TL;DR
- Ask what a GEO score measures, which inputs it sees and which outcome was used to validate it.
- Research found more citation-ranking signal when models used the query as well as the page. That is not a universal performance guarantee.
- Content quality, fit to a question and a live citation are distinct checks. Correlation and AUC values are not citation probabilities.
- Use page grades to locate issues within a disclosed rubric. Validate claims about visibility with saved answers and a repeatable, task-specific test.
A Grade Needs a Defined Meaning
A B2B team may receive a proposal to raise a page from one score to another. Before funding the work, ask whether the score describes text quality, technical access, fit to a buyer question or observed citations. A single number can hide those different jobs. A report should name the job rather than let “GEO-ready” stand for all of them.
Adding the Question Changed the Measured Signal
Bajemon and Rochet’s September 2026 preprint compared page-only scoring with models that also see the query. In its query-disjoint ranking evaluation, the fixed page score had a within-query Spearman correlation of 0.119 with citation allocation. The full fitted model reached 0.368. A single cross-encoder relevance signal reached 0.388.
The test covered 229 distinct queries and 1,269 query–engine groups across ten engine arms. It ranked sources by their share of cited answer text, with position weighting. It did not measure how often a real B2B buyer would see a page.
The values describe how well the methods ordered sources within questions. They do not mean that 11.9%, 36.8% or 38.8% of visitors saw a citation. One minus the coefficient is not an error rate. The simple relevance model did well. Check the test rather than assume that a larger scoring system must do better.
Bajemon and Rochet, 2026, Table 11. Query-disjoint protocol; 229 queries, 1,269 ranking groups, ten engine arms. Positive coefficients shown on a 0–0.5 axis; these are not citation probabilities or effects of editing a page. Study and Table 11.
A Separate Study Keeps the Engine in the Picture
Benjamin Tannenbaum’s September 2026 preprint reports a different result for a lexical match measure called BM25. In a 199-prompt cohort, its AUC for distinguishing own-domain citation was 0.545 for GPT and 0.641 for Gemini. AUC describes how well a measure separates cases under that test. It is not the probability that a specific page will be cited.
This is a different score and outcome from the first paper. The page crawl also came after the saved AI results. Some content may have changed. A score needs a named question, engine, date and outcome before a buyer can judge what its prediction means.
| Direct research | Actual result and unit | Boundary for a buyer |
|---|---|---|
| Bajemon and Rochet, 2026 | Within-query rank correlation: page score 0.119; full model 0.368; relevance signal 0.388 | Candidate-source ranking, not audience reach or a universal citation chance |
| Tannenbaum, 2026 | BM25 match vs own-domain citation: AUC 0.545 GPT, 0.641 Gemini; 199 prompts | Different test and engines; retrospective page crawl limits interpretation |
Figure 2. Evidence comparison. The first team works at TW3 Partners/Citead; the second author leads Aiso. Both develop commercial tools. These are preprints from different groups, not audits of every GEO tool. The second paper cites the first but reports a separate dataset.
The results address different parts of the problem. One tests source order when an engine is given candidate texts. The other links page match to saved engine results. Do not average their coefficients. Neither shows that raising a grade causes more citations.
What Would Make a Score Claim Worth Inspecting?
Start with the inputs. A score based on page text alone cannot know which buyer question you mean. Adding the question makes the task more specific. It still does not prove that a live system will find, choose or cite the page.
Then check the outcome. Was a citation a displayed link, a share of cited answer text, or a match to the site’s domain? Was the model tested on questions it had not seen while being fitted? The results can differ because the tests ask different things.
The first team found and fixed a test flaw: related query records had crossed the training and test groups. The final test keeps query groups apart. A model can look strong if it has seen parts of the test questions before. Ask for the corrected result, rather than the most striking early number.
Keep a practical distinction too. A page check can reveal a missing specification or an unclear answer that a buyer needs. Fix that defect for the reader, even if no citation gain has been shown. A grade change cannot replace a check of the published answer or the result of the work.
A Completed Record for a Score-Based Proposal
This example is fictional. A sensor software supplier receives a report saying that a 94/100 page score means a 94% chance of AI citation. The rubric reads page text but has no buyer question or live engine output. The team reviews the claim before approving a rewrite.
| Review field | Completed fictional record | Decision |
|---|---|---|
| Claim being sold | 94/100 means a 94% citation probability | Reject the probability claim; the scales have not been linked |
| Buyer question | Which tools export sensor logs as CSV without custom code? | Use the actual task to review page fit |
| Source fact | Approved product record says CSV export requires the paid plan | Keep that condition visible and check the public wording |
| Observed answer sample | Three of 20 usable saved answers cite the page under one named test setup | Report 3/20 for that sample; do not equate it with the rubric score |
| Revised deliverable | A clearer export explanation plus an evidence record of the tested answers | Accept content work for a defined defect; assess later visibility separately |
The buyer now has a page fix tied to a real task and a report that matches its evidence. This does not prove that the fix will change citation rates. Keep the old sample, failed checks and counting rule before making a later comparison.
If a score is only a consistent internal rubric, say so. Its repeatability can help spot a regression in a content pipeline. If a provider calls it predictive, ask for validation on the intended questions and engine setup, with limits on where the result transfers.
Our AI visibility measurement guide covers answer-side counting and sample records. Here the narrower task is to keep a page assessment from turning into a visibility promise without evidence.
GEO Page Scores: Common Questions
Are static page scores useless?
No. A clear rubric can support editorial checks and consistent reviews. Its value depends on the job it was designed and tested to do. A text-quality grade should not be presented as a live citation measurement.
Does query relevance guarantee retrieval?
No. A relevance measure assesses fit under its own method. A live system can use other sources or show no link to your page. Record the actual source output for the engine and conditions you intend to report.
Can we compare scores from two providers?
First compare their inputs, scale, outcome and validation. Equal-looking numbers may represent different constructs. Keep the provider and method version beside the score rather than combine them into a shared ranking.
What should count as completion of a page-scoring project?
Agree on the content defects to fix and how the final page will be checked. Keep any claim about citation change as a separate measured outcome. The studies do not establish that a target rubric grade guarantees success.
Bring the Score and the Claim Behind It
If a content plan promises AI visibility from a higher page grade, bring the rubric, buyer questions, page and validation report to SemShift’s GEO service discussion. We can discuss which proposed edits solve real content gaps and what evidence would be needed to assess the claimed visibility result.