Measuring AI referrals and visibility

0:00 Measuring AI referrals and visibility

Someone in a meeting says, we're number one in ChatGPT now, and holds up a screenshot. What do you say? In this lecture you'll learn how to measure AI visibility properly, with three layers of evidence: referral traffic, first-party AI reports from Google and Bing, and a repeatable prompt panel. By the end, you'll be able to replace screenshots with numbers that survive a sceptical finance director.

0:29 Three layers of measurement

Let's frame it with three layers. Layer one is referral traffic from assistants, which you can measure in analytics, with caveats. Layer two is visibility in answers: mentions and citations across a set of prompts, which you have to sample. Layer three is business impact: leads, sales and brand demand influenced by AI exposure, which you mostly infer. Each layer answers a different question. Traffic asks, did they visit? Visibility asks, were we in the answer? Impact asks, did it matter to the business?

1:06 Layer 1: AI referrals in GA4

Layer one first. Many assistants pass a referrer when someone clicks a citation, and ChatGPT adds utm source equals chatgpt dot com to its links. In GA4, create a custom channel called AI assistants, with a regex matching assistant domains, and put it above the referral channel, because GA4 evaluates channel rules in order. The exact regex is in the lesson text. Caveats matter. Not every click carries a referrer: apps, privacy settings and copy-paste strip it. And clicks from Google's AI Overviews and AI Mode arrive as ordinary Google organic traffic, so you can't separate them in GA4.

1:49 Layer 1b: first-party AI reports

Now a big change in twenty twenty-six: first-party AI reports. Search Console's generative AI report shows impressions of your pages in AI Overviews, AI Mode and AI Overviews in Discover, by page, country, device and date. No queries, no clicks. Bing Webmaster Tools' AI Performance report, launched in February twenty twenty-six, shows citations of your pages in Copilot and Bing's AI answers, which pages were cited, and the grounding queries Copilot generated to find them. Those grounding queries are gold. They're real sub-queries used to retrieve your content.

2:27 Layer 2: the prompt panel

Layer two: sampling answers with a prompt panel. Because answers vary, you measure with a fixed set of prompts, run repeatedly. Build thirty to a hundred prompts from real customer questions: recommendations, comparisons, problem-solution questions, brand questions and local ones. Source them from Search Console queries, Bing grounding queries, sales calls and support tickets. Include market and language variants: English, Arabic and Urdu; UK, UAE, Saudi Arabia and Pakistan. Run each prompt three to five times per surface, in clean sessions, on a fixed monthly cadence, and note the location.

3:06 Log it, then measure

For every run, log the date, surface, market and language, the prompt, whether your brand was mentioned, where it appeared, whether you were cited and which URL, which competitors appeared, any accuracy problems, the sentiment, and the sources cited. From that log come five metrics. Mention rate. Citation rate. Share of voice against competitors. Accuracy, meaning the share of mentions with correct facts. And source mix, which third-party sources appear when you're mentioned. Accuracy deserves special attention: being mentioned with the wrong price can hurt more than not being mentioned.

3:45 Example 1: Lahore wedding photographer

Worked example one, simple. A wedding photographer in Lahore can't afford tools. She builds a panel of fifteen prompts, like best wedding photographer in Lahore for small nikkah ceremonies, runs each three times in two assistants every month, and logs results in a spreadsheet. It takes about an hour. After two months she notices she's mentioned mostly when answers cite one particular wedding directory, which prompts her to complete that profile properly. Small panel, clear insight.

4:18 Example 2: Manchester HR SaaS (illustrative)

Worked example two, with illustrative details. A Manchester HR software company sets up measurement in a week. The GA4 channel shows AI assistant sessions are a small share of traffic but convert better than average organic. Search Console's AI report shows most AI impressions landing on three comparison guides. Bing shows Copilot citing the pricing page for queries about HR software cost per employee in the UK. And a forty-prompt panel gives a baseline mention rate and share of voice against four competitors. Those four numbers become the quarterly dashboard.

4:57 Tools and business impact

What about tools? A growing number of tools automate prompt tracking across assistants. They save time, but methods differ: some use APIs, some use the consumer apps, with different locations and logged-in states. So numbers aren't comparable across tools, and API answers can differ from what users see. Treat vendor visibility scores as proprietary indexes, useful for trends within one tool, not as ground truth. Keep a small manual sample alongside any tool. And for layer three, business impact, add AI assistants to your how did you hear about us question, and ask sales to log when prospects mention them.

5:40 Watch me do it: AI measurement setup

Watch me do it. I'll set up AI measurement for a small SaaS in one sitting. Step one: GA4. I go to Admin, Data display, Channel groups, copy the default group, and add a channel called AI assistants. The condition is session source matches the regex from the lesson. I drag it above Referral and save. Step two: I build an exploration with that channel as a dimension, and sessions, engaged sessions and key events as metrics, for the last ninety days. The channel is small but converts well. Step three: Search Console. I open the generative AI report and export pages with AI impressions. Three comparison guides account for most of them. Step four: Bing Webmaster Tools, AI Performance. I export cited pages and the grounding queries. I copy ten grounding queries into a new panel sheet. Step five: I add ten non-brand queries from Search Console and five questions from sales calls. That's a twenty-five-prompt starter panel. Step six: I run each prompt three times in two assistants, log mention, citation and accuracy, and compute the baseline. The dashboard now has four numbers: AI sessions and conversion rate, AI impressions by page, Copilot citations, and panel mention rate.

7:07 Common mistakes

Common mistakes. Reporting a single screenshot as we rank number one in ChatGPT. Comparing numbers from different tools as if they were equivalent. Forgetting that Google AI feature clicks are mixed into organic. Treating the Search Console AI report as a traffic report. And only measuring clicks, not mentions and accuracy.

7:29 Recap and try this now

Recap. Measure in layers: referrals in GA4, first-party AI reports from Google and Bing, a repeatable prompt panel, and business impact signals. Log every run and compute mention rate, citation rate, share of voice, accuracy and source mix. Try this now. Create the AI assistants channel in GA4 using the regex in the lesson text, then pull ten grounding queries from Bing Webmaster Tools, or ten non-brand queries from Search Console, and turn them into the first ten prompts of your panel.