A press release distribution company published a study in July 2026 claiming that 99.3% of press releases get cited by ChatGPT or Claude. Marketing teams have been forwarding that number to justify wire spend ever since. A separate analysis of 4 million real AI citations, run the same year by the digital PR platform BuzzStream, found press releases account for just 0.32% of news citations across ten industries. Both numbers are accurate. They measure different things entirely, and the gap between them explains why a release can get cited almost every time it's asked about and still barely move a brand's AI Share of Voice.
What the 99% figure actually tested
The 99.3% statistic comes from Notified, the company behind GlobeNewswire, working with the AI-visibility platform Profound. Researchers tracked around 8,000 GlobeNewswire releases distributed over one week in May 2026, then asked ChatGPT and Claude questions specifically about the news in each release. Under that narrow test, 99.3% of releases turned up as a cited source at least once, an average of 8 hours after distribution.
That's a legitimate thing to measure, and it says something useful about whether AI engines can find a properly distributed release at all. It's also the easiest number in the industry to misread as citation share, and Notified sells an AI Press Release Optimizer the same report was partly built to promote. The study did surface one genuinely actionable detail: well-structured releases earned 2.6 times more citations than poorly structured ones. What it doesn't say is what portion of a real AI answer, the kind a prospect reading about a purchase decision actually sees, ends up built from press releases rather than anything else.
What citation share looks like once the lens widens
Citation share is the percentage of all AI citations a source type earns across realistic prompts, not prompts about itself. BuzzStream's Citation Labs team analyzed roughly 4 million citations pulled from 3,600 prompts across ChatGPT, Google AI Overviews, Google AI Mode, and Gemini during the week of January 27, 2026, spanning ten industries from finance to travel. Press releases accounted for 0.32% of news citations in that dataset. Newswire content specifically, the raw distributed copy rather than anything a brand controls afterward, made up 0.21% of the full analysis.
The two studies aren't contradicting each other. They're answering different questions. Notified asked whether an AI engine can find a release if you ask about it directly. BuzzStream asked what share of everything an AI engine actually cites, across normal buyer research, comes from press releases. Almost no real customer prompt says "summarize this specific press release," which is why the second question matters more to a brand's AI Share of Voice. It echoes a finding this blog covered before: 85% of AI citations don't come from a brand's own website, whether the source in question is a press release, a product page, or anything else a company publishes about itself.
Syndication splits what little share press releases get
A single press release rarely appears just once online. Wire services syndicate it across dozens of partner sites, financial portals, and aggregators like Yahoo Finance and MSN within hours, and AI engines that deduplicate near-identical pages tend to credit only one version of the text.
This blog has described the same mechanic under a different name: content cannibalization, where a domain publishing several near-duplicate pages on one topic gets an AI engine to cite just one of them, and it's rarely the version the brand would have picked. Syndicated releases run into a sharper version of that problem, because the duplicate copies sit on someone else's domain entirely, competing with the brand's own version for whatever citation share exists in the first place.
BuzzStream's data backs that up in a useful way. Press material published directly on a brand's own domain, rather than distributed through a wire, accounted for 3% of citations overall and 18% of citations on ChatGPT specifically, several times the rate of wire-syndicated copies. An AI engine appears to treat a brand's own announcement, hosted at a URL the brand controls, as a stronger source than the same paragraph reposted across a syndication network.
Earned media still dominates AI citations by a wide margin
Coverage a brand didn't write and didn't pay for remains the largest single source AI engines cite. Muck Rack's Generative Pulse team tracked more than 25 million links cited by ChatGPT, Claude, and Gemini across three editions of its ongoing citation report between July 2025 and May 2026. Earned media held between 82% and 89% of all citations throughout that period, landing at 84% in the May 2026 edition. Journalism alone accounted for 27%, roughly 80 times the citation share BuzzStream measured for press releases.
Earned media's share of AI citations held in an 82 to 89 percent range across three consecutive editions of Muck Rack's tracking study, a level of consistency its researchers read as deliberate system behavior rather than a fluke.
That pattern lines up with what independent, third-party validation seems to earn across AI engines generally: something closer to editorial authority weighting than brand-authored copy keeps winning out. It's consistent with a related pattern this blog has covered: why AI engines trust a brand's Wikipedia page more than its own website for basic facts.
What this means for a press release strategy
None of this makes press releases worthless for AI visibility. It narrows their job to something specific: a fast, factual anchor soon after news breaks, not a durable source of ongoing citations.
The 8-hour time-to-first-citation Notified measured is a real advantage while an AI engine waits for slower sources, independent write-ups, analyst commentary, review-site mentions, to catch up and eventually outweigh it. Well-structured releases earning 2.6 times more citations than poorly structured ones is also worth acting on, regardless of the low overall share, because it costs nothing extra to write a release an AI engine can parse cleanly. What a release can't do is substitute for coverage a brand doesn't control, and it can't fix inaccurate or outdated facts an AI engine already has on file, the kind of problem that raises real liability questions when AI engines get brand facts wrong.
Where to look if press releases aren't moving the number
If a brand's AI Share of Voice isn't shifting despite regular press releases, the gap usually sits elsewhere: in what independent sites say about the brand, whether its own facts are current, and whether near-duplicate content is splitting citations that should consolidate onto one page.
This blog's guide to tracking brand mentions across ChatGPT, Gemini, and Perplexity walks through which source types are worth monitoring beyond a brand's own newsroom.
A separate analysis of 126 million prompts found the same pattern holds industry-wide: brands get mentioned by AI engines far more often than they get cited as an actual source, and press releases, easy to distribute and quick to get picked up, sit much closer to the mention side of that gap than the citation side.
Frequently Asked Questions
Do press releases actually get cited by AI chatbots?
Yes, but rarely, and almost never as a meaningful share of citations. A Notified/Profound study found 99.3% of GlobeNewswire releases were cited by ChatGPT or Claude when tested against prompts about that specific release. Broader usage data tells a different story: BuzzStream's analysis of 4 million citations found press releases accounted for just 0.32% of news citations across ten industries.
What's the difference between a press release being "cited" and having "citation share"?
Being cited means an AI answer references the release at least once, usually when a prompt asks directly about that announcement. Citation share measures what portion of all citations across broad, realistic prompts come from press releases. A release can be cited 99% of the time it's specifically asked about while still accounting for under 1% of total citations, because most real prompts never ask about it directly.
Do syndicated press releases perform differently than releases on a brand's own site?
Yes. BuzzStream's data found owned newsroom content, press material published directly on a brand's own domain rather than distributed through a wire service, made up 3% of citations overall and 18% on ChatGPT specifically, far outperforming syndicated wire copies. AI engines appear to treat a brand's own announcement as a stronger source than the same text reposted across dozens of syndication partners.
What share of AI citations comes from earned media instead?
Earned media, coverage a brand didn't write or pay for, accounted for 84% of citations across 25 million links tracked by Muck Rack from July 2025 through May 2026, holding in an 82-89% range across three editions of the study. Journalism alone made up 27% of citations, roughly 80 times the share attributed to traditional press releases.
Should brands stop issuing press releases for AI visibility?
No, but they shouldn't expect a release alone to move AI Share of Voice. Notified's own data shows well-structured releases earned 2.6 times more citations than poorly structured ones, and releases still work as a fast, factual anchor AI engines can point to soon after news breaks. The bigger lever is still earning independent coverage.
How fast do AI engines pick up a new press release?
In the Notified/Profound study, GlobeNewswire releases were first cited by ChatGPT or Claude an average of 8 hours after distribution. That speed makes press releases useful for establishing a factual baseline while an AI engine waits for slower, more heavily weighted sources like journalism and third-party review sites to catch up.