does not say that blocking it removes a site from anything. That is a documentation gap rather than a behavioural claim, and this page will not fill it with a guess. If you are deciding whether to block Claude-SearchBot, you are deciding without a published consequence, and that is worth knowing before you make the decision rather than after.

The reverse asymmetry is on IP handling. OpenAI's per-agent IP lists make verification straightforward: you can check whether a request claiming to be GPTBot came from an OpenAI address. Anthropic offers no equivalent verification list on the crawler page and actively discourages the IP-based approach to blocking. If your operations team's instinct is to solve crawler questions at the firewall, one vendor supports that instinct partially and the other tells you it will not hold.

What both of them do the same way

Both separate training from everything else, and both make that separation addressable by name. This is the single most useful fact on the page, and the most commonly discarded.

A site that adds User-agent: * / Disallow: / to protect against AI training has also blocked the agents that make it findable. On the OpenAI side that has a documented cost. On the Anthropic side it removes the search-quality crawler along with the training one. Neither vendor requires that trade, and neither benefits from it.

The correct move is boring: name the agents individually, decide training and findability separately, and write the file so that a future reader can see which decision was which.

Both vendors also attach citation behaviour to their developer search products. OpenAI's web search guide requires that "inline citations must be made clearly visible and clickable in your user interface."[^oai-citation-display] Anthropic states that citations are always enabled for web search, each carrying the source URL, the title and up to 150 characters of the cited content.[^anthropic-citations-always] Different mechanisms, same direction of travel: products built on either vendor's search are expected to show where the material came from.

A robots.txt that expresses the actual decision

If the decision is "be findable, do not be training data", the tokens involved are OAI-SearchBot and Claude-SearchBot on the allow side, GPTBot and ClaudeBot on the disallow side, with ChatGPT-User and Claude-User handled as the user-initiated cases they are and OAI-AdsBot decided only if you run ads on ChatGPT.

Two operational notes. OpenAI's roughly 24-hour propagation means a change is not wrong just because it has not taken effect within the hour.[^oai-robots-lag] Anthropic publishes no equivalent window,[^anthropic-no-propagation-window] so the only honest procedure there is a repeating check rather than a deadline.

The user-initiated fetch is the hardest case on both sides

Of the seven agents across the two vendors, the two that cause the most confusion are the ones triggered by a person: ChatGPT-User and Claude-User.

The confusion is reasonable, because the situation is genuinely ambiguous. A robots.txt file expresses a site's policy toward automated crawling. When a human being asks an assistant to read a specific page, the request is not really automated crawling — it is closer to that person opening the page in a browser, with software in between. Both vendors have had to take a position on that, and they have taken subtly different ones.

OpenAI's position is a hedge: because the actions are initiated by a user, robots.txt rules may not apply. The word doing the work is "may". It tells you the rules are not guaranteed to be honoured, and it does not tell you when they are.

Anthropic's crawler article takes a simpler line, listing Claude-User among the agents covered by its general commitment to honour industry-standard robots.txt directives, without carving out an exception for user-initiated requests.

For a site owner, the practical guidance is the same either way, and it is uncomfortable: a robots.txt disallow is a policy signal, not an access control. If a page must not be readable by a third party under any circumstances, the mechanism for that is authentication, not a directive in a text file that some agents describe themselves as possibly ignoring. Conversely, if your goal is visibility rather than protection, this whole category of agent is on your side — a person asked for your page by name, which is the closest thing to intent that exists in this stack.

Both vendors deserve some credit for documenting the ambiguity rather than leaving site owners to discover it in their logs. Neither has resolved it, and neither is likely to, because the underlying question is a policy question rather than a technical one.

What neither vendor publishes

Neither document explains how a brand comes to be named in an answer. Both explain retrieval permissions. Neither explains selection.

That boundary is worth being pedantic about, because the gap between "my pages are crawlable" and "my brand is mentioned" is where most of the money in this category is spent. The first is a documented, verifiable, one-afternoon fix. The second is unpublished on both platforms and observable only by sampling real prompts repeatedly and recording the date, the prompt set and the sampling depth alongside every number.

For the case where one platform gates answers behind its own crawler and the other inherits eligibility from ordinary search, see Google AI Overviews versus ChatGPT search. For a platform whose user-initiated fetcher explicitly ignores robots.txt, see ChatGPT search versus Perplexity. And for what each vendor hands a developer once a source has been used, see Claude versus Gemini on citations.

The limit of this page

Everything above is drawn from OpenAI's crawler and web search documentation and Anthropic's crawler support article and web search tool documentation, all fetched on 3 September 2026. Both vendors revise these pages without notice, which is why every claim carries its retrieval date and why this page is reviewed quarterly rather than treated as settled.

Where a vendor has not published something — what blocking Claude-SearchBot costs, how long an Anthropic robots.txt change takes, whether OpenAI honours Crawl-delay — this page says so instead of filling the gap. A comparison that guesses at the missing cells reads better and is worth less.

[^oai-searchbot-purpose]: OpenAI crawler documentation, retrieved 2026-09-03. [^anthropic-claudebot-purpose]: Anthropic crawler support article, retrieved 2026-09-03.

Where Attensira fits, and where it does not

Attensira is relevant only once both robots.txt decisions are made and you need repeated, dated sampling of whether your brand is actually named in answers on each platform — with the fix arriving as a pull request against your own repository for review rather than as a recommendation in a report.

See how Attensira compares to both

Questions people ask

For OpenAI, the documented agents are OAI-SearchBot, GPTBot, OAI-AdsBot and ChatGPT-User. For Anthropic, they are ClaudeBot, Claude-User and Claude-SearchBot. Seven tokens across two vendors, each with a different job. A single wildcard rule cannot express the decision most teams actually want, which is yes to being findable and no to being training data.

Not according to what is published. Anthropic documents ClaudeBot as collecting content that could contribute to training, and documents Claude-SearchBot separately as navigating the web to improve search result quality. They are different tokens. What Anthropic does not state is what blocking Claude-SearchBot costs a site, which is the one place OpenAI's documentation is more explicit than Anthropic's.

Anthropic advises against it, stating that alternate methods such as blocking IP addresses may not work correctly or persistently guarantee an opt-out. OpenAI publishes per-agent IP lists, but their documented purpose is verification rather than blocking. The reliable control on both sides is robots.txt; the IP lists are for confirming that a request claiming to be a given crawler really was one.

On the Anthropic side, yes — the crawler article states support for the non-standard Crawl-delay extension in robots.txt. OpenAI's crawler documentation, as retrieved on 3 September 2026, documents no equivalent rate control, so the choice there is allow or disallow.

OpenAI documents roughly 24 hours between a robots.txt update and its systems adjusting, as of 3 September 2026. Anthropic's crawler article publishes no equivalent window, so the honest procedure on that side is to re-check on a schedule rather than to expect a change by a particular time.

Neither vendor publishes anything that would answer that. Both document what their crawlers do and how to allow or block them, and both document citation behaviour in their developer APIs. Neither documents how a brand comes to be named in a generated answer. Pick the platform your audience actually uses, satisfy both sets of crawler rules since they are cheap and independent, and measure rather than theorise.

Sources

Every claim on this page, with the page it came from and the date that page was read. Prices and feature lists change; these are what the source said on the date shown, not timeless facts.

  1. OpenAI documents OAI-SearchBot as the user agent used to surface websites in search results in ChatGPT's search features.used to surface websites in search results in ChatGPT's search featureshttps://developers.openai.com/api/docs/bots — read 2026-09-03
  2. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answershttps://developers.openai.com/api/docs/bots — read 2026-09-03
  3. OpenAI documents GPTBot as used to make its generative AI foundation models more useful and safe, and states that disallowing it indicates content should not be used in training those models.used to make our generative AI foundation models more useful and safehttps://developers.openai.com/api/docs/bots — read 2026-09-03
  4. OpenAI documents ChatGPT-User for certain user actions in ChatGPT and Custom GPTs, and notes that because these actions are initiated by a user, robots.txt rules may not apply.Because these actions are initiated by a user, robots.txt rules may not applyhttps://developers.openai.com/api/docs/bots — read 2026-09-03
  5. OpenAI documents OAI-AdsBot as used to validate the safety of web pages submitted as ads on ChatGPT, and states its data is not used to train generative AI foundation models.used to validate the safety of web pages submitted as ads on ChatGPThttps://developers.openai.com/api/docs/bots — read 2026-09-03
  6. OpenAI publishes separate IP range lists for each documented agent, including openai.com/searchbot.json, openai.com/gptbot.json, openai.com/adsbot.json and openai.com/chatgpt-user.json.https://developers.openai.com/api/docs/bots — read 2026-09-03
  7. OpenAI documents that it can take roughly 24 hours from a robots.txt update for its systems to adjust.it can take ~24 hours from a site's robots.txt update for our systems to adjusthttps://developers.openai.com/api/docs/bots — read 2026-09-03
  8. OpenAI's web search guide states that when displaying web results to end users, inline citations must be made clearly visible and clickable in the interface.inline citations must be made clearly visible and clickable in your user interfacehttps://developers.openai.com/api/docs/guides/tools-web-search — read 2026-09-03
  9. Anthropic documents ClaudeBot as collecting web content that could potentially contribute to training, to help enhance the utility and safety of its generative AI models.helps enhance the utility and safety of our generative AI models by collecting web content that could potentially contribute to their traininghttps://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03
  10. Anthropic documents Claude-User as supporting Claude AI users, stating that when individuals ask questions to Claude it may access websites using a Claude-User agent.supports Claude AI users. When individuals ask questions to Claude, it may access websites using a Claude-User agent.https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03
  11. Anthropic documents Claude-SearchBot as navigating the web to improve search result quality, analysing online content to enhance the relevance and accuracy of search responses.navigates the web to improve search result quality for usershttps://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03
  12. Anthropic states it will respect do-not-crawl signals by honouring industry standard directives in robots.txt, and gives a User-agent ClaudeBot Disallow example.respect 'do not crawl' signals by honoring industry standard directives in robots.txthttps://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03
  13. Anthropic states that alternate methods such as blocking IP addresses may not work correctly or persistently guarantee an opt-out.alternate methods like blocking IP address(es)...may not work correctly or persistently guarantee an opt-outhttps://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03
  14. Anthropic states that it supports the non-standard Crawl-delay extension to robots.txt.https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03
  15. Anthropic states that citations are always enabled for its web search tool, with each citation carrying the source URL, the title and up to 150 characters of the cited content.Citations are always enabled for web searchhttps://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool — read 2026-09-03
  16. As retrieved on 3 September 2026, Anthropic's crawler support article does not state how long a robots.txt change takes to take effect, where OpenAI's crawler documentation does.https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler — read 2026-09-03