How to measure AI search visibility

What counts as a mention or citation?

Brand mention

The response names the business. Check that it means the correct business, not a similarly named company. Record positive, neutral, negative or unclear context.

Linked citation to your site

A visible source link points to your domain or one of its pages. Open it and save the actual URL. A source link alone does not prove that the response recommends the business.

Third-party source about the business

The answer cites another site that discusses the business. Record that separately from a link to your own site, then check whether the third-party page really supports the statement.

Recommendation

The answer explicitly suggests the business as a fit for the user's request. A name in a source list or a warning is not a recommendation.

Referral or enquiry

A visit or contact that can be attributed to the source using available evidence. Missing referrer information and consent limits mean not every visit can be identified. A prompt test cannot measure sales.

These categories can overlap. A response can recommend a business without linking to it, or cite a page without naming the business in the answer. Do not add the categories together and call the total unique visibility.

Choose questions that a customer might really ask

Begin with sales conversations, support questions and relevant Search Console queries. Group questions by the service, place and buying decision they represent. Label suggested prompts as candidates unless you have actual evidence that customers ask them.

Illustrative prompts for an independent search consultant:

• What should a local SEO audit include for a homestay in Kodagu?
• How can a small Karnataka business compare SEO consulting proposals?
• What is the difference between AEO and GEO services?
• How should a business measure whether ChatGPT cites its website?

Keep branded questions in a separate group. “Tell me about Fawaz BH” tests identity and accuracy; it does not test whether the business is discovered without a brand cue. Do not improve the apparent result by adding your name or asking the engine to recommend you.

A manageable pilot could use 20 questions with two repeats per engine. That is a practical starting sample, not a statistical standard. Preserve question IDs and a sample version so a changed set cannot silently replace the baseline.

A repeatable manual test

1. Record the conditions. Save the engine, displayed model if available, search or browsing mode, date and time, language, region and whether the session is signed in. Record unknown fields as unknown rather than guessing.
2. Use a fresh conversation. Enter the exact saved prompt without earlier messages that name the business. A fresh conversation reduces conversational influence but does not eliminate personalisation.
3. Save the full answer and sources. Keep a screenshot or permitted response export with an evidence ID. Avoid uploading confidential customer information. Store raw responses in a private workspace, not in a public audit folder.
4. Classify what happened. Mark mentions, direct-site citations, third-party citations and recommendations independently. Check the linked page and any factual errors.
5. Keep failures visible. Distinguish a timeout, refusal, unavailable search feature and a valid answer with no citation. A failed request is not proof of absence; a valid uncited answer is a real observation.
6. Repeat on the same schedule. Use comparable settings and keep the negative results. Record site changes and prompt-set changes alongside each run.

This guide provides a manual workflow and a static worksheet. It does not connect to AI APIs or run automated prompts. If you later use a monitoring product, check its collection method, terms, locations and repeat policy before comparing its results with your manual sample.

Use a denominator people can inspect

For one engine, one mode and one sample version:

• Mention rate: valid responses mentioning the correct business ÷ all valid responses × 100.
• Direct citation rate: valid responses linking to your site ÷ all valid responses × 100.
• Recommendation rate: valid responses explicitly recommending the business ÷ all valid responses × 100.
• Test completion: valid responses ÷ attempted tests × 100.

Illustrative calculation: 20 questions tested twice create 40 attempts. If four attempts fail and nine of the 36 valid responses mention the business, the observed mention rate is 25%, with 90% test completion. It is not a 25% share of all customer prompts. Repeats are not necessarily independent observations, so this simple rate is descriptive rather than a confidence estimate.

Show engine-level counts before any summary. Keep answers without citations in the valid-response denominator. If you choose a different definition, state it and use it consistently. Report branded and unbranded groups separately.

Compare equivalent periods, but do not automatically attribute a change to your latest article. Engine updates, source availability, test location and ordinary response variation can change the result. Combine this log with relevant search data and qualified enquiries rather than replacing them.

Download a blank AI visibility log

The CSV contains a header row only. Add one row per prompt attempt, including failures. It is free to download without a form, email address or account.

Use valid, timeout, refusal or unavailable for status. For outcome columns use yes, no or unclear. Store evidence privately and place its reference in the sheet. If an answer has several citation URLs, separate them inside the cell with a vertical bar.

When importing text from an AI response or an external source into a spreadsheet, treat it as plain text. Some spreadsheet applications interpret cells beginning with symbols such as = as formulas. The downloadable blank file contains no formulas, scripts or external connections.

Sources and limits

Google Search Central's AI features guidance explains its own website and reporting context. The test protocol and calculations above are a proposed practical method, not an official platform metric or an independently validated market-share model.

For the service definitions, read SEO vs AEO vs GEO. For implementation support, see GEO consulting and AEO consulting.