Brand mentions
Record whether the business is named in a comparable completed response, along with the question, context and company-name variation. Keep branded control queries separate from unbranded discovery questions.
AI responses change with questions, platforms, time and context. Our monitoring process focuses on consistent research panels, manually reviewed sources and readable findings that support business decisions.
AI visibility monitoring is the recurring observation of how a brand, its competitors and cited sources appear across a defined sample of AI-assisted answers. A monitoring engagement repeats important buyer questions under documented conditions, separates brand mentions from provider recommendations, inspects referenced URLs and explains changes with appropriate uncertainty.
It is different from a one-time manual AI visibility audit. The audit establishes scope, questions, sources and initial findings; monitoring revisits a comparable panel to identify what changed, what stayed the same, and which differences deserve a closer investigation. It is also different from a software dashboard that publishes a single “AI rank” without explaining the underlying prompt sample or sources.
Businesses often want to know how to monitor AI search visibility in ChatGPT, Gemini, Claude, Perplexity and Google's AI experiences. The practical answer is to define a relevant question panel, repeat what is feasible in each selected interface, capture dated responses, and report observations separately by platform and intent. Availability, retrieval behavior and interface features vary; a missing test should remain missing, not become a zero mention.
Our measurement guide explains the underlying metrics for teams that want to perform their own research. This service page explains the scope and deliverables of an ongoing, human-reviewed monitoring engagement.
An answer mentioning a company is not equivalent to a buyer recommendation. We keep different observations separate so decision-makers can judge where attention is needed.
Record whether the business is named in a comparable completed response, along with the question, context and company-name variation. Keep branded control queries separate from unbranded discovery questions.
Look for a provider actually suggested or shortlisted in answer to a buyer's needs, not an incidental mention in a general explanation. Capture shortlist position where visible and the stated selection context.
Save cited URLs and identify whether they are first-party site pages, social profiles, independent third-party content or competitors. Inspect relevance rather than treating each citation as an endorsement.
Record which other providers appear for the same buyer needs. Compare category and service positioning without assuming every competitor mentioned is a verified better choice.
Check whether a response accurately describes the brand's real services, eligibility, geography and limitations; flag contradictions or outdated third-party facts for follow-up.
Review observable referrals, form completions and buyer-reported discovery routes where tracking is available. Do not treat an AI answer or a crawler request as an actual lead.
Questions that require a detailed technical diagnostic, source comparison or remediation plan should be scoped as a separate manual audit or implementation project.
We begin with the business's real buyer journey, not hundreds of arbitrary prompts designed to inflate activity. A useful panel has documented scope and stable wording, while new questions are clearly labeled as additions rather than silently pooled with an earlier baseline.
| Question group | Example question | Reason to track |
|---|---|---|
| Unbranded discovery | Which providers offer [verified service] for [buyer type]? | Observe whether a brand is offered as a possible solution without being named in the prompt. |
| Commercial comparison | How should I compare agencies that offer [service]? | Examine shortlist criteria, competitors, sources and decision factors. |
| Problem-led selection | Who can help with [specific urgent business problem]? | Test if the brand appears for the buying problem rather than a broad category. |
| Information and definitions | What does [specific service] include? | Measure reference/citation context separately from commercial recommendations. |
| Branded control | What services does [actual company name] provide? | Test entity accuracy; exclude from unbranded recommendation-share calculations. |
Include geography, language and interface mode only when relevant to the real target market. Three repeated runs per prompt can highlight volatility, but they are not a representative sample of all users or all model responses. Exact wording and sample counts must accompany any percentage.
The how-to-check guide explains a self-service test protocol; the MentionGauge methodology explains how source evidence and uncertainty are handled.
Repeat tests are useful only when the collection method and reported limitations remain visible. Automation can assist collection, but a manual reviewer still needs to interpret sources and business context.
Agree on services, regions, competitors, question groups, selected platforms, test cadence and responsibilities. Retain original prompt wording and label any additions or removals over time.
For each actual run, capture platform, mode where available, location/language, question text, time, response and source links. Mark platform downtime and missing access rather than silently replacing results.
Separate mentions, provider recommendations, website citations, third-party references, competitor names and inaccurate claims. Save enough context for a reviewer to verify the classification.
Open key citations, compare them with previous periods and the business's updated service information, and flag sources that are inaccessible, unsupported or inconsistent. Do not infer hidden model queries from a public answer alone.
Compare the same prompt families, platform/mode and repeat counts where possible. Segment branded and unbranded results, and explain any material change in sampling conditions or model availability.
Report observations, notable competitor/source shifts, known site changes, uncertainty and recommended investigations. Keep recommended content or technical fixes distinct from confirmed monitoring results.
A well-defined baseline makes future comparisons more meaningful. Our manual audit service can define an appropriate initial question panel and evidence standard.
Hypothetical example, not a client result: consider a ten-question panel with two branded identity checks and eight unbranded questions. If each unbranded question is run three times on one selected platform, the maximum comparable unbranded denominator for that platform is 24, provided every run is actually completed.
| Metric | Example record | Why it matters |
|---|---|---|
| Sample coverage | 8 unbranded questions × 3 completed repetitions = 24 responses. | Prevents an omitted or failed test from being mistaken for a brand absence. |
| Brand mention rate | 6 of 24 comparable responses mention the business (25% of this illustrative sample). | Describes a defined panel, not share of all AI user queries. |
| Recommendation rate | 2 of 24 explicitly suggest the business as a provider (about 8.3%). | Distinguishes shortlist inclusion from incidental naming. |
| Source mix | Keep own-domain, social, independent and competitor URLs in different fields. | Shows what source type was cited without treating a link as an endorsement. |
| Action | Investigate whether provider qualifications are clearly explained on service pages. | A hypothesis to test, not a proven reason for the model's selection. |
The numbers in this table are deliberately invented to demonstrate calculations, not evidence about MentionGauge or any client. A real report must use actual captured responses, disclose denominators and retain the underlying observation log.
Use calculations that are meaningful to the buyer question and label the population actually tested. Never merge unlike platforms or branded and unbranded questions without showing the segments.
Number of comparable completed responses containing the brand divided by all comparable completed responses in the selected panel. Record spelling variants, context and intent before counting.
Responses explicitly recommending the brand as an option divided by comparable completed commercial-selection responses. A category overview that mentions a name is not automatically a recommendation.
Count actual cited URLs by domain, source type, intent and platform. State whether the denominator is cited links, cited responses or distinct source pages; each answers a different question.
Track attributable enquiries where available and ask what problem a customer was trying to solve. A response with no website click may still influence a buying decision, but the underlying attribution is often incomplete.
For worked definitions, formulas and limits, use the distinct educational resource How to Measure AI Visibility. This commercial page focuses on our monitoring scope rather than replacing that guide.
Answers may differ by prompt wording, available sources, interface features, region, date, underlying system changes and model variability. A change in sampled recommendation frequency does not automatically mean the website gained or lost a stable “rank.”
Keep a stable core set of questions, note when platforms change behavior, and record failures or missing citations. Label new keyword clusters as new measurements rather than retroactively adjusting the historical denominator.
If a site's content changed before a recommendation appeared, the sequence is interesting but not proof of cause. The answer system may have changed independently, so controlled comparisons and repeated observations matter.
Google says its AI features build on ordinary Search eligibility; no special schema is required. See Google's AI search documentation rather than presuming an undocumented ranking shortcut.
Only test platforms and modes that are actually accessible under acceptable terms and permissions. Report incomplete runs honestly. A screenshot from an ordinary search page is not equivalent to a logged answer from a different AI product.
| Service | Primary question | Decision output |
|---|---|---|
| Manual audit | What do the initial buyer-question answers and sources show? | Diagnosis, issue list and baseline proposals. |
| AI visibility monitoring | What changes over repeated, comparable sampling periods? | Dated observation log, citation/competitor changes and recommended follow-up. |
| AI SEO implementation | Which verified technical/content problems should we fix? | Approved page and site changes with QA. |
| AEO | Are customer questions answered clearly and completely? | Answer coverage and useful page revisions. |
| GEO | Which source and competitor patterns appear in generative answers? | Source investigation and evidence-led optimization direction. |
A suitable engagement should define the number and type of questions, platforms, repeat runs, markets, competitors, reporting interval and who reviews sources. No standard monthly price or response volume is claimed on this page. Scope and cadence must be agreed before any commitment.
A versioned prompt panel, actual response captures, source link log, brand and competitor classification, changes since the prior period, significant inaccuracies, limitations and a practical investigation queue. Deeper technical repairs or content creation should be identified separately.
How many buyer journeys, platforms, language/market segments, repeat tests and competitors are tracked; the time needed to review source pages; reporting depth; data permissions; and whether follow-up technical or editorial work is part of the contract.
Tell us your priority service, target market, expected question types and which competitors matter. We can discuss an audit-first scope and whether periodic manual observation is appropriate.
Choose important buyer questions, record exact wording and platform settings, repeat actual test runs and keep complete answers and citation URLs. Separate branded accuracy checks from unbranded discovery and hiring questions; review comparable observations on a defined cadence.
These interfaces can be included in a scoped research plan where the relevant modes are accessible and permitted. We do not claim an always-on universal API or guaranteed access to every model, region, version or conversation.
An audit establishes initial questions, source evidence and prioritized findings. Monitoring repeats comparable observations and reports changes over time. A technical SEO implementation is a different service that makes approved fixes.
Yes, competitor names and shortlist context can be recorded in a defined panel. They should be interpreted together with the user's intent and cited evidence; an incidental mention is not a verified endorsement.
We distinguish cited links, brand appearances, provider recommendations and where possible qualified enquiries. Not all interfaces show identical citations, so unavailable data remains explicitly unknown.
There is no universal number. Start with enough distinct questions to represent the main buyer journeys, keep a stable core, label intent and repeat count, then expand only when genuinely new information needs justify it. Small samples are directional, not population-wide market share.
Cadence depends on the commercial decision, available platform access and sampling workload. Monthly reviews may be useful, but the service schedule is not fixed here and a more frequent report does not automatically create better evidence.
No. Monitoring reports what sampled answers do and what related sources show. Technical and content changes may be recommended separately, but no service can guarantee that a particular AI platform will cite or recommend a provider.
Request a scoped manual audit to understand how your business appears across relevant AI-assisted search experiences and what can be improved.