How to diagnose AI visibility without overreading the score

Where does the brand appear in AI answers—and what evidence should you act on?

An AI visibility diagnostic asks whether selected answer engines mention the brand, cite its domain, mention configured people, or surface competitors for a grounded set of prompts. It also preserves the answers and cited sources behind the summary so the score is not the only thing you can inspect.

The useful decision is not “How do I force the percentage up?” It is “Which important prompt is the brand losing, what source does the engine already trust, and is there an evidence-backed PR action worth taking?”

Free, standard, and deep scans answer different versions of the question.

A scorecard with the underlying answers still attached.

• Mention evidence: brand mention rate, and on paid scans an unaided rate that isolates non-brand discovery, topical, and competitor prompts.
• Citation evidence: whether the brand’s own domain was cited and which third-party URLs appeared in losing answers.
• Per-engine evidence: results grouped by ChatGPT, Claude, Gemini, Perplexity, Exa, or Grok when included in that scan.
• Competitive evidence: configured competitors that appeared in successful answers and the prompts where they surfaced.
• Change context: a comparison with an earlier completed paid scan when a comparable baseline exists.
• Action evidence: losing prompts and cited sources that can be reviewed before an outreach draft enters the normal approval queue.

Read the numerator, denominator, prompt type, and engine before the percentage.

Mention rate and citation rate are not interchangeable

A brand can be named without its domain being cited, or cited without being the central recommendation. Open the answer and source evidence before treating either metric as a win.

Brand prompts and discovery prompts measure different things

Direct brand prompts test what engines say when the brand is already named. Non-brand prompts test whether it is discovered when a buyer asks about the category, problem, competitors, or topic without supplying the answer.

Disagreement is a finding

If one engine mentions the brand and another does not, that is not an error to average away. Review the different answers, sources, and retrieval behavior. A partial scan should stay labeled partial.

A delta is observed change, not attribution

A later scan can show that a mention or citation changed. It cannot by itself prove that a specific pitch, article, site edit, or campaign caused the change.

Prioritize one defensible losing query at a time.

1. Choose an important non-brand query where the brand is absent and the engine call completed successfully.
2. Open the answer and verify the cited page is real, relevant, current, and realistically influenceable through earned media or source outreach.
3. Decide whether the best action is coverage, a correction, a source contribution, an owned-page improvement, or no action.
4. If outreach is appropriate, review the proposed source and draft in RunPR’s pitch queue. Edit, hold, reject, or approve it.
5. Re-scan on the normal cadence and describe any movement as observed change unless stronger causal evidence exists.

Run the live 10-query audit before committing to a paid scan.

Use the free AI Visibility audit with a real company domain. The report shows the generated queries, completed-answer count, brand mention and citation rates, losing queries, winning queries, and cited URLs. That makes it a useful small-sample orientation—not a substitute for the recurring 24-prompt paid workflow.

For the full product context, see AI Visibility and the RunPR feature overview.

AI answers are sampled, changing, and provider-dependent.

• Results are a point-in-time sample of selected prompts and engines, not a census of every question a buyer could ask.
• Answer engines and underlying models change. The same prompt can produce different answers later.
• Prompt wording and grounding matter. An unreadable homepage may require a user-provided description, which changes the input source.
• A provider can fail, omit citations, or return an answer that needs manual context. Failed calls should not be counted as successful misses.
• Mention and citation detection are useful indicators, not measures of revenue, sentiment, purchase intent, or guaranteed recommendation quality.
• Scan-to-scan movement is not proof that RunPR—or any single PR action—caused the change.

RunPR surfaces evidence and prepares action; a person owns the claim and the send.

A visibility result does not authorize outreach. A human should verify the answer, cited source, proposed angle, factual support, recipient, and final language. Any resulting pitch stays inside RunPR’s normal review boundary, where it can be edited, held, rejected, exported, or approved according to the organization’s configured workflow.