Your brand appears in an AI answer. That is an observation worth recording, but it is not a website visit or a qualified lead. Useful AI visibility tracking keeps those events separate and shows how each was measured. Otherwise, a report can turn a small prompt experiment into a business-impact claim it cannot support.
TLDR: Use a consistent prompt panel and record the system, date, context and result for each run. Track brand mentions, linked citations, website referrals and CRM outcomes as separate measures. Preserve missing results and do not infer a buyer journey without evidence connecting its steps.
Sources checked September 7, 2026. The prompt panel below is fictional. Its counts illustrate a measurement method and are not Selworthy visibility, traffic or lead results.
Define brand visibility metrics before collecting AI answers
We recommend the following working definitions. Adapt them to your measurement tools and keep the definitions with the report.
| Measure | What you record | What it does not prove |
|---|---|---|
| Brand mention | The observed answer names the brand | A link, visit or recommendation |
| Linked citation | The observed answer links to a specific URL | That someone clicked it |
| Referral visit | Analytics records a visit with the relevant observable source | Every visit influenced by AI |
| Form or contact event | The defined event appears in the measured system | A qualified sales opportunity |
| Qualified outcome | The CRM owner applies the agreed qualification rule | That one AI answer caused the outcome |
Keep a recommendation separate from a neutral mention if that distinction matters to your business. Record the exact context and evidence rather than letting an automated sentiment label make the final judgment.
Our HubSpot KPI gap-analysis guide offers related measurement context. This panel is a repeatable observation process, not a substitute for your full reporting model.
Build a fixed prompt panel
Select questions your intended buyers might ask and classify their intent. Include the exact wording, market, language and relevant context. Avoid quietly changing a prompt when it produces an inconvenient result.
We recommend separating branded, category, comparison and problem-solving questions. A branded question that names your company in the prompt should not be combined with an unbranded discovery question without a visible distinction.
For each run, preserve the system and version label available, timestamp, settings you can observe, answer evidence and linked URLs. If the interface does not expose a model version, record “not exposed” rather than inventing one.
Keep missing runs out of negative-result counts
This fictional panel contains four prompts evaluated on two hypothetical systems. The labels identify example runs, not real products or measured brand results.
| Prompt and system | Run state | Brand mentioned | Brand URL cited |
|---|---|---|---|
| P1 / System A | Completed | Yes | Yes |
| P1 / System B | Completed | Yes | No |
| P2 / System A | Completed | No | No |
| P2 / System B | Completed | Yes | Yes |
| P3 / System A | Completed | Yes | No |
| P3 / System B | Not collected | Unknown | Unknown |
| P4 / System A | Completed | No | No |
| P4 / System B | Completed | Yes | Yes |
There are eight planned runs and seven completed observations. Five completed answers mention the fictional brand and three cite its URL. The uncollected run is missing data, not evidence of absence. These counts describe only this example panel.
If you calculate a rate, disclose the completed-run denominator and the panel composition. Do not call it market share or overall AI visibility. The questions you chose and the systems you tested define the scope of the result.
Connect website evidence without overstating attribution
Review the actual traffic-source information your analytics tool exposes. Define the domains or classifications included in the AI referral view, preserve the date range and document any filters. Keep an unknown source as unknown.
We recommend checking the landing page, observed event and CRM outcome separately. Where your approved measurement setup connects them, retain the evidence for that connection. Where it does not, avoid presenting an inferred path as an observed conversion journey.
Google's current generative AI guide directs site owners to a Generative AI performance report in Search Console. Check which reports are available for your property and verify their definitions before using them. This article does not establish the dimensions, limits or data available in your account. Google: generative AI measurement guidance.
Google's older AI-features guide also describes AI Overviews and AI Mode traffic within the overall Web performance aggregate. If you use that aggregate, do not label it AI-only. Keep a search-performance report distinct from website referral analytics. Google: AI features and Web measurement.
Use those search reports alongside analytics and CRM evidence, with each source's definitions intact. Do not add their counts together as if they measured the same event.
Turn AI search visibility gaps into a review queue
For a useful weekly review, we recommend recording the question, observed answer, cited destination, issue and responsible owner. A missing citation might prompt an evidence review, but it does not establish that a particular technical fix will change the next answer.
Group work by the problem you can verify: an outdated fact, missing explanation, broken destination, unclear source or untested page access. Keep source-quality evaluation and content recommendations in their own review fields rather than hiding them inside one visibility score.
Our Loop Marketing overview offers broader planning context. The measurement log should still distinguish a hypothesis from a completed change and an observed result.
Make comparisons reproducible
When you compare periods, report whether the prompt panel, systems or collection method changed. Preserve the prior version when adding prompts. A larger panel can alter a percentage even if the answers to the original questions remain the same.
We recommend reviewing repeated observations before drawing conclusions from a single answer. Record the limits of the method and the decisions the data is sufficient to support. More decimal places do not make a small, selective panel representative.
Choose AI visibility tracking tools for the evidence you need
Start with the measurement question, then assess whether the tool exposes the observations needed to answer it. A polished visibility score is less useful when the team cannot inspect the underlying prompts and responses.
For each candidate, check platform coverage, market and language settings, prompt history, cited URLs, export options and the treatment of failed runs. Confirm which capabilities belong to the actual plan under consideration.
Do not treat an AI visibility checker, a full monitoring product and website analytics as interchangeable. They may collect different evidence on different schedules.
Inspect the underlying AI-generated answers
The report should let you review what the system actually returned. A brand name in a source list differs from a recommendation in the answer body.
Retain the relevant text and linked citation. If the tool only provides a summary, document that limit. Do not fill missing answer context with an assumption about why the brand appeared.
This is especially important when reviewing brand sentiment. A quoted criticism, a neutral comparison and the model's own recommendation should not silently become the same category.
Keep AI platforms separate in the report
Google AI Overviews, Google AI Mode and conversational assistants do not need to expose identical interfaces or reporting data. Use the actual surface and collection method as part of the observation.
For ChatGPT, Perplexity or another assistant, retain the visible mode and context the tool supplies. If a setting is not exposed, mark it unknown.
Avoid combining platforms into one trend without explaining the weighting. Adding a new platform can change a combined score even when nothing changed in the platforms previously tracked.
Preserve geography and language settings
If multi-country tracking matters to the business, keep the country, language and any location context with each run. Do not present observations from one market as evidence for another.
A local service question can produce different relevant answers depending on the place described. The prompt panel should reflect the actual audience without quietly changing those details between periods.
Use a collection log that survives changes in the tool
Retain the prompt identifier, exact wording, system, timestamp, result state, brand mention and cited URLs. Store enough context to compare the next observation without depending entirely on a dashboard's current layout.
Keep historical versions when prompts are edited. An improved question may be worth adopting, but it begins a changed measurement scope.
Classify missing data before calculating rates
A failed collection, unavailable model, incomplete response and completed answer without a mention are different states. Preserve the reason when possible.
Use the completed-run denominator for a rate based on completed observations, and disclose the missing runs separately. If missingness is concentrated in one platform or prompt type, flag that imbalance.
A clean percentage can hide poor coverage. The underlying collection log should make the gap visible.
Compare a stable panel before adding discovery prompts
A fixed panel supports period-to-period comparison. Exploratory prompts help discover new buyer questions. Both can be useful if their roles remain separate.
Do not add successful exploratory runs to the baseline panel after seeing their results and then call the increase an improvement. Keep the original panel intact and document any planned expansion.
Retain competitor research as context
AI competitor research can reveal which sources appear in the same observations. Open the cited pages and determine what they actually contribute.
A competitor citation can support a useful research question. It does not prove that copying the page's headings, length or phrasing will produce the same outcome.
Connect findings to a bounded content decision
Look for a reproducible issue: an outdated service claim, missing evidence, an inaccessible source or an unanswered buyer question. Assign a correction based on that issue.
Keep the task's completion evidence separate from the next visibility observation. The content may be corrected even if the selected AI response remains unchanged.
Report referral traffic with its collection limits
Use the observable source information in the approved analytics setup. Preserve unknown or unclassified traffic rather than force it into an AI category.
A visitor may be influenced by an AI answer without arriving through a recorded referral. That possibility is a limitation, not permission to estimate an unobserved visitor count.
Where a recorded visit connects to a CRM outcome, retain the evidence for the connection. Do not infer the full journey from a brand mention and an unrelated contact record.
Explain what the report can support
A prompt panel can support decisions about which answers to inspect and which content gaps deserve research. It does not automatically establish market share, revenue impact or a causal ranking improvement.
State the observed change in plain language, followed by its scope and limitations. That makes AI visibility data useful to business leaders without asking them to trust more than the method measured.
Choose AI visibility tools around the measurement question
An AI visibility tracker should help you reproduce an observation. Before choosing a platform, ask whether you can retain the exact question, model, date, location and source URL behind a reported result. A brand visibility score without those details is difficult to investigate.
Use AI search visibility metrics to identify questions worth reviewing, not to imply that every potential buyer received the same answer. Mentions, linked citations and referral visits remain different observations. Compare like-for-like runs and preserve unavailable observations instead of counting them as negative answers.
If a tool reports visibility gaps, inspect the underlying answers before adding pages to the backlog. The useful next action may be correcting an existing source or clarifying an offer, rather than creating another article.
Frequently asked questions
Does a brand mention count as traffic?
No. Record the mention as an answer observation. Website traffic requires separate analytics evidence, and a lead outcome requires its own definition and record.
Should a failed collection count as no mention?
No. Mark it as missing or uncollected and preserve the reason. Use a clearly stated denominator for any calculation.
Can we call the panel result our AI market share?
We recommend describing it as the result for the specific prompt panel and systems observed. Do not generalize it to the whole market without a method that supports that claim.
What should happen after a visibility change?
Review the underlying answers, citations, collection conditions and related website evidence. Document a hypothesis and next check instead of assigning causation immediately.
Measure what you can defend
Our consulting services can help you define the panel, measurement boundaries and review process.
Talk with Selworthy with the questions you want to track and the business decision the report needs to support.
