Most brands discover an AI visibility problem after the fact. A sales call where a prospect mentions they found a competitor through Perplexity. A quarter where organic traffic holds but demo requests drop. Neither moment gives you data. Without a measurement system, you cannot tell whether you have a citation gap, a content problem, a technical access issue, or an authority gap.
This playbook gives you a repeatable way to measure AI search visibility before you buy software or commission a full audit. It covers what to track, which platforms matter, how to build a prompt set, what to record, and how to score the results. A spreadsheet and 30 minutes get you a real baseline.
Key Takeaways
- Track four metrics separately: mentions, citations, competitor presence, and answer accuracy.
- Start with ChatGPT Search, Perplexity, and Google AI Overviews, the three platforms that drive most AI-influenced purchase decisions.
- Build a prompt set of 20 to 50 queries across five categories: branded, category discovery, problem-aware, comparison, and proof-seeking.
- Run each prompt at least three times per session to account for non-deterministic variance.
- Even brands with strong Google rankings appear in only 18% of AI answers on average — a citation rate below 10% means effective invisibility.
- The pattern in your data tells you which layer to fix first: access, understanding, or authority.
What you are actually measuring
"AI visibility" is too vague to act on. There are four distinct metrics, and each points to a different root cause.
| Metric | What it means | Why it matters |
|---|---|---|
| Mention | AI names your brand in the answer text | Signals entity recognition: the model knows you exist |
| Citation | AI links to a URL on your domain as a source | Signals source authority: the model trusts your pages as evidence |
| Competitor presence | Competitors appear in answers where you should | Reveals relative visibility: what you're losing, not just what you're missing |
| Answer accuracy | AI describes your brand, offer, and audience correctly | Signals understanding: wrong answers can damage buying decisions |
These metrics are not interchangeable. A brand can have a high mention rate and a low citation rate. ChatGPT names you, but it's pulling information from a competitor roundup that happens to reference you. A brand can have strong citations on one educational article while the service page is never retrieved. A brand can appear correctly in branded prompts and be entirely absent from every category and problem-aware query a real buyer would actually use.
Track all four separately before drawing any conclusion about what to fix. Yext's AI visibility framework maps to the same structure: presence, sentiment, and comparative position, three signals that correspond directly to mentions, accuracy, and competitor presence above.
The gap is larger than most teams expect. Google's May 2026 AI SEO guide found that even brands with strong Google rankings appear in only 18% of AI answers. Solid SEO does not translate automatically to AI visibility. That is exactly why tracking all four metrics separately matters: the problem is rarely in one place.
Which platforms to track
The same brand can perform very differently across AI search platforms. Frase's AI tracking guide and Search Engine Land both recommend testing across at least three: ChatGPT, Perplexity, and Google AI Overviews.
| Platform | How citations surface | Priority |
|---|---|---|
| ChatGPT Search | Inline citations and Sources panel when Search mode is active | Primary |
| Perplexity | Source panel on nearly every answer | Primary |
| Google AI Overviews | Embedded links in the Overview block above organic results | Primary |
| Gemini | Citations vary by query; web references appear in some answers | Secondary |
| Microsoft Copilot | Bing-indexed pages; inline citations in most answers | Secondary |
Start with the three primary platforms. They account for the largest share of AI-driven category and purchase-intent queries. Add Gemini and Copilot once your primary baseline is established.
Cross-platform divergence is normal and worth treating as its own finding. Being well-cited in Perplexity does not mean you appear in Google AI Overviews. The platforms use different source signals, different crawlers, and different retrieval logic. Treat each as a separate measurement track rather than collapsing them into a single score.
One thing most tracking guides skip: both Perplexity and ChatGPT depend on the Bing index for real-time answers, not just Google. If your site is not indexed in Bing, a significant share of AI-sourced citations are unavailable to you regardless of how well you rank on Google. Bing Webmaster Tools is currently shipping GEO reporting features before Google Search Console — right now it is the most actionable free platform for tracking AI-driven impressions and diagnosing access gaps.
For ChatGPT specifically, verify whether the answer was generated with Search mode active. Citations only appear in that mode. Running the test without Search measures training-data recall, not live web retrieval. Both are worth tracking, but they answer different diagnostic questions. See how to optimize for Google AI Overviews for the access checks specific to that platform.
How to build your prompt set
The prompt set is the core of the measurement system. A single branded query tells you almost nothing about real buyer-facing visibility. You need a structured set that reflects how buyers actually use AI search.
Target 20 to 50 prompts for a reliable baseline. Below 20, response variance makes patterns unreliable. Above 50, manual tracking becomes unmanageable without software.
Build across five categories:
Prompt category breakdown
| Category | Example | What it reveals |
|---|---|---|
| Branded | "What does [brand] do and who is it for?" | Entity recognition and description accuracy |
| Category discovery | "Best [service] for [audience]" | Whether you appear when buyers don't know you yet |
| Problem-aware | "Why is my brand missing from AI answers?" | Whether educational content is citable at the problem-recognition stage |
| Comparison | "[Brand] vs [competitor]" | Whether third-party evidence supports your brand against a named alternative |
| Proof-seeking | "Which sources explain [topic] well?" | Whether AI can find authoritative evidence about your offer |
Non-branded prompts matter more than branded ones for conversion-stage visibility. If you only appear in branded queries, you are invisible to buyers who don't already know your name.
Phrase prompts to push AI into recommendation mode. "Best [category] for [audience]" returns different results than "What is [category]?" The second pulls definitions. The first pulls vendor recommendations. Test both, and track them separately.
For guidance on running branded and category prompts inside ChatGPT specifically, see how to check if your brand is cited in ChatGPT.
How to run and record your results
A spreadsheet is sufficient to start. One row per prompt per run.
| Column | What to record |
|---|---|
| Prompt | Exact wording: do not paraphrase across runs |
| Platform | ChatGPT Search, Perplexity, Google AIO, Gemini |
| Mode | Search on, default, deep research |
| Brand mentioned | Yes, no, or partial |
| Brand cited | Your domain, third-party source about you, or no citation |
| Cited URLs | Every source URL shown in the answer |
| Competitors named | Which competitors appeared |
| Competitor URLs | Sources used for competitor mentions |
| Answer accuracy | Correct, incomplete, outdated, or wrong |
| Notes | Anything unusual: location context, mode variance, missing Sources panel |
AI responses are non-deterministic. The same prompt can produce different answers across sessions. To account for variance, run each prompt at least three times and record the majority result. If results split evenly, note "mixed" and continue.
Keep platform and mode consistent within a run. Mixing modes makes trends unreadable. If citation rate drops month-over-month, you want to know whether visibility changed, not whether you forgot to enable Search.
After your first full run, calculate three numbers:
- Mention rate: prompts where brand was mentioned divided by total prompts run
- Citation rate: prompts where your domain was cited divided by total prompts run
- Accuracy rate: prompts where the answer described your brand correctly divided by total branded and category prompts
Those three numbers are your baseline.
How to score and interpret your data
Benchmarks make the numbers actionable.
Search Engine Land and Averi AI both note that for B2B brands, a citation rate below 10% across tracked prompts signals effective invisibility to AI-assisted buyers. A rate of 20 to 30% is a reasonable operational target. Above 40% is category-leading.
| Citation rate | Diagnosis |
|---|---|
| Below 10% | Invisible: significant access, understanding, or authority gaps |
| 10 to 20% | Emerging visibility: gaps likely in category and problem-aware prompts |
| 20 to 30% | Solid foundation: target specific prompt categories where gaps appear |
| Above 40% | Category-leading: maintain freshness and monitor for accuracy drift |
Two data points add useful context. An Ahrefs analysis of 1.4 million ChatGPT prompts found that 88.46% of ChatGPT's cited pages come from Google's index — so Google crawlability is foundational, but being indexed is necessary, not sufficient. Separately, niche brands are currently outperforming larger brands in AI citation rates. Content bloat and vague positioning hurt more than low domain authority. A focused, clearly positioned 20-page site can outperform a sprawling enterprise content library in AI answers.
After scoring, identify the pattern behind the gaps. Each pattern points to a specific root cause:
| Pattern in the data | Likely root cause | What to inspect first |
|---|---|---|
| Site never cited when Search is active | Access | Crawler permissions, robots.txt, WAF, indexability |
| Brand mentioned but described incorrectly | Understanding | Homepage, schema, entity definitions, external profiles |
| Competitors cited from sources you're not in | Authority | Reviews, directories, comparison pages, media |
| Visible in branded prompts only | Category association | Service page language, use-case pages, category content |
This diagnostic frame (access, understanding, authority) is the same one used in an AI Visibility Audit. When you can identify the pattern from your tracking data, you know which layer to address first. If you rank in Google but are absent from AI answers, the gap is almost always authority or access, not content volume.
How often to check
Frase recommends weekly or bi-weekly testing for active campaigns. For most teams without dedicated resources, monthly is practical and sufficient to detect meaningful movement.
AI answers are not stable week-to-week. Citation status can shift significantly between refreshes. That is an argument for regular, consistent AI search visibility monitoring, not a reason to avoid tracking. The goal is to detect patterns across runs, not to treat any single result as reliable.
A practical cadence for how to track AI search visibility over time:
- Month 1: Run the full prompt set weekly. Four runs across a month give you enough variance data to establish a real baseline.
- Month 2 onward: Monthly runs. Record the date, platform, and mode. Compare month-over-month, not run-to-run.
- After content changes: Run the relevant prompt categories within 10 days of publishing or updating a page. AI search retrieves fresh content faster than most teams expect.
Thirty data points, three runs across ten prompts, is enough to see whether a pattern is real. One run is never enough to act on. Amplitude's comparison of AI visibility monitoring tools notes that automated platforms add anomaly detection and behavior analysis once you have a baseline, but the baseline itself starts with manual measurement.
When manual tracking points to a professional audit
Manual tracking answers what is happening. It is harder to determine why from a spreadsheet alone.
Some patterns are clear enough to act on directly. If your site is never cited when Search is active, start with the crawler and access layer. If branded prompts return accurate answers but category prompts return nothing, the gap is usually category association — your service page language or use-case content is not connecting to how buyers phrase the problem. If one competitor appears consistently across every prompt category, pull their cited URLs and find which sources you're absent from.
Other patterns need more investigation. A low citation rate despite clean access usually points to the source ecosystem outside your site — reviews, directories, comparison pages, or community discussions that AI systems use to corroborate a brand. A brand that is mentioned incorrectly even after you updated the homepage is often pulling from third-party sources carrying stale information; fixing only your own pages won't resolve it. Inconsistent results across platforms with no directional pattern almost always require systematic source analysis to untangle.
An AI Visibility Audit builds on the manual baseline. It runs a larger validated prompt set, checks technical access for priority URLs, maps the off-site source ecosystem, and produces a prioritized fix roadmap. The methodology and sample audit show exactly what that looks like before you commit.
If you have the tracking data and the pattern is unclear, that is the right moment for a structured audit. Not as a replacement for measurement, but as a diagnostic extension of it.
FAQ
Can I track AI visibility without a paid tool?
Yes. Build a spreadsheet with the columns described above, define a prompt set of 20 to 50 queries across the five categories, run each across ChatGPT Search, Perplexity, and Google AI Overviews, and record mentions, citations, competitors, and accuracy. Manual tracking is reliable enough to establish a baseline and identify where gaps sit.
What is the difference between an AI mention and an AI citation?
A mention means the AI named your brand in the answer text. A citation means the AI linked to a URL on your domain as a source. A brand can be mentioned without a citation. A page can be cited without the answer recommending your brand. Track both separately because they point to different problems.
How often should I check AI search visibility?
Weekly for active campaigns, monthly for ongoing monitoring. Run each prompt at least three times per session to account for response variance. Thirty data points, three runs across ten prompts, are enough to detect a meaningful pattern. One run per month is not sufficient.
Which AI platforms should I monitor first?
Start with ChatGPT (Search mode active), Perplexity, and Google AI Overviews. These three account for the largest share of AI-driven category and purchase-intent queries. Add Gemini and Microsoft Copilot once your primary baseline is established.
How many prompts do I need for a reliable baseline?
20 to 50 prompts spread across branded, category discovery, problem-aware, comparison, and proof-seeking categories. Below 20, response variance makes the data unreliable. Above 50, manual tracking becomes impractical without software.
How do I know if my AI visibility is improving?
Compare mention rate, citation rate, and accuracy rate month-over-month across the same prompt set and platforms. Improvement is a consistent directional shift across multiple runs, not a better result on one prompt in one session. Track the platform-specific breakdown, since gains on one platform can mask losses on another.
Measurement comes before optimization. Define the four metrics, build a prompt set that reflects real buyer behavior, record consistently, and score against the benchmarks. The pattern in the data tells you which layer, access, understanding, or authority, needs attention first.
Ready to go beyond the spreadsheet?
The AI Visibility Audit runs a validated prompt set across ChatGPT, Perplexity, and Google AI, checks technical access for priority URLs, maps your source ecosystem, and delivers a prioritized roadmap.