Two AI engines, two different security vendors, 72.7% of the time.

The measurement

The three agreements are the tell. Identity and access management, container security and threat intelligence — the three categories with one obvious canonical answer. Everywhere a category had genuine competition, the two platforms picked different winners.

All twenty questions, including the nine that did not count

Nine questions are excluded from the 72.7% because one or both platforms contradicted itself across three identical calls. That guard was set before retrieval, and instability is never reported as disagreement. They are published here anyway, because a reader cannot check a rate whose exclusions are invisible.

Penetration testing is the honest row. Perplexity declined to name a leader both times, with the same reasoning each time — there is no single best provider. ChatGPT named Rapid7, then nobody, then CrowdStrike. The platform that refused to answer gave the better answer.

What was registered, and what it cost me

Four predictions with bands, written before any retrieval. Two came out wrong and one of them fired a falsifier I had written specifically to kill my own hypothesis.

The P2 falsifier read: P2 dies if citation overlap is at or below 16.7%. That result would mean the corpus split is a property of how these platforms retrieve, not of thin coverage in entertainment — a stronger and more general finding than the one I am predicting.

It came in at 3.8%. I expected the dense commercial corpus of a B2B category to pull the platforms together. It pushed them three to four times further apart than entertainment did. The prediction was wrong in the direction of a stronger result, which is the only kind of wrong worth having.

What the engines were actually reading

On best EDR platform for a mid-size company, ChatGPT’s first answer led with Seceon, sourced to Seceon’s own press release on a paid newswire — a company growth announcement, syndicated, surfaced as the top recommendation to a buyer. These were the nine domains behind it:

A paid newswire, a vendor’s online store, a Japanese SaaS listing, an Italian software-licensing reseller, two SaaS directories and one security publication. Perplexity’s five, identical on both calls:

A US buyer asking ChatGPT which security product to buy is being answered, in part, from Japanese and Italian software reseller listings. That is also the mechanism behind the 17.2% stability: a long tail of thin directory pages gives it a different draw every time.

Review databases, not trade press

Six source types were registered before retrieval as the ones a model would reach for. Four hit and two did not, and the two that missed are the pure security trade titles.

The registered set appeared on 11 of 20 questions, 26 marker hits in total. Models answering buyer questions reach for review and comparison sites, which are commercially maintained and structured for exactly this query shape, and skip editorial trade press, which is organized around news rather than selection. G2 is not accessible consumer press — it is a buying-decision database, and that is why it is there.

What this supports, and what it does not

Supported: in cybersecurity, on buyer-intent questions, the two platforms usually name different leaders and almost never cite the same sources. A parallel entertainment measurement was the weaker version of this same result, which means it is not an artifact of one category.

Not supported: which platform is right. Nothing here measures accuracy. Perplexity is more stable and better sourced; that is consistency and provenance, not correctness.

What a vendor can act on: the engines read review and comparison databases and do not read the security trade press. If being named matters to your pipeline, that is where the asymmetry is — and it is two separate problems, not one, because the two engines share 3.8% of their sources.

What this study cannot tell you

• Nothing here measures which platform is right. Perplexity is more stable and better sourced. That is consistency and provenance, not correctness, and this study does not establish that either engine names the better product.
• Platform-default budget-tier models — gpt-4.1-mini and sonar. Buyers are usually served the flagship. Nobody in this category is measuring the flagship either, and this study does not close that gap.
• One row of 100 was lost to a 30.1-second timeout on the third container security call. That question is one of the three agreements and rests on two calls rather than three. Had the missing call named something else, the disagreement rate would be 80% rather than 72.7%, so the published figure is the conservative one. It was not re-run, because the stopping rule said one pass.
• Nine of twenty questions are excluded from the headline rate, because one or both platforms contradicted itself across identical calls. Those are reported as instability and never as disagreement. That guard was set before retrieval, and all nine are published in the table above rather than dropped.
• One pass, one day, one location. October 19, 2026. A platform that retrieves differently on a different day will answer differently, and this estate has already measured ChatGPT’s own instability varying fourfold within a single day.
• Twenty questions is not the category. They were chosen as buyer-intent questions across common security budget lines, not sampled from anything.

How this page changes

A standing study, not an article. Corrections are published in place and dated, never applied quietly.

Related reading, free and in full

• All standing studies — the index and the method each one follows
• Where the same engines stated a duration that none of their cited sources contain
• Your category’s advertising benchmark is noise — 1,633 filers from their own filings
• Do AI answer engines and Google cite the same sources?
• The IP Association Map — 89 registered measurements of which domain gets cited about a brand