Track AI Mentions, Citations and Recommendations Separately

An AI answer can name your company without recommending it, cite your article without naming your company, or recommend your product without linking to your website. Counting all three as the same success hides what happened.

Define the events before counting

Use an explicit coding guide. A mention is a recognizable reference to the brand. A citation is a source link attached to the answer. A recommendation is a favorable selection for the user's stated need, which requires reading the surrounding context.

Also record whether a citation points to your own domain or a third-party page discussing you. Those events suggest different follow-up actions.

Use a compact answer record

Record each reviewed answer in the same short format. One row per answer keeps the three outcomes apart and makes later counts easy to check.

Do not infer a website visit from a citation. A displayed link and a recorded referral session are different observations collected through different systems. First-party citation data follows the same rule: the Bing AI Performance report counts citations in Copilot and Bing answers but shows no clicks.

Worked example: a negative mention

Imagine a hypothetical answer saying, “Example Desk lacks the offline mode you need; consider another option.” A simple name counter records visibility. A decision-focused review records an unfavorable fit for that prompt.

This may be an accurate and useful answer. The next step is to verify the product fact, not to treat every unfavorable mention as a problem to suppress. If offline mode exists, investigate which public sources are outdated. Our guide to auditing AI brand sentiment explains how to code tone and accuracy separately.

Calculate rates with clear denominators

For a fixed set of recorded answers, mention rate is answers containing the brand divided by valid answers reviewed. Citation rate needs its own definition: answers with at least one own-domain link, or total links pointing to the domain. Do not switch between them mid-report.

Record failures and unanswered prompts separately. Removing failed runs can change the denominator and make two periods appear different even when brand behavior is unchanged.

Add a human calibration pass

Have two reviewers independently code a small shared sample, especially conditional recommendations and ambiguous brand names. Resolve disagreements and update the guide before processing the larger set.

AI can assist with labeling, but check borderline cases and a sample of ordinary cases. Store the original answer and source links so labels remain auditable.

Report the three outcomes separately, then explain their business meaning. Being cited as a technical reference may support authority; being recommended for a relevant buyer task may support consideration. Neither alone proves revenue impact.

Why one score misleads

A single "AI visibility score" combines three events that have different causes and call for different responses. The moment they are averaged, the number can move for reasons that have nothing to do with what the team is trying to change.

Consider a hypothetical month in which own-domain citations rise because a documentation page was rewritten, while recommendations fall because a competitor launched a feature that answers the buyer's constraint better. A combined score might not move at all. The team learns nothing, when in fact two important things happened at once, one of them good and one of them a product signal.

Separated, the three series each point somewhere specific:

The table is a guide to where to look first, not a proof of cause. But a combined score cannot even offer that.

A coding guide small enough to follow

Reviewers disagree most on recommendations, because "recommended" depends on reading the answer's reasoning. A short written guide with examples resolves most disagreements before they happen. A proposed minimum to adapt for your panel:

1. Mention: the brand name or an unambiguous product name appears. Misspellings that a reader would recognize count; a different company with a similar name does not.
2. Citation, own domain: a link to any page on a domain you control. Subdomains count; a partner's page about you does not.
3. Citation, third party: a link to a page whose main subject is your brand or product.
4. Recommendation, unconditional: the answer selects the brand for the user's stated need without a qualifier.
5. Recommendation, conditional: selected with a stated condition. Record the condition verbatim; it is often the most useful text in the whole answer.
6. Not recommended: the brand is mentioned and explicitly set aside for the stated need, with the reason recorded.

Add one real, anonymized example under each category as reviewers encounter them. The guide grows to a page and stops; a longer guide is not followed.

Report so that someone can act

A report that says "mention rate 34%, citation rate 12%, recommendation rate 9%" is a scoreboard. A report that helps a team decide something adds, for each series, the direction of change, the prompt families driving it, and one proposed check. For example: "Conditional recommendations rose; the condition in most cases was 'if you need offline access', which the product supports. Check whether the documentation page for offline access is reachable and current."

Keep the raw answers linked from the report. When a stakeholder questions a number, the fastest resolution is reading the three answers behind it, and that is only possible if they were stored with their date, prompt version, and platform.

Put this into practice

Copy the worksheet columns below into a spreadsheet and keep one row per item you check. The filled row is an illustrative example, not a reported customer result; replace it with your own verified records.

Use the following prompt only after supplying the records it requests:

Label each supplied answer using this coding guide. Return mention, own-domain citation, third-party citation, recommendation status, condition and supporting answer passage. Do not infer visits or conversions.

Research context

Brand appearances and source links describe different parts of an AI answer. The related Ahrefs starting points are Do AI Assistants Link When Mentioning Brands? For Ahrefs, Only 28% of the Time and How to Monitor and Win Brand Mentions in AI Answers. This guide’s checklist, examples and proposed workflow are independently written; they are not results of a SEOVision experiment.

Continue with the next task

These guides cover the next steps once you track the three events separately.

• GEO vs SEO: What Changes in AI Search—and What Does Not
• AI Search Optimization: A Practical Guide Beyond GEO Hype