How to measure GEO: what an AI visibility report should show

You need to know which questions were tested, whether the questions changed, and what the AI said about your business. A higher mention count can include a recommendation for a service you stopped offering.

Generative engine optimization, or GEO, is work intended to improve how a business appears in AI search answers. A useful GEO report documents the questions tested and preserves the responses. It separates factual accuracy from visibility, then connects the findings to work someone can verify.

Start with the questions your customers ask

A question that names your company tests whether an engine knows something about you. A question that asks for a provider tests whether it brings you into the answer. Keep those results separate.

For example, "Does Acme offer bookkeeping for restaurants?" and "Who handles bookkeeping for independent restaurants in Tampa?" test different things. Combining them can hide weak discovery behind strong recognition of a name the tester supplied.

Choose questions from sales calls, support requests and inquiries. Keep the customer's location or buying constraint when it affects the answer. Record where each question came from.

Save the exact wording before the first run. A revised prompt belongs in a new comparison series; it should not quietly replace an underperforming question.

Make the visibility calculation inspectable

Consider this illustrative example, with invented numbers:

In each reporting period, a team tests the same 20 questions on two separate days, collecting 40 completed responses from one engine per period. The business appears in eight baseline responses and 14 later responses.

The mention rate rises from 20% to 35%. That is a gain of 15 percentage points, or a 75% relative increase in the sampled mention rate.

The arithmetic is valid. Calling it "75% more AI visibility" leaves out the size and scope of the test.

Those six additional mentions could reflect run-to-run variation. Repeat the same protocol before treating the change as a durable gain.

Ask to see the numerator and denominator beside the percentage. Check whether the team used the same engine and settings in both periods. Failed requests belong in the collection log, with an explanation of how the report handles them.

Keep results by engine. Pooling a different mix of engines each month can change the combined score even when performance within each engine stays the same.

Read the answer before counting the win

A company name in an answer can mean several things:

A citation to your tutorial may help a reader solve a problem without recommending your company. A recommendation may describe your business correctly while linking to another source. Both can be useful, but they answer different reporting questions.

Open the cited page. Find the passage that supports the claim. An answer that links to your homepage while inventing a price still needs a correction.

Define these classifications before reviewing results. Keep ambiguous cases visible so another person can inspect the judgment.

Keep enough context to repeat the check

For each observation, retain the exact prompt, the full answer and its visible source links. Record the date and time, engine, available model or mode label, language, and whether the session used search.

Include account context, memory and location settings where available. Mark unavailable details as unknown. A fresh conversation helps control prior-chat context, but does not make two runs identical.

Repeat the agreed questions on separate days. Preserve the runs where the business is missing as well as the favorable ones. The record should let someone trace a summary number back to the answers counted.

Describe the result as a sample of the chosen questions. It does not establish how often all buyers encounter your company.

Compare the sample with platform reporting

Prompt tests let you inspect the answers produced under your test conditions. Platform reports provide a different view of how your pages appeared.

Google's Generative AI performance report in Search Console reports impressions of links to your site in AI Overviews and AI Mode, with breakdowns including pages and countries. It does not count unlinked brand mentions. Use the report alongside your prompt log rather than treating a manual mention rate as a substitute for Google's measurements.

Microsoft's AI Performance reporting in Bing Webmaster Tools tracks citations across supported AI experiences. Microsoft explains that citation counts do not establish a page's placement or role within an individual answer.

Keep the measurement labels intact when presenting results. A Google impression, a Bing citation and a recommendation observed in a test are different events.

Review website inquiries separately. Record identifiable referrals and ask prospects how they found you. Treat self-reported discovery as self-reported; some buyers will visit directly after seeing an AI answer.

Ask what changed on the website

Ask for the changed URL and the date the edit went live.

For example, an answer gives an old service area. The team finds that area on an outdated location page, corrects the page and records the release date. Later checks show whether the same error recurs.

Document that sequence so you can investigate further without assuming that the edit caused every subsequent change.

A work log can stay short:

Check whether the proposed repair addresses a documented problem. Google's guidance for generative AI search says there is no special schema required for these features and that Google Search ignores llms.txt. Those deliverables alone do not demonstrate an improvement in Google AI visibility.

Use the report to choose the next job

Suede AI SEO, Suede AI's search visibility practice led by Jason Colapietro, describes its AI visibility monitoring around repeated buyer questions and separate records for mentions, accuracy and citations.

Use the findings to choose a specific repair. You might need to correct an outdated fact or explain a service that buyers cannot evaluate. A source repeatedly cited for an inaccurate claim deserves a closer look.

Before approving another month of work, choose one reported improvement and trace it back to the original and follow-up answers. If the report credits a website repair, inspect the changed page and release date too. If you cannot follow the record, ask for what is missing.