AI Visibility Metrics: Mentions, Citations and Enquiries

Define four different events

A business can appear in an AI answer without receiving a click. A visitor can arrive without becoming an enquiry. An enquiry can be unsuitable for the service. Treating those events as one success metric makes it difficult to decide what to improve.

Define a qualified enquiry before reviewing results. For a fictional maintenance company, the criteria might include a service it offers, an address in its operating area, and a genuine request for work. Use criteria relevant to your business and apply them consistently.

Keep the observation conditions

For AI answer checks, record the exact prompt, platform, visible model or mode when available, date, location context, and whether the conversation was fresh. Save the answer and cited destinations. Without these details, a later difference may be impossible to interpret.

Separate discovery prompts from brand-specific prompts. “Who can help with this service?” tests a different situation from “What does this named company offer?” Both can be useful, but combining them into one visibility percentage obscures what was tested.

Use a stable prompt sample for comparison and keep exploratory prompts in a separate set. If you add new prompts, record the change. A larger or easier sample can move a percentage without any change to the website.

Understand referral and attribution limits

Profound's analysis of ChatGPT referral links describes how a platform's link presentation can affect observed referral behavior. This is a reason to annotate platform changes in a report. A change in recorded traffic is not automatically evidence that your latest content edit caused it.

Analytics also sees only part of the journey. Someone may read an answer, search the business name later, and contact it from another device. A referrer can be absent or uninformative. Voluntary feedback such as “How did you hear about us?” can add context, but it is still imperfect recall.

Use analytics appropriate to your privacy practices and publish an accurate explanation of it. Do not add tracking simply to make a dashboard look complete. A small business needs a measurement process it can maintain and understand.

Read a small example without overstating it

The following numbers are fictional and illustrate interpretation, not a FoundUp result. Suppose a business reviews twelve discovery prompts in one round. Its name appears in four answers, its website is cited in two, analytics records eighteen identifiable AI referrals that month, and three enquiries mention an AI tool.

Those observations do not form a single trackable funnel. The twelve prompts are a small test set, not the population of customer questions. The eighteen visits may come from questions outside that set. The three enquiries may overlap with those visits or may have arrived through another route.

A useful report would state each observation separately and note what is unknown. It might add that one answer used an old service area, which creates an actionable correction. “AI generated three customers” would go beyond the evidence if the records only show enquiries.

Add first-party reports where you have access

As of September 28, 2026, Google's Generative AI performance report reports impressions for supported Search AI features. Review the available page, country, device and date dimensions. A missing report or insufficient data is not a measured zero. These impressions are not ChatGPT citations, clicks or qualified enquiries.

Also review the property's Search generative AI control. Technical crawl access alone does not establish the property's inclusion setting. Keep its verified status with your indexing evidence rather than assuming a public robots.txt file answers every eligibility question.

Bing's AI Performance reporting provides another platform-specific view of citations and grounding queries. Preserve each platform's definitions and coverage. Do not combine their metrics into a universal GEO score or claim a customer conversion from an impression.

Use a record that preserves missing data

Copy these fields into a worksheet for each answer observation. Keep the full answer in a separate record and use its reference in the final field; do not put private customer information in the worksheet.

observed_at,prompt_id,prompt_text,platform,model,market,language,
prompt_type,brand_mentioned,owned_source_annotation,owned_link_in_text,
fact_check,referral_visits,qualified_enquiries,data_status,answer_record

Our branded FoundUp API observation had a text link but no structured source annotations and an incorrect GEO expansion. Reporting only “brand visible” would hide the actual problem. Reporting zero referrals would also be wrong: that API call did not measure website visits.

For exports from platform dashboards, keep the platform's definitions and any missing-data flags alongside the numbers. Google's Generative AI performance documentation warns that unavailable values displayed as ~ or - become zero in a downloaded report. Save the visible status and reporting period with the export before treating a zero as a measured absence. This is a documentation-based precaution; it is not a claim that FoundUp's Search Console report has been inspected.

Run a short monthly review

1. Check business outcomes. Count enquiries using the same definitions and inspect whether poor fit has a repeated cause.
2. Review accuracy. List incorrect service, location, pricing, or contact claims found in the prompt sample.
3. Inspect sources. Save the URLs used for important answers and note whether they are owned pages or external references.
4. Annotate changes. Record website releases, campaigns, seasonal changes, prompt edits, and known platform changes.
5. Choose a small next action. Repair an incorrect fact, clarify a service boundary, or investigate a broken page before expanding the report.

For low enquiry volumes, avoid reacting to every monthly percentage swing. Read the actual requests and keep uncertainty visible. A useful measure should improve a decision, not supply a confident-looking narrative for a small sample.

Use the prompt testing guide to design the observation set and the correction workflow when an answer contains a factual error.

An observation record you can inspect

These are observations from FoundUp’s own API tests, not customer results or records of real user searches. The table shows what was asked, what the response contained and what the evidence does not establish.

The full question set and method preserve the original sample. A monitoring proposal should specify the platforms or models, questions, markets, languages, frequency, reporting period and who acts on the findings. This example does not establish a standard FoundUp monitoring schedule.

Download a record you can reuse

Download the blank observation template (CSV) or inspect the completed FoundUp example (CSV). The example preserves the September 28 branded diagnostic above, including its incorrect expansion of GEO. It is a historical observation, not a result for the latest website release.

Keep one row per answer. Record the exact question, test time, platform and model, language, location setting, brand mention, structured source annotations, text links and limitations. A question written by your team is an evaluation prompt, not evidence of search demand. Leave unavailable fields blank rather than recording a measured zero.

In this example, the answer names FoundUp and includes a website text link, but returns no structured source annotation. These are separate fields in the download. The source URL points to this explanatory record. Retain full answers privately when needed, with customer information and credentials excluded from any public export.