Measuring AI Search Visibility: Citations, Visits and Leads

Measure AI search through observations you can explain: crawler access, platform-reported citations, referral visits and useful customer actions. Each reveals a different part of the journey.

Results vary by product, question, location and time. A useful report states what was measured and how the sample was collected, so you can compare changes without treating a limited observation as a market-wide ranking.

This guide shows how to combine those measures, document gaps and connect visibility with lead quality and sales outcomes.

The Mithril measurement model

We divide AI-search measurement into four layers:

1. Platform-owned visibility data
2. Site-owned traffic and conversion data
3. Machine-access data
4. Synthetic observation

Each answers a different question.

Do not merge them casually.

Platform-owned data: Google

Google provides a dedicated Generative AI performance report in Search Console. Use its current definitions when reviewing your site’s visibility.

The reporting can show visibility from generative AI features in Search and Discover, including metrics such as:

• impressions,
• pages,
• countries,
• devices where available,
• and trends over time.

This is important because it gives publishers direct first-party visibility from Google instead of forcing third-party tools to guess. Use it. Also remember what it is measuring. An impression is not a citation count from another platform.

It is not a conversion. It is not a universal AI-search rank. Keep the metric attached to the product that created it.

Platform-owned data: Bing

Bing’s AI Performance report is currently one of the most useful windows into grounding behavior.

It can expose:

• total citations,
• cited pages,
• grounding queries,
• citation trends,
• and expanded dimensions such as topics, intents, and citation share.

Bing explicitly warns that citation counts do not indicate ranking, authority, or importance. That sentence should be printed directly above any dashboard built from the data. A citation means the content was visibly referenced.

It does not mean Page A “ranks #1 in AI.”

Grounding queries are especially valuable

Bing describes grounding queries as grouped phrases associated with content that was retrieved and cited. They are not necessarily the exact user prompt. That distinction matters.

Grounding queries can tell you:

• which themes your pages support,
• which URLs are associated with those themes,
• where citation activity is concentrated,
• and where expected coverage is absent.

This is closer to observing retrieval behavior than traditional keyword ranking. Do not over-interpret it. Do use it.

Site-owned data: AI referrals

When people click through from AI products, your analytics can capture some of that traffic.

OpenAI says ChatGPT search referral URLs include:

utm_source=chatgpt.com

That creates a relatively clean signal for those visits. Other AI products may expose referral information differently. Build a reporting channel that identifies known AI referrals separately.

Track:

• sessions,
• landing pages,
• engaged sessions,
• calls,
• form submissions,
• conversions,
• revenue where available,
• and assisted outcomes.

A citation that never produces a click can still have brand value. A referral that produces a lead is easier to value. Measure both.

AI traffic will never be perfectly attributed

Some users will:

• see an answer,
• remember the brand,
• search it later,
• type the URL directly,
• or call from another device.

That means AI influence can appear as:

• branded search,
• direct traffic,
• organic traffic,
• or offline conversion.

This is not unique to AI. Marketing attribution has been annoying for decades. Do not demand impossible certainty from one channel while happily accepting view-through attribution elsewhere.

Machine-access data

Crawler data answers a different question:

Are AI-related systems accessing the site?

Sources can include:

• server logs,
• Cloudflare AI Crawl Control,
• CDN logs,
• security logs,
• and crawler dashboards.

Useful metrics:

• verified crawler requests,
• top URLs,
• status codes,
• blocked requests,
• robots compliance,
• and trends.

Crawler activity does not equal visibility. It is a prerequisite and operational signal. Keep it separate.

Synthetic prompt tracking

Third-party AI visibility tools can be useful.

They can repeatedly ask selected prompts and record:

• whether the brand appears,
• which sources appear,
• how responses change,
• and how competitors compare.

This can reveal patterns. It is not ground truth.

Why?

Prompts vary. Models vary. Location can vary. Personalization can vary. Answers can vary between runs. Retrieval may vary. Third-party tools do not have secret access to Google’s or OpenAI’s internal ranking systems.

Google itself now warns site owners to be cautious with third-party tools claiming internal AI metrics or guaranteed ranking insight. Use synthetic tracking as directional research. Not a census.

Measure the pipeline, not just the citation

Remember our model:

Different tools observe different stages.

This is why one tool cannot give you the entire story. It is also why a measurement strategy can diagnose where the problem lives.

Example: Lots of crawling, no citations

Possible questions:

• Is the content indexed?
• Does the page answer useful intents?
• Is it commodity content?
• Are important facts explicit?
• Are there stronger competing sources?
• Is the crawler training-related rather than search-related?
• Do Bing grounding queries show adjacent topics but not the target?

Do not respond by increasing crawl frequency. The access layer may already be fine.

Example: Citations, no traffic

Possible explanations:

• The answer satisfies the user without a click.
• The citation is visually secondary.
• The topic is informational and low commercial intent.
• The platform sends limited referral traffic.
• The cited page lacks a compelling reason to visit.

This may still produce awareness. But if business value is the goal, investigate the outcome layer.

Example: AI referrals convert extremely well

Now we have something interesting.

Analyze:

• Which landing pages?
• Which topics?
• Which service lines?
• Which AI sources?
• Which queries or citations can be associated?
• Can those canonical resources be strengthened?

This is where AI-search work becomes normal marketing again. Very refreshing.

Branded search can be a supporting signal

If AI systems repeatedly expose a brand, users may search for the brand afterward. That is difficult to attribute cleanly.

But you can monitor:

• branded query volume,
• direct traffic,
• assisted conversions,
• and customer-reported source.

Do not claim causation from a correlation. Use it as context.

Ask people how they found you

Radical analytics innovation:

Ask. Lead forms can include optional discovery-source fields. Intake teams can ask naturally.

CRM notes can record:

• ChatGPT
• Copilot
• Google AI Overview
• Gemini
• Traditional Google
• Referral
• Other

Self-reported attribution is imperfect. So is everything else. It can still reveal channels that analytics miss.

The KPI hierarchy

For a service business, we would prioritize:

The bottom layers help diagnose. The top layers pay the bills.