Can a page score tell you whether AI search cites it?
A page score describes the properties that its scoring method checks. To know whether a particular search product cites that page, observe that product. A small study of live AI search illustrates why those are different measurements. It does not establish a universal citation rate or a formula for increasing visibility.
Read the study within its boundaries
Scoring With the Engine, published in September 2026, compared displayed citations for commercial prompts across search products. Citation sets differed substantially. The study used a narrow vendor-owned dataset, did not release the full raw log, and could not observe hidden retrieval. Its author sells AI-visibility software. Those limits make independent replication important.
The result supports a question for your own measurement process. Does the label on a report match what the report actually observed?
Write the question before choosing the metric
Consider an illustrative report for a consultancy. Its website passes a checklist for headings, contact information, and clear service descriptions. The report calls that an excellent AI visibility score. Yet the report contains no saved AI answer.
The checklist may still identify useful page improvements. It cannot tell the owner which services an assistant recommends, what the assistant says about the consultancy, or whether the owner receives qualified inquiries.
A more useful proposed report would give each measurement its own name. A page check can record missing or inconsistent facts. A citation check can record whether a saved answer links to the page. An accuracy check can compare the answer's claims with the business's verified information. An inquiry record can show whether a visitor contacted the business.
Keep unknown values empty and label them unknown. For example, a citation check that sees the final answer cannot establish whether the system also read uncited pages. Do not fill that gap with an inferred retrieval score.
Make repeated observations comparable
Start with questions that reflect an actual buying decision. Include the constraints a customer would provide, such as location or required service. Keep the original wording in the record.
For each observation, record the product, date, account or session conditions, answer, and cited URLs. Save an unsuccessful check too. A login error or an unavailable response is different from an answer that omits your business.
Choose the products that matter to your audience and report them separately. If one product changes, the report should make that visible. A combined average can conceal the location of the change.
This is a proposed measurement design, not a validated optimization tactic. Its value is that another person can inspect the saved answers and understand the claim.
Use the report to choose the next action
- Rename each existing score to describe its actual input and calculation.
- Add saved answers for a fixed set of relevant questions.
- Compare citation presence, answer accuracy, visits, and inquiries as separate outcomes.
If a page edit is followed by a change, investigate the explanation before assigning credit. Keep the observation useful even when the cause remains unknown.