GEO LAB
We change one thing, measure, and publish the result
GEO sells results without showing evidence. Here we do the opposite: one change per experiment, measured before and after with the same questions, and published even when it does not work.
How we run each experiment
- One hypothesis, one changeFor example: “adding price and currency to product structured data increases appearances”. Nothing else is touched during the experiment.
- Questions fixed before we start10 to 20 real buying questions from the sector, written and published before the change.
- Repeated measurementEach question is run several times, on more than one AI, before and after. We publish the share of appearances and its range, not a single number.
- A control storeA similar store we leave untouched, measured the same way. Without it, the effect of the change cannot be told apart from the normal variation in AI answers.
- WaitProtocol window: 3 to 4 weeks after the change before the final measurement. It is not a guaranteed update time for the assistants.
- Always publishResult, method and limits. Including experiments where nothing changes or things get worse.
The sheet for each experiment
HypothesisWhich change is expected to improve what.
ChangeWhat was modified, with before and after.
QuestionsThe exact list, fixed before starting.
MeasurementAI used, repetitions and dates.
ResultAppearances before and after, with range, plus the control store.
Conclusion and limitsWhat can be claimed and what cannot.
Experiments
FINISHED · 27/10/2026 · NO EFFECT ATTRIBUTABLE TO THE CHANGE
Experiment 0 · Does publishing clear, structured information make AI answer better about it?
HypothesisAdding an llms.txt section and FAQPage markup about the «Legible por IA» badge makes ChatGPT answer better about the badge.
ChangeOn /verificar/: a badge section in llms.txt and FAQPage markup (JSON-LD) with 3 questions and answers, in Spanish and English. Control: this page, /laboratorio/, left untouched during the experiment.
Questions10, fixed on 5/10/2026 before measuring: 5 about the badge and 5 about the Lab.
MeasurementChatGPT without signing in, a new chat for each answer, 3 runs per question: 30 answers per measurement. Dates: 5/10 (before the change), 12/10, 19/10 and 27/10.
Correct = accurately describes what Comprable’s site says. Final measurement: /verificar/, 9 correct, 3 partial, 2 wrong and 1 «does not know»; /laboratorio/, 12 correct and 3 «does not know».
The change does not explain the improvement. The control page, which was not touched, improved as much or more (from 0 to 12 of 15) than the changed page (from 0 to 9 of 15).
What the data does show: in three weeks, ChatGPT went from finding nothing about Comprable to answering 21 of 30 questions correctly and linking to comprable.co in 21. The simplest explanation is that ChatGPT found and read the pages, not that the added markup helped.
Remaining errors: in 2 of the 15 answers about the badge, ChatGPT confused it with the EU AI Act labelling of AI-generated content, in 1 it attributed it to another company, and in 2 it made up part of what the badge checks.
Limits: one site, one AI and 30 answers per measurement, with a control page instead of a control store. Declared deviation: the final measurement was taken on 27/10 instead of 26/10, and the one planned for 2/11 was dropped. The figures describe this case, not a general rule.
SHA-256 of the pre-registration: 3c34a97dd14ba75d4e821cac65114d8d3f82a4901c694f067a60f261e9eabf33
The next experiment will be published here once it has its baseline measurement.
What we can claim and what we cannot. This is not causal proof. With few stores and AI answers that vary between runs, what we publish is “in this case, after this change, this happened”. We do not say a change works in general and we do not promise rankings.
Want your store to be a case?
If you would like to take part, write to us from the site. Always with your written permission to publish data.
Write to Comprable