How to Measure AI Search Visibility Without Inventing a Vanity Metric

Measure AI search visibility in five separate layers: technical eligibility, a controlled sample of answers, cited-source evidence, search and referral data, and commercial actions. Do not compress them into one “AI visibility score” unless the inputs, weighting and uncertainty remain visible.

AI answers are not a stable set of ten blue links. Wording, model, mode, location, account context and time can change the response. Measurement therefore needs repeatable sampling and honest limits, not false precision.

The goal is not to know every answer ever produced. It is to know whether relevant systems can discover the business, whether the brand and its sources appear more often in important answer journeys, and whether that visibility leads to relevant enquiries.

Checked: 14 July 2026 · Measurement methods reflect current Google and OpenAI reporting limits

The five-layer AI visibility framework

Report each layer. A technically eligible site can have low sampled visibility. A frequently mentioned brand can have inaccurate citations. A small number of high-intent referrals can be more valuable than a large mention count.

Start with a measurement question, not a tool

Useful questions include:

• Are we included when UK buyers ask for this category and narrow by our real strengths?
• Which sources influence answers about our products?
• Are answer engines confusing us with another company?
• Does our new comparison content earn citations?
• Are local answers using accurate location and service information?
• Do AI referrals reach commercial pages and take action?

“What is our AI visibility?” is too broad. The answer depends on the market, prompt set and system.

Layer 1: verify technical and factual eligibility

Before tracking mentions, check whether the source can participate.

Google AI features

Google's AI features guidance says a page must be indexed and eligible to show a search snippet to appear as a supporting link in AI Overviews or AI Mode. It states that no extra technical requirement, special schema or AI text file is necessary.

Check:

• Googlebot is allowed by robots and infrastructure;
• important pages are indexable and canonical;
• useful content exists in rendered text;
• internal links make it discoverable;
• structured data matches the visible page;
• Search Console shows the correct preferred URL;
• Merchant Center and Business Profile facts are current where relevant.

ChatGPT search

OpenAI's publisher guidance advises sites not to block OAI-SearchBot if they want content included in ChatGPT summaries and snippets. Check the live robots.txt, CDN rules and response status for the crawler.

Entity consistency

Record the official:

• organisation and brand names;
• services or products;
• website and relevant profiles;
• founders or experts;
• locations and areas served;
• contact facts;
• relationships between brands, people and offers.

Then find contradictions across the site and credible external profiles. Eligibility includes being understood as the correct business, not merely being crawlable.

Layer 2: build a controlled prompt set

Prompts should represent buyer journeys, not vanity searches for the company name.

Use six groups:

1. Category discovery: “Which companies provide [service] in the UK?”
2. Problem-led: “How should a [buyer] solve [specific problem]?”
3. Comparison: “[Option A] vs [Option B] for [situation].”
4. Suitability: “What is best for a business with [constraints]?”
5. Local: “Who provides [service] near [place]?”
6. Brand facts: “What does [brand] do?” and questions about specific evidence.

Build the set from Search Console queries, sales questions, site search, customer interviews and genuine comparison behaviour. Do not pad it with 500 synthetic variations to make the dashboard look complete.

Record enough context to repeat the sample

For every run, store:

• date and time;
• engine and product mode;
• logged-in or neutral environment where known;
• location or market setting;
• exact prompt and relevant follow-up;
• full answer or a compliant evidence extract;
• whether the brand appears;
• role: recommended, compared, cited, mentioned or excluded;
• cited URLs;
• factual accuracy;
• material competitor mentions.

Keep the core prompt wording stable between periods. Add an exploration set separately so new buyer language does not destroy comparability.

Layer 3: measure mentions and citations separately

A mention says the brand appeared. A citation shows which source supported the answer. They are not interchangeable.

Useful sampled measures include:

Brand inclusion rate

prompts containing an accurate brand mention ÷ eligible prompts tested

Segment by discovery, comparison, local and brand-fact groups. One overall percentage can hide complete absence at the decision stage.

Citation inclusion rate

prompts citing an owned or attributable brand source ÷ prompts tested

Also record citations to independent sources that accurately corroborate the business.

Source share

Count which domains and page types recur across citations. The useful finding may be that answer systems rely on trade bodies, customer communities, product documentation or comparison pages rather than agency blogs.

Accuracy rate

answers with no material factual error about the brand ÷ answers mentioning the brand

An inaccurate recommendation is not positive visibility. Record the exact error and likely source.

Role quality

Classify appearances:

• named recommendation;
• valid option in a comparison;
• source cited without brand recommendation;
• passing mention;
• confused with another company;
• present but unsuitable for the stated need.

Do not assign a secret weight and hide the categories. Leaders should see what kind of visibility changed.

Layer 4: use Search Console and analytics for observable discovery

Google Search Console

Check whether the property has Google's Generative AI performance report. It is available to a limited set of sites with sufficient impressions. Where it is absent, use ordinary Web reporting and state that you cannot isolate AI activity. Do not label every search change “AI traffic”.

Track:

• non-brand discovery queries;
• comparison and question queries;
• impressions and clicks to pages designed for answer journeys;
• branded query growth;
• page-level changes after content or technical work;
• countries and devices relevant to the offer.

Google recommends using Search Console for pre-click search performance and Analytics for on-site behaviour. The two systems use different metrics and will not match exactly; its Search Console and Analytics guide explains the roles.

AI referral traffic

In analytics, group identifiable referrals from ChatGPT, Perplexity, Copilot, Gemini and other relevant sources. Preserve the original source and landing page before rolling them into a channel group.

Measure:

• sessions or users, subject to consent;
• landing pages;
• engaged visits;
• service, product or pricing-page views;
• forms, calls, bookings and purchases;
• assisted conversions;
• query context when a customer volunteers it.

Referral totals are a floor, not a complete count. Apps, privacy controls, copied links and zero-click answers create missing attribution.

Layer 5: connect visibility to commercial action

Add one simple question to the sales or enquiry process:

Where did you first hear about us, and what did you check before getting in touch?

This catches journeys analytics misses: an AI answer introduces the company, a review confirms it, a later branded search receives the click.

Track by month:

• qualified enquiries mentioning an AI answer or search assistant;
• sales conversations containing the monitored problems or comparisons;
• landing pages used in those journeys;
• close rate and value where sample size permits;
• factual objections caused by inaccurate answers;
• branded search and direct visits around visible answer changes.

Do not force attribution where the customer cannot remember. “Unknown” is better than a fabricated channel.

Create a baseline report

A useful baseline contains:

1. scope: market, products, engines and date;
2. technical eligibility and crawler access;
3. entity facts and material conflicts;
4. prompt-set design and limitations;
5. raw answer observations;
6. brand inclusion, citation, accuracy and role by prompt group;
7. recurring sources and competitors;
8. Search Console and identifiable referral baseline;
9. commercial tracking gaps;
10. a prioritised implementation plan.

Save the raw evidence. A chart without the prompts and cited URLs cannot be audited later.

Reporting cadence

Monthly

• eligibility or crawler changes;
• fixed core-prompt sample;
• citation and source changes;
• Search Console page-query trends;
• referral landing pages and actions;
• work completed and next decisions.

Quarterly

• refresh the buyer-question research;
• review platforms and modes worth sampling;
• assess entity and external-source changes;
• compare commercial outcomes;
• retire prompts that no longer represent demand;
• add a separately labelled exploration set.

Daily rank-style monitoring can create noise without action. Use higher frequency only for material brand-risk or launch events.

Common measurement mistakes

• Searching only for the brand. That tests recognition, not category discovery.
• Changing prompts every run. The sample stops being comparable.
• Treating one answer as a ranking. Outputs vary.
• Counting mentions without accuracy. Confusion looks like success.
• Counting citations without inspecting the source. A cited page may be outdated or unrelated.
• Calling all Google changes AI traffic. Use the dedicated generative-AI report where available; ordinary Web totals do not isolate the cause of a change.
• Ignoring consent and dark referrals. Analytics undercounts some journeys.
• Inventing a magic score. Weighting hides the evidence leaders need to decide.
• Monitoring without implementation. Observation alone does not create visibility.

A transparent dashboard

Keep it small:

Put methodology and caveats beside the dashboard, not in a forgotten appendix.

Frequently asked questions

Can Google Search Console show AI Overview traffic?

Some properties have a dedicated Generative AI performance report. Check access and the report's metric definitions before using it. If your property lacks that report, use Web search trends and explain that they do not isolate AI Overview activity.

Can I track ChatGPT referrals?

Yes, identifiable referrals can appear in analytics, and OpenAI says publishers can track ChatGPT referral traffic. Consent, apps and copied links mean the visible total will not capture every journey.

What is share of model?

It is usually a vendor-defined estimate of how often a brand appears across a prompt sample. It can be useful if prompts, engines, frequency, weighting and raw answers are visible. It is not an industry-standard census.

How many prompts should I track?

Start with 25–50 high-value prompts across distinct buyer stages. A smaller curated set is easier to review for accuracy and commercial meaning than thousands of generated variants.

How often should AI visibility be measured?

Monthly is sufficient for many businesses, with quarterly prompt research. Measure more often for sensitive launches or reputation issues, not because the tool can run daily.

Keep the uncertainty visible

AI visibility can be measured well enough to make decisions. It cannot be reduced to a permanent rank. Use controlled samples, preserve raw evidence, connect the layers and let commercial action—not dashboard drama—set the priority.

Related: AI visibility audit deliverables · Run a DIY ChatGPT citation audit · From SEO to AEO · AEO pricing UK · What AI search cites · Google Search Console guide