Which crawler is the question actually about?
“Does ChatGPT read press releases” contains a hidden ambiguity, because OpenAI does not operate one crawler. It operates three, each with a different job, and a site can allow or block them independently.
GPTBot, introduced in August 2023, collects content that may be used to train future models. Blocking it is a statement about training corpora and has no bearing on whether a release can be surfaced in a search result today.
OAI-SearchBot supports search functionality in ChatGPT, including the links and citations shown alongside answers. For anyone whose interest in a press release is whether it can be cited back to a user, this is the relevant agent.
ChatGPT-User covers fetches made because a user asked for something specific in a conversation. OpenAI's documentation notes that because these requests are initiated by a person rather than by automated crawling, robots.txt may not apply to them in the usual way.
The three publish separate IP ranges — at gptbot.json, searchbot.json and chatgpt-user.json on OpenAI's site — so a site operator can verify that a request claiming to be one of them genuinely is. The practical consequence is that “we block AI crawlers” and “we are invisible to ChatGPT” are different claims, and a site can easily be one without being the other.
Classification from OpenAI's published crawler documentation
Source: OpenAI, crawler documentation for GPTBot, OAI-SearchBot and ChatGPT-User.
Bars encode whether standard robots.txt exclusion is documented to govern the agent, not volume of crawling.
What do the wire services actually permit?
Rather than ask the wires, we read what they tell crawlers. On 18 September 2026 we fetched the robots.txt of PR Newswire, GlobeNewswire, Business Wire, AccessWire and EIN Presswire and searched each for named AI user agents.
None of the five named GPTBot, OAI-SearchBot, ChatGPT-User, CCBot or ClaudeBot in any directive. That absence is the finding, and it matters because of how the robots exclusion standard works: an agent with no rule addressed to it falls through to the wildcard User-agent: * block, and anything not disallowed there is permitted.
In the four files we could read, the wildcard rules restrict administrative and machine-readable plumbing rather than editorial content. PR Newswire disallows eight narrow paths covering page templates, a widget landing page and its multimedia directory. GlobeNewswire disallows its API, JSON data, stock-quote endpoints, embeddable widgets and a number of language directories. AccessWire disallows admin, user, account, API and RSS paths and sets a five-second crawl delay. EIN Presswire maintains named rules for Microsoft's msnbot and for Googlebot, and applies a thirty-second crawl delay to everyone else.
In no case is the news-release content itself placed out of bounds. GlobeNewswire makes the point explicit with a line reading Allow: /news-release/. On the stated permissions, press releases on these wires are open to OpenAI's crawlers by default.
Checked 18 September 2026; count of AI user agents named in any directive
Source: Media Morpheus original measurement, 18 September 2026, direct retrieval of each robots.txt.
All bars are zero: no wire named any AI crawler either to permit or to block it. Business Wire's file could not be read from our vantage point, so its contents are unknown rather than empty.
What could we not see?
Business Wire's robots.txt returned an HTTP 403 “Access Denied” from its content delivery network, both to a standard desktop browser user agent and, with a connection failure, to a request declaring itself GPTBot.
It would be easy and wrong to report that as evidence Business Wire blocks AI crawlers. Edge protection commonly refuses traffic from data-centre addresses regardless of the user agent presented, and a request from our vantage point is not a request from OpenAI's published ranges. The honest statement is that we could not read the file, so its contents are unknown.
This is worth dwelling on because it is the same error the rest of this subject is full of. A tool reports that something could not be fetched, and the result is reported as a deliberate policy. Absence of evidence arrives looking like evidence.
The same caution applies to our positive findings. A robots.txt records what a site says crawlers may do. It does not record what any crawler did, and a site can permit everything while a crawler ignores it entirely, or permit everything and simply never be visited.
Does being crawlable mean being cited?
No, and this is the gap where most claims about AI visibility are made. Permission to crawl is a precondition for citation, not a cause of it.
Between a permitted page and a cited answer sit several selection steps that none of the parties involved publish: whether the crawler visited, whether the content was retained, whether a retrieval system judged the page relevant to a particular question, and whether the model surfaced it with a link rather than absorbing it silently into an unattributed sentence.
The training case is weaker still. Content used in training does not exist in the resulting model as a retrievable document with a source attached. A model that read a company's announcement during training has no mechanism to cite that announcement later, and asking a model whether it has seen your press release produces a fluent answer with no evidential value whatsoever.
The search case is the more plausible route, and it is the one worth reasoning about. A wire release that ranks for a distinctive query and sits on a domain search crawlers are permitted to read is, at least, eligible to be retrieved and linked. Eligibility is a real thing to have bought. It is simply not the same as an outcome, and nobody in the chain can promise the outcome.
What does an llms.txt tell us?
PR Newswire's robots.txt contains a line most readers would skim past: a sitemap entry pointing to llms.txt, sitting under the comment “# AI Context File”.
We fetched it. The file returns successfully and is written for machine readers rather than people: a short description of the organisation, then labelled links to its chronological news feed, its multimedia gallery, its public company news, and topic feeds for automotive, business and technology, consumer products, health, travel, and policy. A metadata block states the content type as press releases and corporate news and the update frequency as real-time, with multiple updates per minute.
The closing line is the clearest statement of intent: “Full Index: Large-scale crawlers should refer to our Sitemap Index.” This is not tolerance of AI crawlers. It is an invitation, structured to make the archive easy to ingest.
It is worth being precise about what this does and does not show. llms.txt is an emerging convention, not a standard any model is obliged to honour, and publishing one does not oblige any crawler to read it. What it demonstrates is the wire's posture: at least one major distributor treats being read by language models as a service to its customers rather than a leak to be plugged.
What this means for anyone buying distribution
The mechanical answer to the title question is yes, most likely, for the crawling stage. Press releases on four of the five wires we checked sit on paths that OpenAI's crawlers are permitted to read, one wire says so explicitly, and one actively advertises an index built for machine consumption.
The useful answer is that this is the least valuable link in the chain, and it is free. Permission is the default state of the open web; it required no purchase and confers no advantage, because it applies identically to every release on the same wire. Whatever distinguishes a release that gets cited from the thousands alongside it, crawl permission is not it.
The questions that do discriminate are ordinary ones. Does the release contain a specific, checkable claim that someone might search for? Is there a number, a date, a named source? Is it the only place that fact can be found? Retrieval systems select for documents that answer questions, and most press releases are not written to answer anything.
Media Morpheus sells press release distribution. What it buys is publication on a wire whose pages AI crawlers are generally permitted to read — which is real, and is not the same as being read, retained, retrieved or cited. Any seller who tells you a release guarantees presence in AI answers is describing a chain whose later steps nobody publishes and nobody controls.
Methodology and limitations
The robots.txt findings in this report are original measurements taken on 18 September 2026. For each of prnewswire.com, globenewswire.com, businesswire.com, accesswire.com and einpresswire.com we requested /robots.txt over HTTPS and searched the response for the user agents GPTBot, OAI-SearchBot, ChatGPT-User, CCBot and ClaudeBot, then read each file's wildcard rules in full. PR Newswire's llms.txt was retrieved separately and returned HTTP 200. Five limits apply. First, Business Wire's file returned HTTP 403 from its content delivery network to every agent we presented, so its contents are unknown and its row should not be read as an absence of directives. Second, robots.txt reflects a single moment; these files change without notice and the findings should be re-checked rather than assumed. Third, requests were made from one network vantage point, and sites may serve different files to different origins. Fourth, robots.txt records stated permission only: it is evidence about policy, not about crawler behaviour, retention, retrieval or citation. Fifth, the descriptions of OpenAI's three crawlers are taken from OpenAI's published documentation and are not independently verified here.
Conclusion
Asking whether ChatGPT reads press releases turns out to be three questions wearing one coat, and the crawling question — the only one with a checkable answer — resolves in the affirmative almost everywhere we looked. No major wire service names an AI crawler in robots.txt at all; the wildcard rules that govern them by default restrict plumbing rather than editorial content; one wire explicitly allows its news-release path; and PR Newswire publishes a file whose closing line directs large-scale crawlers to its full index. What none of this establishes is the thing actually being sold. Permission to crawl is the web's default setting, identical for every release on the wire and available to anyone who publishes a page. The steps that determine whether a document is retrieved and cited are not published by anyone, and a claim that distribution secures a place in AI answers is a claim about machinery its seller cannot see.
Frequently Asked Questions
Does ChatGPT read press releases?
For the crawling stage, most likely yes. Four of the five major wires we checked permit OpenAI's crawlers by default, because none names them in robots.txt and the wildcard rules do not disallow news-release content.
How many crawlers does OpenAI operate?
Three, controllable separately: GPTBot for training, OAI-SearchBot for search results and citations, and ChatGPT-User for user-initiated fetches, which OpenAI notes robots.txt may not govern.
Do the wire services block AI crawlers?
None of the five we checked named any AI crawler in any directive. GlobeNewswire explicitly allows /news-release/, and PR Newswire publishes an llms.txt inviting large-scale crawlers to its sitemap index.
Does being crawlable mean ChatGPT will cite my release?
No. Crawl permission is a precondition, not a cause. Whether a page is visited, retained, judged relevant and surfaced with a link involves selection steps that none of the parties publish.
If a model was trained on my press release, will it cite it?
There is no mechanism for it to. Training content is not stored as a retrievable document with a source attached, and asking a model whether it has seen your release produces an answer with no evidential value.
What is an llms.txt?
An emerging convention in which a site publishes a machine-readable guide to its content for language models. It is not a standard any crawler is obliged to honour.
Sources
- vendor research Media Morpheus, 18 September 2026. Original retrieval of robots.txt for prnewswire.com, globenewswire.com, businesswire.com, accesswire.com and einpresswire.com, and of prnewswire.com/llms.txt. Method and limitations stated above.
- official OpenAI. Crawler documentation for GPTBot, OAI-SearchBot and ChatGPT-User, including published IP ranges at openai.com/gptbot.json, /searchbot.json and /chatgpt-user.json.
- official Robots Exclusion Protocol, RFC 9309, for the rule that a user agent without a matching group falls through to the wildcard group.