Do press releases show up in AI answers?

Someone asks an AI assistant what your company does. Does the press release you paid for turn up in the answer?

Nobody can tell you. Including us. There is no public citation data, the same question asked twice returns different sources, and every vendor selling "AI Search Visibility" as a line item is quoting you a number they cannot check either.

What can be checked is the layer underneath — and it is the more useful question anyway. Is an assistant even permitted to fetch the page your release lands on? Every publisher answers that in public, for free, in a file at the root of its domain. We read all 21 of ours on 9 September 2026 and ran each one through the matching rules the standard specifies.

The answers are unflattering to this industry, ours included.

The short version

• Ten of our 21 outlets block at least one major AI crawler from the path a release lands on. They are mostly the famous ones.
• Four are closed to every AI crawler we tested: AP News on its press-release path specifically, plus USA Today, MSN and The Globe and Mail across ordinary article paths.
• Two will not serve us their robots.txt at all — the file itself returns 403 behind a bot challenge — so their policy is unknowable from outside.
• Nine carry no directive that touches assistant retrieval.
• Permission is per-crawler, not per-outlet. MarketWatch allowlists OpenAI's three crawlers by name while leaving Anthropic's under a whole-domain block. "Can an assistant read this page" has a different answer depending on which assistant is asking.
• None of that is a citation. Being allowed to read a page is a precondition for being cited, never a cause of it. We sell the first thing. We will not put a price on the second.

If someone quotes you a figure for AI visibility, ask them one question: what would have to be true for that number to be wrong, and how did you check? We have never heard an answer.

The evidence, outlet by outlet

Every publisher states, in a file at the root of its domain, which automated agents may fetch which paths. The format is a published standard — RFC 9309 — the rules are matched per named agent, and the longest matching rule wins. We evaluated every file against nine crawlers: OpenAI's GPTBot, OAI-SearchBot and ChatGPT-User; Anthropic's ClaudeBot, Claude-User, Claude-SearchBot and anthropic-ai; Google-Extended; and PerplexityBot.

What we counted, so you can check us. A restriction counts above only if it blocks one of those nine crawlers from the path a release lands on. Blocks aimed at SEO tools, ad-verification bots, scrapers or training-corpus crawlers are not counted, and several outlets have them — TechBullion blocks Common Crawl, Amazon and ByteDance while explicitly permitting Google-Extended and PerplexityBot. Every file above is public. Read them yourself; it takes about a minute each.

One thing that surprised us while doing it: whether you can read the policy at all depends on how you introduce yourself. Fortune and USA Today both served us their file under one user agent and returned 403 under another. We report what the files say to a reader who can get them. We cannot promise every agent sees the same bytes.

And a correction, since this article is where we make the case for checking things. An earlier version of this table said MarketWatch was off-limits to everything. That was wrong: the wildcard block is real, but OpenAI's crawlers are allowlisted underneath it by name. That version also placed TechCrunch, MSN and The Globe and Mail among the unrestricted outlets, and all three carry whole-domain blocks today. We cannot tell you whether those three changed in the eleven days between our sweeps or whether we read them wrong the first time — which is exactly why the table above carries the date we read it, and why you should re-check it rather than trust it.

The row that matters commercially

Read the AP News row again. AP is the outlet this industry sells hardest, and AP fences its press-release section off from automated retrieval by name, for every agent, in the group that applies to everybody.

We saw it enforced in production: a search query scoped to apnews.com through one major assistant's search API came back as an error saying the domain is not accessible to its user agent. Not "no results". Refused.

It goes further than the assistants. When we searched for the full text of one specific AP-syndicated release, the text turned up on Morningstar, Yahoo Finance, MarketScreener, 3BL Media, CSRWire and the issuer's own newsroom — and not on apnews.com at all. On this evidence the AP copy of a release is the copy least likely to be found by anything.

Three publishers go further still, and keep paid content out of their own index. USA Today and Toronto Sun each exclude their sponsored-story paths from Google News specifically, aimed at Googlebot-News, the separate token Google documents for its news crawler. Fortune goes furthest: Disallow: /sponsored/ sits in the group that applies to every agent. That is a publisher telling the whole web not to treat the thing you bought as news.

The one number anybody has

Meltwater measured press-release citation share in AI search falling from 0.4% to 0.2% between April and May 2026 — a halving in a single month.

We did not run that study and cannot reproduce it; it is one vendor's measurement over one month, and you should weigh it accordingly. It is also the only published measurement of this that exists, and it points down. Anyone selling you AI visibility as a growth channel is selling against the only number on the table.

The limits of our evidence, stated plainly

A robots.txt file is an instruction that well-behaved crawlers follow. It is strong evidence about permission, not proof about every system in existence:

1. It governs live retrieval only. It says nothing about whether a model absorbed a page during training, which is unobservable from outside.
2. Compliance varies, and we cannot audit it. Google says as much about its own format: the instructions in a robots.txt file cannot enforce crawler behaviour — it is up to the crawler to obey them. We can only show that at least one production system enforced it, which we did.
3. It is per-agent, and vendors run several. Blocking GPTBot is not blocking OAI-SearchBot; blocking ClaudeBot is not blocking Claude-User. Which crawler an assistant reaches for is the assistant's business, not ours.
4. It changes, and so do we. Four outlets this article once listed as unrestricted are not, eleven days later. We cannot separate "the publisher changed it" from "we read it wrong", and neither can you from outside. A domain that is open today can close tomorrow, without notice, and this article will be out of date again.

That is still a far better basis for a purchasing decision than a promise about AI citations, for one reason: it is checkable by you, today, in a browser, for free.

A page that stops existing cannot be cited by anything

Retrievability is worthless if the URL dies, and press pages die more often than the sales material suggests.

Digital Journal has deleted an entire tier of its press archive. We checked its section indexes wire by wire: the paths for ACCESS Newswire, GlobeNewswire and 24-7 Press Release returned 200, while five reseller-wire sections returned HTTP 410 Gone. A 404 means not found; 410 is the status a publisher returns to tell search engines a page is permanently removed and should be de-indexed. Individual articles behave the same way — reseller-wire URLs still sitting in search results return 410 when opened.

We also found one MSN placement still findable in search that renders "This story is unavailable", and one MarketWatch URL still cited in a competitor's live case study that returns "Page Not Found". A third outlet changed the published slug on a delivered release, so the URL in that customer's report now redirects rather than resolving.

Against that, plenty of pages last. A release published on Street Insider on 26 February 2026 was still live when we opened it six months later. CoinMarketCap articles from June and July were still live in August. Persistence is real. It is simply not universal, and nobody can promise it — we do not sell permanent links and we will not, because we have watched links stop being permanent.

Archiving does not rescue this either. We asked the Wayback Machine's availability API for the fifteen specific real placements in our verification study. Two of the fifteen had any snapshot at all.

What we cannot prove, and will not imply

Everything above is about permission. None of it is about citation. Four reasons we cannot bridge that gap, none of which we can engineer around:

1. We have no citation data. We do not see assistants' query logs or retrieval traces. Nobody outside those companies does. Any "we got you cited" claim in this market is either an anecdote or a screenshot.
2. Retrievability is a precondition, not a cause. Everything measured above establishes that a page can be read. Whether a system chooses to surface it — and to attribute it — depends on ranking behaviour none of these companies documents, and which they change without notice.
3. The answers are not reproducible. Ask the same question twice and you can get two different sets of sources. A method that gives different results on Tuesday is not a method anyone can sell a monthly report against.
4. Your copies compete with each other. A release is published verbatim at many URLs on the same day; we watched one submission appear on two major finance sites within hours. A retrieval system that surfaces one of them has no obligation to surface yours, or to name your company rather than the wire.

There is a fifth reason, and it is about us. A claim we cannot verify after the fact is a claim we cannot stand behind when a customer asks. Anyone can re-open a URL in twelve months and see whether a release is still on it — a sample report shows the state of every row on the day it was written. We cannot re-run an AI answer from March and show you what it said.

Why the screenshots are not evidence

You will be shown screenshots of an assistant naming a client. Treat them the way you would treat a screenshot of a stock chart.

There is no timestamp you can trust, no way to know how many attempts produced it, no way to know whether the prompt named the company outright, and no way to reproduce it. It is not falsifiable, which means it is not proof — of anything, in either direction. We have not run that test and we are not going to publish one, because a test we ran on ourselves is not evidence either.

So what should you actually do

1. Buy the pages, not the citation. A release buys published pages at real URLs, carrying your announcement in your words, on domains people and machines can find. That is worth money on its own — it is what an investor finds, what a partner's compliance team files, and what an assistant reads if it goes looking. It is not a ranking and it is not a promise of citation.
2. Weight the open outlets accordingly. Where a publisher permits automated retrieval — nine of our 21 outright, and several more for some crawlers but not others — those pages are readable by an assistant that goes looking. AP News, USA Today, MSN and The Globe and Mail are not, and you should not pay a premium for them on an AI-visibility argument.
3. Ask any vendor quoting an AI number the falsifiability question. What would have to be true for that number to be wrong, and how did you check? The answer tells you what you are buying from them.
4. Judge us on something you can re-open. Every link we report is one you can click, today and in a year. That is the whole promise, and it is the only kind we can keep.

If that is the deal you want, here are the packages and what each one buys. If it is not, you now have the file paths to check anyone else's claims — which is a fair outcome for an article too.