AI Search: 4 Measurement Solutions Beyond Clicks

Summarize this article with AI

In short: In short: Tom Capper (STAT) proposes measuring AI Search with four levers: separate brand mentions from cited URLs, track samples of realistic prompts, investigate the cited sources, then compare presence against business results using the post-purchase survey. Here is how to apply them to an online store.
4tactical solutions from Tom Capper’s (STAT) recap published by Search Engine Journal
1,300+semantic cocoons delivered since 2016, the foundation of my prompt grids
650+clients supported over 30 years of SEO expertise

Why don’t clicks tell the AI Search story anymore?

SEO in 2026 is no longer played on the click. It’s played on presence.

Here’s a scénario I see with my e-commerce clients. The dashboard shows stable organic sessions. Revenue grows. And at support, customers keep repeating the same sentence: « I found your store by asking ChatGPT. » The reports stay blank: no line, no channel, no source.

The problem is the measurement instrument, not AI Search.

Search Engine Journal published the recap of a talk by Tom Capper, of STAT, on this challenge: measuring visibility in AI-generated answers, distinguishing brand mentions from citations, and connecting prompt tracking to business results. The recap, written by Loren Baker, lays out five practical ideas. I grouped them into four solutions, because they roll out naturally in that order.

Why is this framework useful? Because many teams are still hunting for a single number. An « AI visibility score » that fits on one slide. In my view, that number doesn’t exist. An AI answer changes from one day to the next, from one phrasing to the next. A click is an event. A recommendation is a context. They are two distinct things.

Measuring one with the tools of the other gives you empty or misleading dashboards. Neither helps you decide.

What you gain from reading on: a four-layer tracking grid, ready to launch with a simple spreadsheet, and designed for an online store that wants to know whether AI assistants name it, cite it and recommend it.

Before detailing each solution, one clarification. I rely on the points published in the Search Engine Journal recap. The e-commerce applications that follow are mine. I flag them as such.

Solution 1: how do you separate brand mentions from cited URLs?

First distinction in the recap: track brand and sub-brand mentions separately from cited URLs. Two events. Two columns in your spreadsheet.

A mention is your name written in the answer. A citation is a URL displayed as a source. Each can exist independently of the other.

The four possible cases.

SituationWhat it meansPriority action
Mention + citationThe AI names you and relies on your pagesProtect and extend those pages
Mention without citationYour reputation circulates, but other sources carry itIdentify those sources (solution 3)
Citation without mentionYour pages feed the answer, your brand stays invisibleStrengthen the brand name on those pages
NeitherYou don’t exist for this promptBuild the missing semantic coverage

A diagnosis, not a score.

Why this separation? Because the two metrics respond to different levers. The mention depends on your presence in the ecosystem: reviews, press, comparison sites, discussions. The citation depends on the quality and structure of your own pages. Merge them into a single indicator and you no longer know which lever to pull.

The recap also specifies sub-brands. For an e-merchant, that’s a real field question. A store often has several names to track: the site name, the distributed brands, the private-label ranges, sometimes a trade name that differs from the legal name. Each one can be cited, distorted or ignored by an assistant.

E-commerce application (my interpretation): list in a sheet every variant of your name, your ranges, your abbreviations and the common typos. For each tracked prompt, tick three boxes: brand named, sub-brand named, URL from your domain cited.

The principle is simple. What you don’t distinguish, you can’t improve.

One point of attention: a mention is not a recommendation. The AI can name your store to say it’s more expensive, or slower. We come back to this in solution 4. For now, keep the rule: two columns, never just one.

And you — in your current reports, do you know whether your brand is named or only your site is cited?

Solution 2: how do you track realistic prompts when every answer changes?

Second point of the recap: use samples of realistic prompts to understand trends in AI answers that keep evolving.

The key word is trends. Not positions.

In classic SEO, you tracked a keyword and a position. A position is a stable number across a single day. A generated answer, on the other hand, varies. Same question, similar phrasing, different context: the answer shifts. Capturing a single answer and presenting it as « your visibility » is like photographing one wave to describe the ocean.

The logic to adopt is that of a poll. A sample of prompts, repeated over time, read as frequency of presence rather than as a single result. If your brand appears in a growing share of the answers for a group of prompts, the trend is readable. If it appears once on an isolated prompt, you’ve learned almost nothing.

What does « realistic » mean? A realistic prompt uses your customers’ words, not your marketing department’s. A buyer doesn’t type « best ergonomic bedding solution. » They write a long sentence, with a constraint, a budget, a doubt.

Here is where I draw the material to build a sample:

This is where my work on semantic cocoons becomes an asset. A cocoon already maps a market’s questions by intent. Each branch of the cocoon provides a natural group of prompts. The architecture that structures your pages also structures your tracking sample. After more than 1,300 cocoons delivered since 2016, it’s the first reflex I recommend.

Reading rule: group your prompts by intent, then track the frequency of presence per group. A « comparison » group that progresses while a « purchase » group stalls tells you precisely where to work.

A method tip: freeze the sample. Changing the prompts at each reading breaks the comparison over time. Add prompts in waves, keeping a constant core.

Look for a map of your presence zones and your blank zones, not a ranking.

Solution 3: how do you connect your actions to recommendations through cited sources?

Third lever of the recap: investigate the cited sources to connect campaign work to the recommendations that change.

The word is strong. Investigate.

When an assistant recommends a competitor, the answer often shows where the information comes from. These sources are your field of investigation. They tell you what the AI « reads » to build its opinion of your market.

The method unfolds in three steps.

Step 1: inventory. For each group of prompts, note the cited domains. Comparison sites, specialized media, forums, merchant listings, brand sites, videos. You get a map of the sources that matter in your universe.

Step 2: spot the recurrences. Some domains come back from one prompt to the next. These are your most profitable levers of influence. A domain cited everywhere weighs more than a domain cited once.

Step 3: connect to your actions. You ran a campaign: PR, partnership, review program, a new series of buying guides. The question becomes: do the sources touched by this campaign now appear in the answers? And has the recommendation changed?

This is what the recap calls connecting campaign work and the evolution of recommendations. A plausible, documented causal chain, rather than an act of faith.

Keep an action log: date, action, targeted page or source. Next to it, your presence readings by group of prompts. When a recommendation evolves, you trace the thread back through the log. Without a log, you observe a change but can’t explain it.

Caution is needed. Correlation is not causation, and an AI answer depends on many factors you don’t control. So I present this approach as a hypothesis you test, not as proof. Repeated clues across several prompts and several weeks count for far more than a simple coincidence.

For an e-merchant, the value is very concrete. You learn where to invest. If your competitors are recommended thanks to two comparison sites and a forum, your content and partnership budget has a precise target. Instead of producing in every direction, you produce where the AI goes looking.

Which sources does the AI cite when it talks about your product category? If you don’t know, you’re investing blind.

Solution 4: how do you connect AI presence and business results?

Last axis of the recap, and the most important for an executive: measure presence, recommendation context and SERP prominence alongside business results, then compare post-purchase survey answers with analytics attribution.

Three visibility measures. One link to revenue.

Presence. Does your brand appear in the answer? It’s the foundation of solutions 1 and 2.

Recommendation context. How are you presented? First choice, one option among others, cheaper alternative, product with reservations? An enthusiastic mention and a nuanced mention don’t have the same commercial value. Score it on a simple scale: recommended, neutral citation, citation with reservations.

SERP prominence. Classic search results remain a touchpoint. Tracking your SERP position alongside your presence in AI answers avoids pitting them against each other. The two either reinforce or contradict each other. Worth watching closely.

Next comes the link to business results: orders, basket, new customers, brand searches. Not to prove a mechanical causality. To observe whether presence curves and commercial curves move in the same direction.

That leaves attribution. And this is where the post-purchase survey really matters.

Classic analytics struggle to label what comes from an AI assistant. Search Engine Journal points this out in one of the guides featured on the page: calls coming from ChatGPT and Claude are often classified as direct traffic. Same mechanism for visits: a visitor who types your address after reading a recommendation arrives « direct. » The real channel disappears.

The survey fixes this blind spot. One question, at order confirmation: « How did you discover us? » Offer a free-text field, with a few options. Then compare the answers with what your analytics tool says.

Reading the gap: if customers say they found you through an AI assistant while analytics files them under « direct, » the gap measures the invisible share of your acquisition. That gap becomes your best argument in the executive committee.

The two sources complément each other. Analytics gives the traceable volume. The survey gives the declared origin. Between the two, you rebuild a credible attribution story. The recap talks about improving your « attribution story »: the phrase is right. It’s about telling, with evidence, how a customer reached you.

No magic number here. Cross-checked signals. And that’s enough to decide.

Is the click dead? No, it has become incomplete

Here is the point I defend, and it runs against some alarmist discourse.

The click isn’t dead. It’s incomplete.

A customer can discover your store in an AI answer, then search for it by name, compare on a third-party site, then buy. The last recorded click looks like a brand search or direct traffic. The decisive moment, meanwhile, happened upstream, leaving no trace.

The problem isn’t the loss of clicks. It’s the loss of visibility on what precedes the click.

This reading changes how you steer. You gain by accepting a measurement that is imperfect but cross-checked, rather than waiting for the perfect tool. Tom Capper’s recap goes in this direction: no miracle metric, a combination of brand tracking, prompts, sources and declarative attribution.

Another important point: a mention without a click has value. Many e-commerce managers judge it useless because it doesn’t show up in traffic. That confuses what is easy to measure with what matters. A brand named in an answer prepares a brand search, a comparison, a basket. It works upstream.

Third point, more delicate: a citation without a mention can hurt you in the long run. Your pages feed the answer, while a competitor or a marketplace takes the credit. Hence the importance of separating the two columns from solution 1.

What should you take away for governance? Three simple principles.

The dashboards that inspire confidence in committee are the ones that own their gray areas. They distinguish what is certain from what is estimated and from what they are testing. That measurement candor protects you better than any flattering indicator.

To measure AI Search impact on your ecommerce store, follow this four-step process. Each step builds on the previous one, from mapping your brand presence to investigating the sources that recommend or omit you.

4-Step AI Search Measurement Workflow

Map → Sample → Record → Investigate

Where do you start in an online store?

You don’t need a new tool to get started. A spreadsheet, a list of prompts, a survey field. I recommend this order, which follows the four solutions.

StepActionDeliverable
1. MapList all name variants, distributed brands and rangesBrand dictionary
2. SampleBuild realistic prompts by intent (discovery, comparison, purchase)Frozen sample, grouped by cocoon
3. RecordNote mention, citation, recommendation context, cited sourcesSpreadsheet with two columns minimum
4. InvestigateSpot recurring domains, connect them to the action logList of influence sources
5. ReconcileAdd a discovery question to the checkout flow and compare with analyticsDocumented attribution gap

Start small. A modest sample, recorded regularly, beats a large sample recorded once.

Two mistakes to avoid from the start. The first: changing prompts at each reading, which makes any comparison impossible. The second: waiting three months before looking at the first results. Trends read better with a regular rhythm, and the first blank zones appear from the very first readings.

On the organization side, one person is enough to maintain the spreadsheet, as long as the post-purchase survey is reviewed by the same person who reads the analytics reports. Connecting the two creates the value. Alone, each falls short.

And on the content side? Each blank zone you identify becomes a precise editorial project. A « comparison » group of prompts with no presence calls for a structured comparison page. A citation without a mention calls for stronger brand naming and entity signals on the page. I build these projects from the site’s existing architecture, branch by branch, so that every page has a clear role.

Quick test: pick ten questions your customers really ask before buying. Put them to an AI assistant. Note whether your brand is named, whether your pages are cited, and who is named in your place. You’ll have your first map in a single session.

Tom Capper and the Search Engine Journal recap say it well: it’s a discipline more than a tool. Separate, sample, investigate, reconcile.

So, among this month’s customers, how many discovered you in an AI answer while your reports kept not a single trace of it?

Your AI Search presence, in a live audit

I take your ten most frequent customer questions, run them through the AI assistants, and show you live who is named, who is cited, and on which sources. I don’t sell you the method. I show you the pages.

Book a strategic call — 45 min

Frequently Asked Questions

What’s the difference between a brand mention and a citation in AI Search?

A mention is your name in the answer. A citation is your URL as a source. Each can exist independently of the other: they belong in two separate columns.

Why use samples of prompts rather than a single reading?

AI answers change with the phrasing and the moment. The same test, repeated regularly, shows trends. A single reading shows only an instant.

How do you connect a campaign to the evolution of AI recommendations?

Inventory the sources cited per prompt, spot the recurring domains, then connect them to your action log (press, partnerships, guides). It’s a lead, not proof.

Why does traffic from AI assistants show up as direct in analytics?

Search Engine Journal observes it: calls from ChatGPT and Claude often arrive as direct. A visitor who types your address after a recommendation works the same way.

Which question should you ask in a post-purchase survey to measure AI Search?

Ask: « How did you discover us? » Offer a free-text field and a few options — AI assistants included. Compare the answers with the attribution from your analytics tool.

Stéphane Jambu

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn