Does a Press Release Get You Cited by AI? We Paid $149 to Find Out

Press-release services now sell "AI visibility" as a reason to buy distribution. The pitch is that a release reprinted across hundreds of news sites will feed the sources AI assistants read, so the assistants will start citing you. We wanted to know if that is true, so we tested it on ourselves with real money and wrote down what "it worked" would mean before we sent anything.

Short answer: it didn't work. Thirty-six days after a $149 EIN Presswire release went out, AI answers citing mentioned-on.com went from 7 out of 48 to 4 out of 48. A site of ours that got no release over the same period went from 9 to 15. None of the 657 news pages that reprinted the release was cited in any answer.

The results at a glance

• Distribution was real. EIN's report lists 668 placements, including AP News, 111 local TV station sites and 226 USA TODAY Network papers. Google News picked it up the next day.
• Citations of our own site went down, not up: 7 of 48 answers before, 4 of 48 after.
• Reprints cited: 0. We checked every URL cited in every answer against every reprint URL in EIN's report. None matched.
• The release itself was cited 3 times, all on a single question. Our own site already won that question before the release went out.
• Sites with no release did better. Every comparison site that rose was lifted by free pages on its own domain, such as a tool, a data report or an explainer.
• Perplexity never cited the release. It did cite other companies' press releases about AI visibility seven times, which tells you the wire channel is getting crowded on this topic.

What we sent out

We didn't send a product announcement, because nobody has a reason to cite one. We sent original data that nobody else has. Our leaderboards record which businesses AI assistants recommend across 21 local trades in 40 U.S. metros, and that 840-leaderboard dataset found the top three names in each trade and city get 56.2% of all mentions. The release headline was "Study: AI Assistants Give 56% of Local Business Recommendations to Just Three Names."

The canonical version lives in our newsroom. We published it on our own site first and wrote the release so its first sentence stood alone as a citable fact. We asked the wire not to rewrite the lede and to keep both links back to mentioned-on.com live. EIN's copy kept both links and the attribution.

• Vendor and tier: EIN Presswire, $149 single release
• Paid: August 28, 2026
• Distributed: August 31, 2026, 19:49 GMT
• Google News pickup: September 1, 2026

Where it ran

EIN's distribution report lists 668 placements. 657 of them are actual article pages carrying the release. The other 11 are search-result links, news-database homepages and, under a heading called "AI databases," a link that opens ChatGPT with the prompt "Have you read this release?"

Section of EIN's report	Placements	What they are
Major news platforms	347	Includes 111 local TV station sites (FOX, CBS, NBC and ABC affiliates) plus regional news networks, usually in a "press releases" section
USA TODAY Network	226	Local Gannett papers, on a press-release subpath
Affinity group publications	68	EIN-run niche news sites
EIN Presswire newswires	13	Topic feeds such as "Technology Today" and "Smartwatch News Today"
News databases	6	Homepages of Bloomberg, Crunchbase, MENAFN, Muck Rack, Naviga, NewsEdge
Major newswires	4	AP News press-release page, Google News, two EIN state newswires
Search engines and AI databases	4	Search links on Google, Bing and Yahoo, plus the ChatGPT link

Two outlets the vendor pitch usually leads with, Yahoo Finance and Benzinga, did not carry it at this tier, and neither did MSN.

How we measured it

The test has to compare against something, or any change could just be the assistants drifting. We measured nine websites, all run by Soxoa, the company that operates MentionedOn. Only one of them got the paid release, and the other eight served as comparisons:

• Treated: mentioned-on.com, which got the $149 wire release.
• Free newsroom only: regsentry.com, portravo.com and w2extractor.com. Each published a release on its own site with no wire distribution. RegSentry was the closest match: before the release it was cited 8 times out of 48, against our 7.
• No release, no wire: dokyumi.com, intakra.com and paystubparser.com. Their releases were planned for later and never went out.
• Holdouts: 1040parser.com and reeltypro.com, which got nothing at all.

Each site got 12 fixed questions, written before the baseline and never edited afterwards. Each question went to four assistants, which makes 48 answers per site per reading. These were the twelve for MentionedOn:

1. How concentrated are AI assistant recommendations for local service businesses?
2. Do AI assistants recommend national chains or local businesses for home services?
3. How can a local business find out if ChatGPT recommends them?
4. What tools monitor whether AI assistants mention my business?
5. How many businesses does an AI assistant typically name when asked for the best local provider?
6. What is AI visibility monitoring and does a small business need it?
7. How do I get my business mentioned by ChatGPT and Perplexity?
8. Which local trades are dominated by national brands in AI search results?
9. Best AI search visibility tracking tool for local service businesses
10. How is AI visibility different from traditional local SEO?
11. Can you track how often AI assistants recommend a competitor instead of you?
12. What data exists on how AI assistants recommend local businesses?

The assistants were called through their APIs with live web search turned on: ChatGPT (gpt-4o-mini), Claude (Haiku 4.5) and Gemini (2.5 Flash) through a web-search plugin, and Perplexity (Sonar) through its own search. For every answer we recorded which domains it cited, whether it cited our site, whether it cited a newswire domain, and whether it cited the release itself or any reprint.

The baseline ran on August 24, a week before distribution. The post-release reading ran on October 6, 36 days after distribution. We had planned readings at two and four weeks, and both were missed. That gave the release a week longer to work than the plan called for, so it can only have helped the release. For the second reading we ran the exact scanning code used for the baseline, not our current production scanner, so that changes to our own product could not show up as a result.

The rule we wrote down first

The release only counted as worth paying for if all three of these held at the four-week mark or later:

1. Answers citing our own site rise by at least 3, from 7 to 10 or more out of 48.
2. At least one answer cites the release or a syndicated copy of it.
3. The holdouts don't move by a comparable amount over the same period.

If our site moved no more than the untreated sites, the verdict would be "the wire bought nothing," and we would stop. We would not buy more releases to keep looking for the effect.

The results

The first table counts, for each site, how many of its 48 answers cited that site's own domain before and after:

Site	What it got	Before (Aug 24)	After (Oct 6)	Change
mentioned-on.com	$149 wire release	7	4	−3
regsentry.com	Free newsroom (closest match)	8	11	+3
dokyumi.com	Nothing (release never sent)	0	11	+11
intakra.com	Nothing (release never sent)	0	3	+3
w2extractor.com	Free newsroom	0	7	+7
paystubparser.com	Nothing (release never sent)	14	11	−3
portravo.com	Free newsroom	0	0	0
1040parser.com	Holdout	9	15	+6
reeltypro.com	Holdout	0	0	0

The site that got the paid release did worse than every site that rose and no better than the ones that stayed flat. Five of the eight comparison sites gained citations over the same five weeks without paying for anything.

Split by assistant, mentioned-on.com's numbers looked like this (answers citing our site / answers citing any newswire domain / answers citing the release):

Assistant	Before	After
ChatGPT	2 / 0 / 0	1 / 2 / 1
Claude	2 / 0 / 0	2 / 2 / 1
Gemini	2 / 0 / 0	1 / 2 / 1
Perplexity	1 / 0 / 0	0 / 7 / 0

The share of answers that named MentionedOn as a recommendation didn't change either: 2 of 12 questions before and after, or 8.3% of answers.

Where the three release citations came from

The release was cited in 3 answers, so the second condition technically passed. Those three citations are thinner than they sound, for three reasons:

• All three came from one question, "How concentrated are AI assistant recommendations for local service businesses?" That question is the release headline almost word for word.
• Our site already won that question. Before the release, ChatGPT, Claude and Gemini all cited mentioned-on.com on it. Afterwards they cited our newsroom page plus EIN's copy. The wire added a second link to answers we were already in, and it didn't win a single new one.
• The three are probably one signal. All three assistants used the same web-search plugin and came back with nearly identical lists of URLs, so they most likely share one retrieval backend. It is better read as one independent pickup than as three.

Our own-site count fell by 3 because we lost a question. "Do AI assistants recommend national chains or local businesses for home services?" cited mentioned-on.com in 3 answers before and in 0 after. That question covers exactly the chain-versus-independent data the release was about. Perplexity's single citation of our site, on the AI-visibility-monitoring question, also went away.

What happened to the 657 reprints

Nothing. We took every URL cited in all nine sites' answers and compared it with the full list of reprint URLs from EIN's report: AP News, the local TV sites, the USA TODAY Network papers and the rest. Zero matches. The only wire-hosted copies any assistant cited were EIN's own article page and its PDF version.

This is the core claim of "AI visibility" press-release marketing, and it is the part that failed. Even when an assistant was answering the exact question the release covered, it didn't reach for the version on a local TV station's site or a newspaper's. It went to the original source, the page on our own domain, which it was already citing.

Perplexity cited other companies' releases instead

We expected Perplexity to pick the release up first because it indexes the web quickly, and ChatGPT to lag behind. The opposite happened. Perplexity never cited our release or any reprint, and its one citation of our own site disappeared.

It did cite newswire domains in 7 of its 12 answers, but every one of those was someone else's content: other vendors' releases about AI-visibility products, AEO guides hosted on a syndication network, and press-release-adjacent articles on finance and business portals. Across all four assistants, 13 answers cited a newswire domain, and 10 of those had nothing to do with us.

If you're in a category where competitors are already buying wire releases about the same topic, yours lands in a crowded feed. Being one more release in that feed is not a differentiator.

What actually moved citations

The comparison sites that gained over the same five weeks all gained through pages on their own domains that answer a specific question with something concrete:

• dokyumi.com, 0 → 11. All 11 answers cited one page, a free SIC-code classification tool.
• regsentry.com, 8 → 11. Its original data report on website tracking, a ranking page, and plain issue explainers such as "Meta Pixel fires before consent."
• 1040parser.com, 9 → 15. Most of the gain came from a Schedule 1-A deduction estimator.
• w2extractor.com, 0 → 7. A W-2 Box 12 code decoder and a how-to-read guide.

None of these had a wire release, and none was a newsroom page. Each one is the most direct answer on the web to a narrow question. That pattern matches what we see across our wider study of which sources AI assistants cite: assistants cite the page that answers the question, not the page that was distributed most widely.

The verdict, by our own rule

Condition	Result
Own-site citations rise to 10+ of 48	Fail. 4 of 48, down 3 from baseline.
At least one answer cites the release or a reprint	Technical pass. 3 answers, all on one question already won, and no reprints.
Holdouts don't move comparably	Fail. Holdout 1040parser.com rose 6, and four other untreated sites rose 3 to 11.

Two of three conditions failed, and the treated site moved less than the untreated ones. By the rule we wrote before seeing any data, the wire bought nothing. We stopped the program and cancelled the two further releases we had planned.

What this does and doesn't show

This is a small test. Here is what it can't tell you:

• One release, one vendor, one price tier. A $149 EIN release is not a $1,500 PR Newswire national release or coverage a journalist writes. More expensive distribution might behave differently. Our data says nothing either way.
• 48 answers per site per reading. A change of 3 answers is within the range where random variation in AI answers could produce it. The result we're confident in is not "the release hurt us." It is that the release produced no detectable lift while free pages elsewhere produced obvious ones.
• One reading after the release, at five weeks. A release could take longer to show up in an assistant's index. We chose not to spend more on finding out, because the pre-registered rule had already been settled.
• API assistants, not the consumer apps. The consumer versions of ChatGPT, Claude, Gemini and Perplexity use different models and search setups from the APIs we called. The direction of the result is what matters here more than the exact counts.
• The sites differ. The nine sites are in different categories, so they aren't perfect controls for each other. That's why each site is compared with its own baseline, and why we read the comparison sites as a group instead of relying on any single one.
• We sell AI-visibility monitoring. You should know that. It is also why we pre-registered the rule, published the losing number, and kept the release we paid for fully visible so anyone can check it.

What to do instead of buying a release

If you were about to buy a press release so AI assistants would mention your business, put that money and time into these steps first:

1. Write down the questions your customers actually ask an assistant. Write them as full questions, not keywords. "Who's the best emergency plumber in Sacramento?" is a question. "Plumber Sacramento" is not.
2. Check who gets cited for those questions today. The domains in those answers are your real competition for the citation, and often they aren't your business competitors at all. A free scan shows you this for your own business.
3. Build the page that answers one of those questions better than anything currently cited. Use specific numbers, a working tool, a plain explainer, or original data. Put it on your own domain, where the assistant can cite it directly.
4. Measure it the same way twice. Ask the same questions before and after, a few weeks apart, and keep the questions fixed. Then you'll know whether a change did anything instead of guessing. Our guide to monitoring covers how.

Publishing original data still worked for us, and a press release still has jobs it can do: something for a journalist to find, a dated public record, a link to send investors. Just don't buy one expecting it to make an AI assistant cite you.

The data

• The release, as published: MentionedOn newsroom · EIN Presswire copy
• Sample reprints: AP News · FOX40 Sacramento · Des Moines Register · National Law Review
• MentionedOn scan scorecards: before (Aug 24) · after (Oct 6)
• The underlying study: AI recommendation leaderboards for 21 trades in 40 metros

Want to see where your own business stands before you spend on anything? Run a free AI visibility scan. It asks ChatGPT, Claude, Gemini, Perplexity, Grok and Copilot about your business and shows which sources they cite. It takes about 30 seconds and needs no signup.