AI Engines Don’t Read Your Press Release. They Read the Reporter Who Did.

In the first quarter of 2026, 250 press releases went out over PR Newswire and Business Wire. Over the next 90 days, 203 of them were never retrieved by ChatGPT, Claude, Gemini, Perplexity, or Google AI Overviews. Not at 30 days, not at 60, not at 90.

The 50 partnership and award announcements in that sample scored zero. Fifty releases, five engines, three checkpoints, no citations.

That’s the headline from the Press Release Retrieval Study 2026, published August 12 by Everything-PR’s editorial team. It’s a PR trade outlet, not a university lab, and 250 releases is a modest sample. I’d still take it seriously. It’s the most granular look anyone has published at what happens to a release after it leaves the wire, and three other datasets from this year point the same direction.

The wire is a delivery truck, not a source

The study pulled a stratified sample: 50 releases each across M&A, regulatory filings, product launches, executive appointments, and a partnerships-and-awards bucket, spread evenly over five industries. Each one was checked against the five engines at 30, 60, and 90 days, for 15,000 observations scored by two raters who agreed 94% of the time.

Of the 47 releases that were cited at all, Perplexity cited 40. ChatGPT cited 28, Gemini 22, Claude 17, Google AI Overviews 14.

But look at what got cited. Perplexity linked directly to a wire URL seven times. Every other engine cited journalist coverage of the announcement rather than the announcement itself. The authors put it bluntly: “The engines are not reading the wire — they’re reading the journalist who read the wire.”

So the wire did its job. It delivered the release to newsrooms. What it didn’t do is put the release in front of the model. A reporter did that, and only when a reporter chose to.

This matches what Muck Rack found in its May 2026 “What Is AI Reading?” report, which analyzed more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries. Earned media accounted for 84% of citations and journalism alone for 27%. Paid and advertorial content came in at 0.3%. Meltwater’s June analysis of more than 8 million citations across eight models had press releases falling from 0.4% of citations in April to 0.2% in May. And Loganix ran 100 category queries across ten verticals in March and found press release domains in exactly none of them.

Four datasets from four different shops, one answer.

Which announcements actually get retrieved

The category breakdown is the useful part, because it tells you what to stop wiring.

Release type	Sample	Cited by any engine within 90 days	Rate
M&A / acquisition	50	20	40%
Regulatory / SEC filing	50	15	30%
Product launch	50	7	14%
Executive appointment	50	5	10%
Partnership / award / other	50	0	0%

Writing quality doesn’t explain this pattern. What explains it is whether the release contains a fact of record, something a person will later ask an AI about. “Who bought Acme?” and “What did Acme disclose in its 10-K?” are questions with one verifiable answer, and the release is the primary record of that answer. “Which vendors won a Gold Stevie this year?” is a question nobody types, and no reporter covers it, so nothing ever enters the retrieval path.

Product launches at 14% should worry anyone running a launch calendar through the wire. The study doesn’t say which seven made it, but with editorial pickup multiplying retrieval odds by six, the safe bet is they were the launches someone wrote about. A launch nobody covered is a launch the models never saw.

Press releases shape one kind of query, and it’s not the one you’re paying for

Loganix’s zero deserves a closer look, because the same study found something different when it switched query types.

For category queries like “best CRM for small teams,” the engines pulled from G2, NerdWallet, Clutch, Wirecutter, and vertical comparison publishers. Press wires: nothing. But for brand queries like “What is [Brand]?”, coverage on Yahoo Finance’s news path earned confirmed citations, while the same content on a /press-releases/ path at Fortune or Reuters rated “Rarely,” and prnewswire.com and globenewswire.com rated “No” or “Rarely” on ChatGPT and Gemini. Loganix’s read is that press content “serves the training data layer,” the mechanism models use to build base knowledge about a brand, rather than the retrieval layer that answers recommendation questions.

We split monitoring queries into three types at RivalHound: Discovery (“best X for Y”), Direct (“what do you know about Acme”), and Comparison (“Acme vs. Rival”). Map the evidence onto that split and the picture gets sharp. A press release, when it works at all, moves the Direct answer. It teaches the model what your company is, what it just acquired, and who runs it. It does almost nothing for Discovery, which is where buyers who don’t know you yet decide whether to consider you.

Most PR budgets are justified on Discovery outcomes and spent on Direct mechanics. That’s the mismatch.

Muck Rack’s data adds a wrinkle: press releases show up in industry-trend responses at 3.5 times the rate they appear in best-of queries. So there’s a narrow lane, trend and news context, where wire content still surfaces. It’s not the lane where “which one should I buy” gets answered. The sites that own that lane are the same 15 domains that control most AI citations, and none of them are wire services.

Five things that predicted retrieval

The study checked which features separated the 47 cited releases from the 203 that vanished.

Predictor	Retrieval rate when present	When absent	Multiplier
Editorial pickup by a top-50 publication	38%	6%	~6x
Citable number in the first 100 words	28%	9%	~3x
Five or more named entities in the body	27%	11%	~2.5x
Regulatory or SEC-required	30%	15%	~2x
Schema-marked boilerplate	24%	13%	~2x

Editorial pickup at 38% versus 6% is the whole game, and the study says as much. The other four are worth doing, but read them for what they are: they make the release easier to compress into a sentence a reporter will keep and a model will extract. A number in the first 100 words survives the rewrite. Five named entities give the model something to anchor on. Schema on the boilerplate, with company name, founding year, HQ, sector, and CEO, helps the engine resolve which entity the story is about. We’ve written before about how attribute-rich schema beats generic markup, and the boilerplate is the one block of a release where doing it properly is cheap.

None of that substitutes for a journalist. It just improves the odds that a journalist produces something the engines can use.

The 60-day clock

Retrieval doesn’t accumulate forever. At 30 days, 20 releases had been cited (8%). At 60 days, 35 (14%). At 90 days, 47 (19%). Almost everything that was ever going to be picked up was picked up by day 60, and the study concludes that a release uncited at 60 days is unlikely to be cited at all.

That kills one of the wire’s oldest sales pitches, the permanent public record. The record persists on the wire’s servers. It doesn’t persist in the retrieval pool, where ChatGPT replaces about three-quarters of its cited sources every week. If your PR reports still count “pickups” at 7 days and “impressions” at 30, you’re reading the wrong clock. Check at 30, check again at 60, across engines, then close the book.

What to do with the budget

The study’s own math: at $3,000 to $5,000 per release and a 19% retrieval rate, you’re paying roughly $16,000 to $26,000 per AI citation. The 50 awards and partnership releases in the sample cost somewhere between $175,000 and $250,000 to distribute and returned nothing at the AI layer.

Some of that money should move. My suggestions, roughly in order of impact:

Wire the categories that retrieve. M&A, regulatory disclosures, and executive changes. Keep launches on the wire only when a journalist has already committed to covering them.
Stop wiring awards and minor partnerships for AI reasons. Post them on your newsroom page, share them on LinkedIn, and save the distribution fee.
Move the saved fee into pitching. The Trade Press AI Index, published in May by 5W and Everything-PR, names the three to five outlets per vertical the engines actually cite: PCMag, TechRadar, and CIO.com in technology, STAT and Endpoints News in healthcare, Skift in travel, Bloomberg in financial services. One placement there is worth more than a stack of wire pickups on sites no model reads.
Rewrite the template. Number in the first sentence. Five named entities before the boilerplate. Organization schema on the boilerplate itself.
Measure at 60 days, across all five engines, and treat Perplexity as your early warning. It cited 85% of the releases that got cited anywhere, and it’s the only engine that consistently reads the wire URL directly.

The competitive read

There’s a second use for all this, and I’d argue it matters more.

Your competitors’ releases are a leading indicator of what AI will say about them. A rival’s M&A announcement has roughly a 40% chance of appearing in AI answers within 90 days. Their new-CEO release, 10%. Their “Recognized as a Leader” award release, none. If you’re watching what competitors publish, and you should be, you can now weight each item by how likely it is to reach the model at all.

Then check. Run the Direct queries about the competitor at 30 and 60 days and see whether the announcement shows up. If their launch is being cited and yours isn’t, the difference is almost never the release copy. It’s that someone at TechRadar wrote 400 words about theirs. That’s a media relations gap, and it’s fixable, but only if you can see it. Most teams can’t, because they track their own mentions and nobody else’s. We covered the mechanics in our piece on earned media as the main driver of AI citations. The press release study is what happens when you skip the earned part.

The wire still moves announcements into newsrooms. It just stopped being a shortcut to the answer.

RivalHound tracks your brand’s visibility across ChatGPT, Google AI, Perplexity, and more. Start monitoring to see where you stand.