How to evaluate AEO claims before paying for them

The short answer

Ask what was measured, where, over what period and against which comparison. A vendor’s reported result may be interesting without establishing that its product caused the result or that your business will get the same outcome. The checklist below is our proposed evaluation method, not a validated scoring system.

A technical fix is not a customer outcome

A fast response or a successful bot request shows something about access. It does not by itself show an explicit citation, a human visit or a new customer. Ask for the evidence at each step.

For our own service, the deliverable is a documented website improvement and a repeatable visibility report. We have not established a customer citation lift from that work.

Ask for the missing denominator

“Thousands of citations” needs context: how many questions, how many repeats, which engines, which dates and which citation definition? A crawler request is not a citation. A link to a directory is not automatically a link to your website.

For any percentage increase, request the original counts. Ask whether unsuccessful attempts, unchanged results and declines are included. A chart without those details is not enough to assess the claim.

What brand-correlation research actually shows

Ahrefs studied 75,000 brands and reported a stronger association between web mentions and AI Overview brand visibility than between backlinks and that visibility. This was a correlation study, not a trial showing that buying mentions makes an engine recommend a business.

For a service business, our proposed first step is to check that existing third-party descriptions are accurate. A broad study of brands does not justify a promised citation increase for a local pilot.

• Ahrefs: AI Overview brand visibility factors (May 2025)

Separate the search surfaces

Ask the provider to name the actual surface and collection method. A Google AI Mode observation is not a Google Maps ranking. A Gemini app answer is not an AI Overview. An API sample is not a reproduction of every consumer account’s experience.

For a proposed pilot, write the engine or surface, prompt, location, repetitions and settings into the protocol before collecting the baseline. Mark any later protocol change as a break in comparability.

Check how a study turns observations into advice

Semrush compared content-quality scores for AI-cited pages with a group of Google-ranking pages. Differences between those groups do not tell us how much an individual page will improve after an edit. A percentage difference in a content score is not a percentage increase in citations.

Write down the measured outcome, the comparison group and the collection dates before quoting a headline. Then ask whether the researchers changed anything themselves or simply observed differences.

• Semrush: content optimization study (January 2026)

Treat universal statistics cautiously

Before repeating an industry statistic, find its original study. Check the date, country, sample, definition and whether it describes searches, people, visits or bots. Those units are not interchangeable.

If you cannot trace a number to a suitable source, leave it out of the buying decision. A vendor case study without inspectable methods belongs in the “reported by the vendor” category, not in your forecast.

Use this five-question buying check

• What exactly will you change, and can I approve it?
• Which visibility metric is the primary outcome?
• Can I inspect full answers and citation URLs?
• What happens if the batch fails or results decline?
• What conclusion would you draw from no change?

Choose the next action from the evidence

If access is broken, fix access. If the service information is wrong, correct it. If a completed comparison shows a difference, report the observation and its limitations. Changes in engines, other websites and timing can also affect answers.

A small before-and-after pilot can guide the next experiment. It does not establish causation, market demand or a guaranteed return. Track inquiries separately and retain the source information supplied by the customer.