What metrics measure AI visibility?

Four main ones: mention rate, explicit recommendation rate, listed-only rate, and citation source share. The key is separating strong signals like "AI says you're the best" from weak ones like "merely listed," not conflating them.

How can I tell if AI recommends my brand?

Use a scenario-based question bank, run each question multiple times in a neutral, memory-free environment, and check per question whether AI names you, ranks you, and cites your page — with a screenshot each.

Why can't a single screenshot be a conclusion?

AI answers vary naturally. Ask the same question twice and the brands named and sources cited can differ. One screenshot may happen to mention or miss you, so only multiple runs per question reveal your true position.

How does domestic monitoring differ from overseas?

Same method, different engines. Overseas tests ChatGPT, Perplexity, Gemini, and Google AI Overviews; domestic tests Doubao, Qwen, and DeepSeek. Retrieval pools and e-commerce ecosystems differ, so the question bank is rewritten for each.

Does monitoring guarantee AI will recommend me?

No. Monitoring only measures and records your brand's real position in AI answers. Ranking and inclusion are decided by the platforms, so we make no such promise. To see your current baseline, start with a free audit.

How to Measure AI Visibility: Metrics and a Reproducible Method

Why Measure AI Visibility

More and more buyers open ChatGPT, Doubao, or Perplexity before they buy or shortlist a vendor, asking things like "what should I buy" or "which one is reliable." Whether your brand shows up in those answers, and how it gets described, decides whether you make the shortlist. Yet most brands are completely blind to this layer, and you can't optimize what you can't see.

AI visibility monitoring exists to fix that blindness. AI visibility monitoring is the practice of using a question bank built around real buying scenarios to measure, in a neutral and reproducible environment (no account memory, multiple runs per question), how often a brand is mentioned, recommended, and cited in AI answers. It is not a single screenshot of what AI says about you; it turns "how AI actually answers real questions" into comparable, recomputable data.

Why not rely on one screenshot? Because AI answers vary naturally. Ask the same question twice and the wording, the brands named, and the sources cited can all differ. A single capture may happen to mention you or happen to miss you, and neither can be treated as a conclusion. To judge your true position, you have to run every question multiple times and record the variance honestly, without cherry-picking.

Four Core Metrics: Keep Strong and Weak Signals Separate

One "mention rate" is not enough. We break AI visibility into four metrics, and we always separate strong signals from weak ones:

Mention rate — across multiple runs, does AI mention you at all? This is the baseline threshold.
Explicit recommendation rate — does AI actively call you "the best" or "the top choice," rather than listing you passively? This is the strongest signal.
Listed-only rate — AI simply lists you alongside other brands with no preference. This is a weak signal.
Citation source share — which sites AI cites when it builds an answer, and whether your own site or third-party content is among them.

Measuring AI visibility means separating strong signals from weak ones: "AI actively says you're the best" and "AI merely puts you on a list" are two signals of completely different strength, and they must not be conflated. Counting a weak signal as a strong one overstates your real position and sends your optimization off course. Honest monitoring always records and reports the two categories separately.

A Five-Tier Position Model

To apply one consistent ruler to "how AI actually treats you," we score each question against five position tiers:

Explicitly named best — AI actively calls you the best or top choice.
Top three — you appear in a leading recommendation slot in the answer.
In the list — you show up in the list, but with no clear preference.
Source only — AI doesn't mention you in the body, but cites your page as a source.
Absent — AI neither mentions nor cites you.

Tag every question with one of these five tiers, run a batch, and your brand's true distribution across buying scenarios becomes obvious: where you sit up front, and where you're completely absent. That distribution, not gut feeling, is what decides which content gap to fill first.

A Reproducible Testing Method

Reproducibility is the core of this method: the same question bank, run by a different person on a different device, should produce close results. The flow has five steps:

Build a scenario-based question bank: don't just test brand terms. Lay out the questions real buyers actually ask along the purchase path — category blind tests, e-commerce scenarios, brand comparisons, reputation probes, and entity disambiguation, layer by layer.
Run multiple rounds in a neutral environment: use a clean environment with no history, running each question multiple times so account preferences don't contaminate the result.
Screenshot every question: capture the question, answer, and cited sources for each item. Quoted lines in the report are continuous excerpts from the AI's original text, not stitched together.
Aggregate citation sources: roll up the sites AI actually cites into a distribution, and let that distribution decide where content goes.
Net out drift: on re-test, first subtract the AI's natural variance, then look at the real net lift after content goes live, instead of mistaking noise for effect.

This method scales concretely. In one real monitoring project, the team used a 96-item, five-layer question bank, ran 960 real-device tests across three platforms, and aggregated roughly 2,945 answer citation sources — all figures observed during the monitoring period. The real-device app is the acceptance standard, so anyone can re-test on a phone on the spot.

To see this method fully applied to an e-commerce category — from a five-layer question bank to a 90-day plan — read the e-commerce GEO case study.

Common Mistakes: Three Ways to Get a False Conclusion

The biggest traps in AI visibility monitoring aren't technical; they're methodological. Three common mistakes push conclusions off course:

Feeding boilerplate to AI for scoring — hand AI a self-congratulatory blurb and ask it to score, and you get a confident hallucination: AI invents a flattering conclusion from your input that has nothing to do with your real position. Always run an input hygiene check before scoring.
Testing with an account that has memory — the account you use daily has already been trained on your preferences, so results come back far rosier than reality and the baseline is contaminated from the start. Baselines must run in a clean, history-free environment.
Treating a single run as a conclusion — asking once and concluding is like judging a coin from a single flip. AI answers vary, so a single run is neither reproducible nor comparable.

Avoid these three and your numbers hold up: clean environment, multiple runs, input hygiene first. That's what separates monitoring from a lucky screenshot.

Domestic vs. Overseas: Different Engines to Test

The same methodology tests entirely different targets at home and abroad, and that must be designed for separately.

Overseas monitoring watches ChatGPT, Perplexity, Gemini, and Google AI Overviews, where buyers of global-facing brands do their early research. Citation preferences differ by engine: for the same question, who gets cited and how it's answered in ChatGPT versus Perplexity must be recorded separately.

Domestic monitoring watches Doubao, Qwen, and DeepSeek. Their retrieval source pools and how they connect to e-commerce ecosystems differ from overseas engines, so the e-commerce scenarios and platform mechanics in the question bank must be rewritten for local habits, not translated wholesale from the overseas set.

Home or abroad, the underlying logic is the same: write questions around real buying scenarios, run multiple rounds in a neutral environment, document every question, and separate strong signals from weak ones. The engines and ecosystems change; reproducibility is the line that doesn't. To see how monitoring threads into a full observe-discuss-respond-verify growth loop, read The AI Growth Loop for Global Brands.

FAQ

What metrics measure AI visibility?: Four main ones: mention rate, explicit recommendation rate, listed-only rate, and citation source share. The key is separating strong signals like "AI says you're the best" from weak ones like "merely listed," not conflating them.
How can I tell if AI recommends my brand?: Use a scenario-based question bank, run each question multiple times in a neutral, memory-free environment, and check per question whether AI names you, ranks you, and cites your page — with a screenshot each.
Why can't a single screenshot be a conclusion?: AI answers vary naturally. Ask the same question twice and the brands named and sources cited can differ. One screenshot may happen to mention or miss you, so only multiple runs per question reveal your true position.
How does domestic monitoring differ from overseas?: Same method, different engines. Overseas tests ChatGPT, Perplexity, Gemini, and Google AI Overviews; domestic tests Doubao, Qwen, and DeepSeek. Retrieval pools and e-commerce ecosystems differ, so the question bank is rewritten for each.
Does monitoring guarantee AI will recommend me?: No. Monitoring only measures and records your brand's real position in AI answers. Ranking and inclusion are decided by the platforms, so we make no such promise. To see your current baseline, start with a free audit.

Want to see your brand's current AI visibility baseline? We run a real test using public information and lay out the issues worth tackling first.

Get your free Growth Audit