Two GEO Experiments Found PR Produced 0.2% of AI Citations. Third-Party Listicles Produced 85.8%.

Zeeshan Yaseen, founder of ZeeKnows and CEO of RankViz, ran two structured generative-engine-optimisation experiments back to back and tracked every result by hand. The headline finding is blunt enough to be uncomfortable for a large section of the emerging GEO services industry.

Across 437 source mentions:

• Third-party listicles: 85.8%
• Self-published listicle: 14.0%
• PR: 0.2%

Press releases produced approximately one mention in five hundred.

What was actually tested

Experiment one tracked 15 commercial-intent keywords across four platforms — ChatGPT, Claude, Gemini and Perplexity — over several months, with queries run manually both with and without a VPN. Citation totals came in at ChatGPT 148, Claude 96, Gemini 87, Perplexity 64. Peak keyword presence hit 37.01% on April 29. The source-type split there was listicles 72.4%, PR 24.1%, with guest posts, the owned site and LinkedIn sharing the remainder.

Experiment two was a 30-day cold start — May 30 to June 28 — for a SaaS link building agency with no measurable AI presence at the outset. Six platforms this time, adding Google AI Mode and Grok, again across 15 commercial-intent keywords. It produced 298 appearances: Gemini 104, Google AI Mode 95, Claude 59, ChatGPT 32, Grok 4, Perplexity 4.

Across both, 775 citation events were logged manually.

The concentration is striking. Three sources — Indie Hackers (146), Bruce Jones SEO (69) and the brand's own listicle (61) — accounted for 78% of all mentions.

Why PR collapsed between the two tests

This is the detail worth dwelling on. PR went from 24.1% of citations in experiment one to 0.2% in experiment two.

The obvious explanation is the difference between the subjects. Experiment one involved an established brand with existing coverage and authority; experiment two was a cold start with none. Press coverage for a brand nobody has heard of appears to do almost nothing for AI citation, while press coverage attached to an already-recognised entity does something.

If that reading is right, PR isn't useless for AI visibility — it's not a starting move. Which is close to the opposite of how it's currently being sold to companies with no AI presence who are told a press release campaign will fix it.

The mechanism makes sense

When someone asks an assistant for "the best X," the model reaches for pages that already are lists of the best X. A third-party listicle is a pre-formatted answer to the exact question, written by someone other than you — it carries structure and independence at once.

Your own page claiming you're the best is the same claim without the independence. A press release is the same claim with even less.

Notably, the brand's own listicle still managed 14.0%, and the two strongest keywords in the data were "Best SaaS Link Building Agency in USA" and the same phrase with "2026" appended, at 18 appearances each. So owned assets aren't worthless for high-value commercial terms — they're outgunned roughly six to one.

What I'd take from this

Audit which listicles rank for your commercial terms, and get on them. That's the actionable version of this finding. Not "create thought leadership" — identify the specific third-party pages that answer your buyers' comparison questions and work out how to be included legitimately.

Stop treating GEO as a content-production problem. The prevailing advice is to publish more, add more schema, restructure your own pages. This data says the lever is other people's pages. That's a PR-and-relationships job wearing an SEO hat, not a publishing calendar.

Don't lead with press releases on a cold start. 0.2% is close enough to zero to plan around.

Watch Gemini and AI Mode if you're starting from nothing. In the cold-start test those two produced 199 of 298 appearances — two thirds. New entities appear to surface there far faster than in ChatGPT, which dominated the established-brand test. That inversion is one of the more useful practical signals in the study.

Read it with the caveats attached

This is one researcher, two brands, 15 keywords each, tracked manually. It is not a large-scale study and shouldn't be cited as one. Manual checking across platforms is also subject to personalisation and volatility in ways that are hard to fully control, which is presumably why the VPN comparison was included.

But the method is described, the numbers are specific, the two experiments used different niches and reached a consistent conclusion — and it's testable. Anyone can run the same check on their own commercial terms this afternoon.

That's a considerably better standard of evidence than most of what currently circulates as GEO advice, which is why it's worth taking seriously while holding the sample size in mind.

Sources

• Search Engine Land: Two GEO experiments challenge conventional AI visibility advice — Zeeshan Yaseen's original research, published there