We asked two AIs to prove us wrong. One corrected the other.
Before claiming that our schema engine does things nobody else does, we had it examined by Gemini and ChatGPT with explicit instructions to find our faults. Six points held up; three claims we removed before publishing them.
Published: September 13, 2026 · Analyses by Google Gemini and ChatGPT, commissioned by us
The method, because it matters. We described nine features of the engine and asked which already existed elsewhere, with plugin names, versions and links to documentation. The prompt banned compliments and asked where we were worse. Then we compared the two answers: only what they agree on counts.
Six points neither found documented elsewhere
On these points both analyses say the same thing, having looked at Yoast, Rank Math, All in One SEO, SEOPress, Schema Pro, Slim SEO, The SEO Framework, SASWP and Squirrly.
Working out the trade from the site
The others have you pick the type by hand from a menu — Rank Math with its 193, Yoast, AIOSEO, SEOPress, Schema Pro. What is automatic elsewhere is the content type: product → Product, post → Article. Not the trade, inferred from the text.
Weighing different sources
No documentation describes a system that adds up clues with different weights. Where Google Business integration exists, it is there to import the address — not as a clue about who you are.
Stopping when it does not know
The thresholds deciding whether to apply, flag or merely suggest have no documented equivalent. It follows from the point above: without a confidence score there is nothing to stop on.
Saying why, and what it rejected
Others show a code preview or a validator. Nobody shows the reasoning, or the rejected signals — because there is no reasoning to show.
Words that mislead
“image bank” not making it a bank, “applying plaster” not making it an app: that belongs to language-processing libraries, not plugins. None of the competitors parse text to assign the entity, so they never need it.
Not declaring premises that are not there
The others emit the user’s chosen type even with no address, producing warnings in Google’s test. Nobody downgrades the type when the evidence is missing.
One of the two also rewrote the sentence we are allowed to use, and we are adopting it as it stands: “among the main plugins we analysed, we found no documented system that automatically identifies the type of business”. It states what we verified, without pretending to know every plugin in existence.
Three claims we removed
The draft of our announcement contained these three as well. The check stopped them before we did — which is exactly why we asked for it.
- “A register of 171 trades”, said as a first
- “A single JSON-LD graph, unlike the others”
- “We separate the maker from what it sells”
- Rank Math documents 193. We are below it
- Yoast has done it since 11.0, so do Slim SEO and The SEO Framework. And Google accepts separate blocks linked by
@id: one block is not even better - Standard since 2019. Claiming it “would be discrediting”
The first is the most instructive: the number that felt impressive is smaller than a competitor’s. You do not discover that by re-reading your own copy — only by going to look at everyone else. Which is why one outside check is worth ten internal reviews.
Where one corrected the other
Asking both was not caution for its own sake: it paid off. Three relevant facts came from only one of the two, and they change the conclusions.
| Fact | Why it matters |
|---|---|
| Rank Math’s 193 types | Cancels our numerical “first” |
| WordLift, which uses language processing and entities | That is the real benchmark, not Yoast |
| A schema plugin closed in May 2026 over a security issue | It had to come out of the comparison: it no longer exists |
There is also a difference in method worth telling, because it concerns anyone using an AI to get informed: one of the two backed almost every claim with links that were web searches, not documentation pages. Seven out of nine. The other cited real pages, checkable one by one. Their conclusions matched almost everywhere — but only one could be verified.
Four faults we accepted
These are not polite remarks: we are fixing them.
We count the same evidence more than once. Site name, tagline, H1 and title are worth four points each. But on a site called “Rossi Dental Clinic”, with matching H1 and title, that is sixteen points from one editorial decision. The score measures how many times we re-read the same thing, not how many independent clues we have. The fix is to group signals and cap each group.
The premises test is too generous. We accept an address or a phone number or opening hours. But an entirely online company has a phone line and support hours: it is not a place you walk into. It needs two tiers — strong evidence (address, Google profile with an address, coordinates) and supporting evidence (the rest).
“87% confidence” is the wrong word. It is not a probability: it is a sum of weights we chose. Calling it a percentage invites the right question — calibrated against which set of sites? It will become a score, not a percentage.
The manual choice is too weak. If you declare “dental practice”, you should not be outvoted by the sum of title, domain and tagline. It will become a decision that simply wins.
“Don’t step aside”
On one point they agreed flatly, and it is the one that weighs most. Today, if another SEO plugin on the site already builds its own graph, AISA stops emitting the nodes it produces. Sensible, to avoid duplicates — but it means the detection engine, the part that carries the value, produces nothing precisely on sites running Yoast or Rank Math. Which is most of them.
The direction suggested is to go inside their graph rather than sit beside it: reuse their identifiers and add only the entities and properties that are missing. This is not a hack: Yoast itself recommends it to third-party plugins in its developer documentation.
It was already written in our code as the next phase of work. Now we know it is the priority, not an extra.
Including the parts that do not suit us
An analysis commissioned to agree with you is worth nothing, to you or to the reader. We asked to be proved wrong and in part we were: three claims were false, four criticisms were fair. The rest — six points where neither found a documented equivalent — we can now say knowing why.
Try AISA freeAnalyses produced by Google Gemini and ChatGPT on 13 September 2026 at our request, from our own description of the plugin. Comparisons rest on the public documentation of the plugins named, as of that date: third-party features, tiers and pricing can change at any time. The AISA features described are those of version 1.99.999.2; the corrections mentioned are in progress for 2.0.