Skip to content
العربيةLet’s talk

Practical guide

Measuring AI search visibility: mentions, visits and qualified enquiries

Build an AI search measurement plan that separates sampled mentions, reported impressions, website visits, crawler requests and qualified enquiries.

Lava Solutions · · 5 minute read

Printed charts, glasses, pencils and a calculator arranged on a desk for analysis.
Illustrative stock photo. Photo: Cht Gsml / Unsplash.

Measuring AI search visibility requires several separate records. An answer mentioning your company, a link shown to a searcher, a visit to your website and a qualified enquiry each answer a different question. Combining them into one “AI traffic” total hides what changed and what still needs evidence.

Start with a small measurement plan that a marketing owner and a sales owner can both explain. The framework below uses provider documentation checked on 18 November 2026. It describes a proposed method, not an audit of a particular website or a report of customer results.

1. Define what each number means

Write the definition, source, period and owner beside every metric before collecting a baseline. Decide whether you count answers, sessions or individual enquiries. Keep the denominator visible when reporting a percentage.

SignalEvidence to retainWhat it does not prove
Sampled brand mentionSaved answer containing the identified brandThat the brand was recommended or that real buyers saw it
Sampled citationAnswer and exact linked pageThat anyone clicked the link
Reported impressionProvider report and its counting definitionA website visit or a unique person
Attributed website sessionAnalytics source, landing page and collection rulesAn identified buyer or a qualified opportunity
Qualified enquiryReceived record reviewed against agreed sales criteriaThat AI exposure alone caused the enquiry

A citation can exist without the company name being mentioned, and a mention can appear without a link. Record these independently. Also distinguish a positive recommendation from a neutral reference or criticism.

2. Make answer sampling repeatable

Choose a fixed set of genuine buyer questions for each language and market. Keep Arabic and English results separate; a translated prompt is not evidence that both markets have the same demand. Include problem-solving and comparison questions, not only prompts that explicitly name your company.

For each observation, save the exact prompt, provider, product mode, date, prompt language, answer language, stated market context and answer. Record location where known and signed-in or fresh-session conditions where relevant. Use the same procedure for the next round and log changes rather than silently replacing poorly performing questions.

Review whether the cited page supports the answer and whether the business description is accurate. Treat the sample as a controlled set of observations, not a census of all answers or a market-share estimate. If the sample changes, report the old and new sets separately.

3. Use the current provider reports

Google’s Generative AI performance report for Search now documents impressions from AI Overviews and AI Mode, grouped by pages, countries, devices and dates. It may be absent when a site lacks enough impressions. Its data is included in overall Web search performance, so do not add these impressions to that total. This impressions view is not an AI-only clicks, conversion or exact-prompt report.

Google Analytics’ current channel definitions include an AI Assistant channel for sources such as ChatGPT and Gemini. Google’s AI Overviews and AI Mode remain within Organic Search. Therefore, an Organic Search session alone does not identify an AI feature as its origin. Preserve source/medium detail alongside channel labels and record the reporting scope.

Use the actual available export, not an assumed dashboard. Note its timezone, filters and missing values before comparing it with another tool. If a report is unavailable, record that limitation instead of entering a zero or estimating a precise split.

4. Connect visits to enquiries carefully

Define a successful enquiry as a received submission, not simply a click on the contact button. Agree how the analytics event corresponds to the receiving form or CRM record. Test with clearly marked synthetic data and remove test records from business reporting.

Ask sales to define qualification in advance: for example, a relevant service need, a serviceable market and a usable contact route. Record duplicates, spam and enquiries outside scope separately. For workflow ownership after receipt, use our enquiry-to-CRM guide.

Document gaps in consent-dependent collection, missing referral information and cross-device journeys. A prospect’s answer to “How did you hear about us?” is useful supporting evidence, but should remain distinct from measured session attribution. Keep personal information out of analytics URLs and shared reporting sheets.

5. Keep crawler activity in a diagnostic log

OpenAI distinguishes its user agents: OAI-SearchBot supports search discovery, GPTBot relates to potential model training, and ChatGPT-User handles certain user-requested visits. Check relevant provider IP information when validating requests; a claimed user-agent name alone is insufficient identification.

Use logs to investigate access, errors and which pages were fetched. Neither a crawler request nor a user-triggered fetch establishes a human website session, an answer citation or a lead. Keep these counts outside the sales funnel.

6. Read a small report without inventing a funnel

Fictional example: a team saves 20 answers, finds four brand mentions and three linked citations, and separately records 12 AI-attributed website sessions. Two enquiries are received, of which one meets its qualification rules. These are invented numbers for illustrating the method.

The four mentions represent 20% of that answer sample only. The team cannot claim that the three citations produced the 12 sessions without connecting evidence, or attribute both enquiries to those sessions without a valid measurement link. Report unknown connections explicitly.

Review a consistent period, annotate content changes and investigate data quality before explaining movement. Google’s anomalies log, for example, records a logging issue that reduced reported Search generative AI impressions for 9–11 November 2026. A reporting drop is not automatically a loss of visibility.

Use our SEO proposal scope guide to define reporting deliverables. To discuss a measurement baseline through Lava’s SEO services, send your target markets, languages and the business action you want to measure.

Sources & further reading

A useful next step

Bring one real workflow.

Share the process you want to improve, the systems involved and the decision you need to make.

Talk to Lava ↗