Does the GEO score predict whether Perplexity cites a page? A pre-registered test

What we tested

The GEO score measures structural readiness: whether crawlers can reach a page, whether it can be read, quoted and attributed. It does not measure whether an AI system cites you today. We wanted to know whether, on real pages, a higher score goes with a higher chance of being cited.

What we did

• Sample. 341 pages, one per site, from the public Tranco popularity list, drawn across ten rank bands. We kept sites that were reachable, in English and with a sitemap, and chose one informational page per site. Pages whose headline contained the site's brand name were left out.
• Score. Each page was audited with the open-source GEO Optimizer CLI, version 4.19.3, rubric v2.
• Question. For each page we built two questions from its own headline, with fixed wording: "What are the most reliable sources on: [headline]" and "Explain this topic and cite your sources: [headline]". Each was sent once to Perplexity Sonar through its API, on 17 October 2026.
• Outcome. Whether the page's domain appeared among the sources Perplexity returned.
• Analysis, fixed in advance. A logistic regression of the outcome on the score, controlling for the site's popularity rank and the page's length. The plan was written before the full run; its SHA-256 hash is 441fcbb422964fdf2bf2b63eb02181ccbea9e7c5590d3e770399d567cd373ff8.

What we found

• Domain cited, first question: 42% of pages. Second question: 45%.
• Change in the odds of being cited per 10 score points: 0.96 (95% confidence interval 0.78 to 1.18).
• The same after resampling the data 2,000 times: 0.96 (0.76 to 1.20).
• The same, only for sites that do not block AI bots (253 pages): 1.04 (0.83 to 1.30).
• Correlation between score and being cited, controlling for popularity: 0.03.
• Eight of the nine score categories, one at a time (AI discovery had no variation in the sample): none significant.

An odds ratio of 1 means no difference. Ours is close to 1 and its interval includes 1. The interval excludes an increase of more than about 18% in the odds per 10 score points. It does not exclude a small effect.

Popularity matters, which is why the analysis controls for it. The share of pages whose domain was cited was 54% in the four most popular bands and 29% in the four least popular ones (141 and 128 pages), although it does not fall steadily from band to band.

What this does and does not show

• It does not show that structure is irrelevant. It shows that this score did not predict citation in this test.
• It does not cover ChatGPT, Gemini, Claude or Google's AI features. We tested one engine on one date with one model version, and engines change.
• The questions were built from each page's own headline. The test therefore measures whether a page is cited when the question names its own topic, not how it ranks against every competing source.
• The sample leans toward sites that are reachable, written in English and publish a sitemap. Sites that block our user agent are under-represented.
• The data are observational. We make no causal claim.

What we are doing about it

We are not changing the scoring weights because of this result. The score remains a measure of structural readiness, not a prediction of citations.

Data

The pre-registered analysis plan and an anonymised results table (one row per page: rank band, score, the category scores, page length and the outcomes; no domain, no address, no headline) are published in the GEO Optimizer repository so that the figures above can be checked. The audit and API-call scripts are not published.

Sources: the Tranco list at tranco-list.eu (public popularity ranking) and the Perplexity Sonar API.