How to Measure AI Search Visibility

Save the exact wording of your questions. Before each round of observations:

• Record the platform, how you access it, and any model information you can verify. Keep observations from a web interface separate from API results. If the model version is not disclosed, record it as unconfirmed.
• Keep the question wording, instructions, language, target market and available search settings consistent.
• Decide how many times to repeat each question and record that number. Follow the same approach when retesting, starting a fresh, independent session for each repetition.

For Australian businesses, specify Australia as the target market and record the language used. Keep both consistent between rounds.

First, record the brand name and any name variants you intend to recognise. Check each signal separately: record 1 when it is present and 0 only when you have checked the complete response and confirmed it is absent.

Within one answer, count repeated mentions of the same brand only once. For each of the other two signals, count the same main domain only once, even if several of its pages appear or are cited repeatedly. Keep the three counts separate; do not add them into a single total.

For each observation, also save:

• The exact question, platform and access method, model information, market and language.
• The date, repetition number and whether the response completed successfully.
• The complete answer, any source list returned, and the specific page URLs associated with citation markers in the answer.

GEO measurement should check whether the answer describes the brand accurately and in the right category.

If the model or platform does not provide a source list, mark that field as N/A — not observable, rather than 0. If you cannot verify an in-answer citation, mark it as “unverified” or “not observable”, as appropriate. An unconfirmed result is not a zero. A URL appearing in a source list does not, by itself, establish that the answer cited it.

Keep API errors, timeouts and incomplete requests in a separate failure record. Do not count them as instances where the website failed to appear. A citation record identifies a reference in the answer; it does not, by itself, establish that the source supports the claim being made.

Compare baseline and follow-up results only under comparable conditions:

• Retain the original baseline and every subsequent round; do not overwrite them.
• Compare only question, platform and access-method combinations with valid observations in both periods.
• Mark combinations available in only one period as missing or exceptional, and exclude them from the change calculation.
• Retain every repetition, rather than selecting only the best result.

If the model version, question set, market, language, search method or recording rules change, establish a new baseline. Do not interpret a change in testing conditions as an improvement in your website's visibility. If you cannot confirm the underlying model version, retain that limitation alongside the comparison.

A change in one result does not establish a lasting improvement. Check whether it recurs in later observations under the same conditions. Success in one model does not establish improvement across all AI platforms, and a before-and-after difference alone does not show that your content changes caused it.

Visibility is not revenue

Visibility records alone cannot establish that:

• More impressions will lead to more visits.
• More visits will lead to enquiries.
• Enquiries have become sales or revenue.