How to measure AI search visibility: a reproducible observation protocol

A brand mention, a source citation, a recommendation, and a business enquiry are different outcomes. Combining them into one “AI visibility score” can conceal what actually happened. This protocol keeps them separate and makes each observation reviewable.

Publication scope: this is a measurement method and a blank recording template. It does not report a completed 72-response experiment, a market share estimate, or an improvement in Ramfa SEO’s recommendations.

Start with a fixed panel of questions

Choose 12 non-branded questions before viewing any results: four Swedish, four English, and four Persian questions. Include both commercial selection questions and practical troubleshooting questions. Record their exact wording and assign permanent IDs. For example, “How should I test hreflang on a website with Swedish, English and Persian versions?” tests useful source discovery; it does not require an answer naming a consultant.

Use three search-enabled AI products that are actually available in the test environment and repeat the panel twice. That produces 12 × 3 × 2 = 72 planned observations. A web-search API result is not automatically evidence of what a person sees in ChatGPT Search, Copilot, Perplexity, or Google’s AI interfaces. Record the actual product and mode shown on screen.

Brand questions such as “Who is Ramin Amirhaeri?” can check identity consistency, but keep them in a separate dataset. Including them in non-branded discovery rates would inflate the result.

Record the conditions, not just the answer

For every attempt, record the UTC timestamp, prompt ID and text, language, product, visible model or mode, actual test country, login status, and any visible personalization settings. Use a new conversation for an independent observation and avoid adding a follow-up that tells the system to recommend the brand.

Save the response and its source URLs, plus a screenshot or other reproducible evidence reference. If a product is unavailable, requires login, hits a limit, or fails technically, record that failure. Do not silently replace it with a different interface and keep the same label. Do not pretend that translating a prompt changes the geographic location of the test.

Code four outcomes independently

• Person mention: the response explicitly names the person being measured.
• Brand mention: the response explicitly names the business.
• Owned citation: a visible source link points to a website owned by that business. Record the exact destination and which claim it supports.
• Service recommendation: the response actually proposes the business or person as an option for the requested service. A citation in an educational answer is not automatically a recommendation.

Also distinguish service providers from publishers, directories, and software products. A documentation site may be a competing source for an educational question without being a competing consultancy. If identity is ambiguous, mark the row for review instead of guessing.

Use denominators that match the question

For a completed non-branded panel, recommendation rate equals valid responses containing a service recommendation divided by all valid responses in that same panel. Report the numerator and denominator alongside the percentage. Separate results by intent, language, and product; an educational question often should not recommend a provider at all.

Technical and login failures are not valid negative responses. Exclude them from that denominator and show their count separately. In Google AI Overviews, distinguish “no overview triggered” from “overview triggered, brand absent.” Report the trigger rate across eligible searches and presence within triggered overviews separately.

Do not pool different repeat dates, changed prompts, or different products without labeling the comparison. A small panel is a controlled sample of those conditions, not an estimate of every user’s AI search experience.

Keep platform reports and enquiries alongside the panel

Use the reports actually available in Search Console and Bing Webmaster Tools, recording their date range, property, filters, and definitions. A report with no data is not proof that the business never appeared anywhere. An aggregate grounding topic is not necessarily the exact wording a user typed.

Website analytics can record a consented visit or accepted enquiry, but missing referrer information does not establish that the visitor came from AI. An optional “How did you hear about us?” answer can add context. Keep personal enquiry details in the business workflow, not in public analytics event parameters.

Google’s official guidance on AI features describes eligibility and controls for AI Overviews and AI Mode. Technical eligibility does not guarantee that a page will be indexed or cited. A special AI text file is not a substitute for accessible pages, clear content, and accurate claims.

Human review is part of the method

A reviewer should compare each coded outcome with the saved answer and source list. Keep uncertain rows separate until reviewed, and record corrections rather than overwriting the history. Automated extraction can propose labels; the evidence should determine the final label.

Repeat the same panel at days 30, 60, and 90. Version any changed questions and log site releases, new articles, and changes in product availability. A higher citation count after a redesign is an observation; without stronger controls it does not prove that the redesign caused the change.

Download the blank log

Download the AI search observation template (CSV). It contains no prefilled performance claims. Keep raw response evidence in a controlled workspace and publish only material you are entitled to share.

This method belongs with a wider technical SEO measurement plan. For the implementation evidence from our own website, see the Ramfa SEO internal project. Prepared by Ramin Amirhaeri.

Where to go next

If observations will be summarised with an AI tool, use the human-review workflow to check each claim against the recorded response and source. For public background facts, keep the original pages and their readable alternatives together in a source directory.