Start with observable events
AI search visibility does not have one universal ranking position. Answers can vary by product, model, date, market, account state, prompt wording, and available sources. A useful baseline records these conditions instead of hiding them inside a proprietary score.
Google's AI search guidance connects visibility in AI experiences to established search foundations and useful content. Bing Webmaster Tools AI Performance adds citation and grounding-query reporting for Bing AI experiences. These sources provide different views and should not be forced into one metric.
Define a stable question set
Build a question set from real customer and market evidence:
- Category and definition questions
- Problem and workflow questions
- Product and service comparisons
- Alternatives and evaluation criteria
- Implementation and risk questions
- Brand and entity questions
Tag every question by audience, journey stage, market, and intended answer type. Keep a stable core set for trend comparison, but allow a separate discovery set for emerging language.
Do not prompt only for the brand. Non-brand questions show whether the company appears in category discovery.
Record the complete observation
For each check, record:
| Field | Why it matters | | --- | --- | | Surface and model | Different systems retrieve and compose differently | | Prompt | Wording changes the answer | | Date and market | Answers and sources change over time | | Brand included | A mention is an observable event | | Context | Inclusion can be positive, neutral, inaccurate, or irrelevant | | Cited URL | The source page reveals what was retrieved | | Other sources | Source diversity shows the surrounding evidence set | | Accuracy issue | Incorrect representation needs a separate action | | Follow-up action | The observation should support a decision |
Screenshots can preserve context, but structured records make comparisons possible.
Separate mentions, citations, and visits
These are different signals:
- A mention includes the brand or product in the answer.
- A citation links or attributes information to a source.
- A referral visit reaches the site from an answer surface.
- Branded demand may increase without a direct click.
- A qualified outcome is a signup, inquiry, purchase, install, or other valuable action.
A cited page is not necessarily a sales page. An informational source may influence discovery, while a later branded search or direct visit produces the conversion.
Use platform data where it exists
Platform reporting should remain separate from manual prompt checks. Bing's AI Performance reporting can provide citations, cited pages, and grounding queries for supported experiences. Search Console data can show Google Search impressions, clicks, and landing pages, but it should not be relabeled as a universal AI citation report.
Analytics can identify some referral traffic. Referrer handling, privacy controls, apps, and browser behavior may limit attribution. Record unknown and unattributed journeys instead of assigning them to AI without evidence.
Build a decision dashboard
A monthly decision view can include:
- Question coverage by audience and journey stage
- Brand inclusion rate on the stable set
- Cited pages and source diversity
- Accuracy and entity consistency issues
- AI referral sessions and qualified actions
- Branded demand alongside major content or PR activity
- Changes shipped and the next review date
Always keep the raw observations available. A summary percentage without the underlying prompts and dates cannot be audited.
Avoid false causality
AI visibility can move because of model updates, retrieval changes, news, competitor publishing, source availability, or prompt changes. A page update followed by a mention does not prove that the update caused it.
Use cautious language:
- "Observed after" is not the same as "caused by."
- "Cited more often in this question set" is not universal visibility.
- "No citation observed" does not prove the page was never used.
- "Mentioned" does not mean recommended or accurate.
Limitations
No measurement method can observe every answer shown to every user. Manual checks can be personalized or unstable, platform reports cover only their own surfaces, and referral attribution is incomplete.
The goal is not a perfect AI visibility score. The goal is a repeatable evidence set that helps a team improve access, entity clarity, useful content, source quality, accuracy, and qualified discovery without promising citations.
Primary sources
- Google Search AI optimization guide Google Search Central accessed
- Bing Webmaster Tools AI Performance Microsoft Bing Blogs accessed