Measurement

How to measure AI visibility

A proposed measurement protocol for brand mentions, URL citations, answer accuracy, and business relevance in AI search.

The direct answer

Measure AI visibility with a fixed set of buyer questions and recorded test conditions. Report brand mentions and URL citations separately, retain the original answers, and compare repeated observations using the same protocol.

Define the observation

Record the prompt, platform and model when available, search mode, account state, target market, language, timestamp, and run identifier. Save the answer and its cited sources. A failed or unavailable run is not an observed absence.

Separate branded and unbranded questions

A question that already includes Freeways measures brand-information accuracy. A question asking which agency can help a B2B business measures discovery. Keep these groups separate so branded prompts cannot inflate discovery performance.

Brand mention rate

Count valid observed answers that explicitly name the target brand, divided by all valid observed answers in the same group and period. Disclose the numerator and denominator. A mention may be neutral, negative, or irrelevant; classify its context separately.

Citation rate

Count valid observed answers with a link to the target origin, divided by all valid observed answers in the same group and period. Store the cited URL and the claim it supports. Citation and brand mention may occur independently. Do not count crawler requests as citations.

Accuracy and decision relevance

Compare service, pricing, audience, and other material statements with approved business facts. Record errors. Classify whether the answer is useful for the buyer's actual task and whether the brand is discussed as an option, a source, or a recommendation.

Interpret changes carefully

Keep repeated runs rather than selecting the most favorable answer. Log changes to prompts and model conditions. Treat uncontrolled before-and-after differences as observations, not proof that a page caused the change. Connect referral visits and conversions only where the analytics can support attribution.

Frequently asked questions

Is this a published benchmark?

No. This page describes a proposed protocol. It does not report measured performance for Freeways or another company.

Does crawl activity count as a citation?

No. A crawl is a request for content. A citation is a source reference visible in an answer. Track them separately.

What is missing from a single screenshot?

Usually the complete protocol, denominator, repeated observations, unavailable runs, and source history. A screenshot documents one answer, not a population-wide outcome.