Does the GEO score predict whether Perplexity cites a page?
In our test it did not. Across 341 pages, a higher score was not associated with a higher chance of being cited by Perplexity Sonar, once site popularity and page length were taken into account. This page gives the number, the limits and the files.
Result
Logistic regression of “cited” on the score (per 10 points), the log of the site’s Tranco rank and the log of the word count. An odds ratio of 1 means no association. The analysis plan was written before the full run; its SHA-256 is in the repository.
| Outcome | Odds ratio per 10 points | 95% interval | p |
|---|---|---|---|
| Primary: domain cited, question template A | 0.96 | 0.78 to 1.18 | 0.717 |
| Domain cited, template B | 0.97 | 0.78 to 1.19 | 0.744 |
| Exact page cited, template A | 1.06 | 0.84 to 1.34 | 0.643 |
| Exact page cited, template B | 1.05 | 0.83 to 1.32 | 0.693 |
The mean score was 65.5 both for cited pages (144) and for pages not cited (197). None of the nine categories showed an association after correcting for multiple comparisons.
Limits
- — One engine (Perplexity Sonar), one date (17 October 2026), one model version. Nothing here says how ChatGPT, Google AI Overviews or Claude behave.
- — The question was built from each page’s own headline, so the outcome is whether a page is cited when the question names its own topic.
- — The sample favours reachable, English-language sites that publish a sitemap.
- — Observational data. A flat curve does not show that the checks are useless for other goals (access, indexing, clean markup), and it does not show that they are useful for citation.
- — The confidence interval is wide enough to include a modest effect in either direction.
What this means for the score
The score measures how many readiness conditions a page meets: crawl access, markup, structure, entity clarity. We do not present it as a predictor of citations, and we make no promise that raising it changes what an AI engine answers.
Check it yourself
- results.csv: one row per page, no domains or headlines.
- analysis-results.txt: output of the pre-registered analysis.
- PREREGISTRATION.md: the plan, with its checksum.
- scripts/: sampling, audit and analysis code. The raw files with domains stay private.