Measuring AI Search: Tom Capper’s 3 Metrics (STAT)

Summarize this article with AI

In short: A rank tracker measures a list of links. AI answers aren’t a list of links. According to Search Engine Journal, Tom Capper (STAT) proposes tracking 3 things: brand presence, citations and the attribution story. I’ll show you how I deploy them with my clients.
3metrics to track: brand presence, citations, attribution story
5measurement principles detailed in the SEJ recap of October 26, 2026
1 300+semantic cocoons delivered since 2016, the foundation of my prompt banks

Why your rankings report no longer sees anything

I review my clients’ visibility reports every week. They all have the same gap.

One « average position » line. One chart of tracked keywords. And across the table, a client telling me: « We’re cited in ChatGPT, but I don’t know what that’s worth. »

That’s the gap. The rank tracker measures an ordered list of links. An AI answer has no ordered list. It has a paragraph. Sometimes three brands named. Sometimes two links cited. Sometimes no link at all.

Nothing to position. Nothing to rank.

On October 26, 2026, Search Engine Journal published the recap of a talk by Tom Capper, of STAT, on this exact topic: Tactical Solutions For The Biggest AI Search Measurement Challenges. Written by Loren Baker, 5-minute read. Dense. I read it twice.

What stood out: Capper isn’t selling a new miracle tool. He proposes changing what we measure. The real question is the one you ask the tool.

Transparency on my sources. Everything I attribute to Tom Capper comes from the recap published by Search Engine Journal. I didn’t attend the session. The rest of the article is my own observations with my clients, with no invented figures: when I have no measurement, I say so.

Why do rank trackers break down? Three mechanisms, which I observe every week.

The result: the weekly report shows stable positions while demand moves elsewhere. You’re steering with a dashboard that watches the wrong screen.

So can this be fixed? Capper’s three metrics work with tools you already have, plus a spreadsheet and one field in your order form. I’ll walk you through them one by one.

And you: what does your latest rankings report say about what an AI tells people about your brand today?

Metric 1: brand presence, separated from cited URLs

The first principle in the SEJ recap, and the most important: track mentions of the brand and sub-brands separately from cited URLs.

Two events. Two rows in your table.

A model can name your brand without ever linking to your site. It can also link to a page on your site without ever saying your name. If you blend the two in a single counter, you can’t tell what’s moving forward.

SignalWhat it tells youWhat it doesn’t tell you
Brand mentionThe model knows you and associates you with the queryWhether it sends traffic to your site
Sub-brand mentionWhich range or product carries the awarenessWhether the parent brand benefits
Cited URLWhich page the model uses as a sourceWhether the reader remembers your name

Sub-brands are where I see the most blind spots. Take an e-commerce client who distributes third-party brands and also sells its own range: the assistant recommends the in-house range under its product name, never under the retailer’s name. In a combined report, you conclude « weak presence. » Separate the two, and you find the range is well recommended. The retailer simply isn’t associated with it.

Visibility wasn’t the problem: the missing link was between the product and the brand.

How do I build this tracking in practice?

In my view, this is the most underrated metric of the three. It resembles brand awareness, which has been hard to quantify in SEO until now. At last, you can count it.

One honest limit: this presence shows a sample, not an official ranking. I read it as a trend over several weeks, not as an isolated score.

Does your current tool distinguish your brand name from your site’s address? If the answer is no, you’re blending two signals.

How to build a prompt sample that looks like your customers

The second principle: use realistic prompt samples to understand trends across changing answers.

The key word is « realistic. » Not the keywords from your tracking sheet pasted into a chat window. The sentences your customers actually write.

A keyword is « men’s trail shoes. » A prompt is « I’m looking for trail shoes for muddy terrain, I have wide feet, tight budget, what do you recommend? » The second triggers a comparative answer with named brands. The first, rarely.

Where do I source prompts for my clients?

Then I organize the sample by intent: discovery (« what solutions exist for… »), comparison (« X or Y for… »), decision (« where to buy… », « best choice for… »). Each intent has its own reading. Being absent at discovery doesn’t carry the same cost as being absent at decision.

The rule I apply. I never draw conclusions from a single answer. I rerun each prompt several times, note what comes back, and look at the frequency. A brand that appears in almost every run is established. A brand that appears once in five is a lucky break.

Why does this discipline matter? Because answers change. Measure once, and you measure noise. Measure a realistic sample, several times, and you measure a pattern.

The counter-intuitive part? A smaller but better-chosen sample beats a large volume of generic prompts. A few dozen questions that look like your real customers say more than a thousand keyword variations. Representativeness beats quantity.

How many of your tracking prompts would a customer actually have typed?

Metric 2: investigate citations, don’t just count them

The third principle in the recap: investigate the cited sources to connect campaign work to evolving recommendations.

Investigate. Not count.

Counting citations gives you a total that goes up or down. Investigating means asking who is cited in your place, why that source, and what has changed since your last action.

My method, in four steps.

  1. I list all the cited sources for each prompt in the sample. Not only yours: everyone’s.
  2. I sort them by family: brand site, competitors, comparison sites, specialized media, forums, review platforms, merchant listings.
  3. I spot the prompts where you’re absent while a competitor is present, and I read the cited pages. What do they have that you don’t?
  4. I date my actions (article published, partnership, listing updated, PR campaign) and check whether the cited sources change afterward.

The fourth step is the one missing almost everywhere. Without it, you can’t link an action to an effect. With it, you build evidence, even if imperfect.

What I observe with my clients: very often, the model relies on third parties rather than on the brand’s own site. A comparison published elsewhere. A discussion thread. A media buying guide. The official site comes after.

It’s counter-intuitive. The best visibility lever in an AI answer isn’t always your own site. Sometimes it’s a third party’s page that the model judges reliable, and that you can influence by becoming a citable reference on the topic.

That echoes what Lily Ray and Aleyda Solis note: citation follows the perceived credibility of the source, not just the optimization of a single page.

And on the content side, the logic remains the one I’ve defended for years. A cited page is a page that answers a question precisely, within a coherent whole. Architecture matters. People rarely choose an isolated piece of content. A cocoon that covers a topic end to end gives the model several reasons to pick you.

When did you last read the page the AI cites instead of yours?

Being cited isn’t enough: recommendation context and the SERP

The fourth principle in the SEJ recap: measure presence, recommendation context and SERP prominence, alongside business results.

Four layers. Each one informs the others.

Presence: are you there? We covered that above.

Recommendation context: how are you there? Being cited as the first recommendation, as a simple alternative, or in a list of ten names where nobody remembers the seventh are three different things. So is being presented as the pick for a specific need versus a generic option.

For each mention, I note a role, using a simple grid:

This grid takes ten minutes to define and changes how you read the report. Presence that grows but stays confined to neutral mentions doesn’t produce sales. I see it regularly.

SERP prominence: classic search hasn’t disappeared. Results pages remain a place of decision, and that’s where position tracking keeps its usefulness. In my view, the trap is throwing out the rank tracker. It no longer measures everything. It still measures part of the picture.

Business results: sales, quote requests, calls, subscriptions. Capper’s principle, as SEJ summarizes it, is to never look at the first three layers without the fourth.

My standard four-column dashboard. Presence (name cited, frequency across the sample). Context (role of the mention). SERP (prominence on key queries). Business (sales or leads over the same period). One row per prompt group. No global average that hides the gaps.

Why keep all four side by side? Because each combination tells a different story.

Each of these cases calls for a different decision. An isolated number calls for none.

In your current reporting, which of these four layers is missing?

Metric 3: the attribution story, or what your customers actually tell you

The last principle in the recap: compare post-purchase survey responses with your analytics attribution.

It’s the simplest metric. And the most powerful.

The mechanism is familiar to anyone who has watched their direct traffic grow with no explanation. A buyer asks an assistant for a recommendation. They read your name. They open a tab, type your brand, land on the site, buy. Your analytics writes: « brand search » or « direct. » The assistant doesn’t exist in the chain.

SEJ sums it up in one of its guides: visits coming from ChatGPT and Claude are often classified as direct traffic. The channel is real, the reporting line doesn’t exist.

Capper has a simple answer: ask your customers.

How do I deploy it?

  1. One question on the order confirmation page or in the post-purchase email: « How did you hear about us? »
  2. As a free-text field, not a closed list. My opinion: a closed list hands you back your own catégories. A free-text field hands you their words, and they write « ChatGPT, » « Claude, » « DeepSeek, » « an AI recommended it to me. »
  3. A monthly grouping of responses into families: AI, word of mouth, social media, classic search, advertising.
  4. A comparison with the analytics breakdown over the same period.

The gap between the two columns is the attribution story. If 100% of your « direct » becomes « search engine » in the questionnaire, great, your analytics is reliable. If part of that direct turns into « AI » in the answers, you’ve just discovered a channel your reports couldn’t see.

I’m deliberately giving you no typical percentage. I have no reliable, quantified measurement to offer, and I’d rather say so than invent an order of magnitude. Yours will be specific to your market.

But look at what the gap makes possible. It turns an impression (« it looks like AI is bringing us customers ») into an argument you can defend in front of a finance director. You move from intuition to declared evidence. Imperfect, but it works.

The most reliable attribution source in AI search isn’t a tool. It’s a question asked at the right moment, to the right person.

And if you asked that question tomorrow, what would you discover in your « direct »?

Your 5-step plan to replace position tracking alone

Here’s how I’d proceed on a site, in this order. No need to change your tools. A spreadsheet to start.

  1. Separate the counters. One column for brand and sub-brand mentions. Another for cited URLs. No more single total.
  2. Build a realistic sample. Customer sentences, drawn from customer service, Search Console, reviews and your semantic cocoon. Sorted by intent: discovery, comparison, decision.
  3. Run, rerun, count frequencies. Each prompt several times, across several assistants. You read trends over several weeks, not one-day scores.
  4. Investigate the sources. List who is cited, sort by family, read the winning pages, date your actions and watch what moves afterward.
  5. Cross-check with business and declared data. Recommendation context, SERP prominence, sales or leads, and a free-text field in the post-purchase survey to challenge your analytics.

The trap to avoid? Chasing the single number that replaces average position. It doesn’t exist. Tom Capper, according to SEJ, says the opposite: multiple signals, side by side, read together.

What I take from it for my clients. Three rows in the dashboard (presence, citations, attribution), one post-purchase survey, and a monthly meeting to review the gaps. It’s modest. And it lets you decide what to produce, what to restructure and where to invest, instead of staring at a positions chart that tells you nothing about the market anymore.

I’m not selling you the method. I’m showing you the pages: during an audit, I take your twenty prompts closest to your customers, run them in front of you, and we read the cited sources together. You see live who is recommended, who is cited in your place and on which page.

That’s often where the decisions become obvious.

So the next time a client tells you « we’re cited in ChatGPT, » will you be able to answer with a number, a source and a sale?

Audit your visibility in AI

I run your 20 prompts closest to your customers live, we read the cited sources together, and you leave with your 3 metrics ready to track.

Book a strategic call — 45 min

Frequently Asked Questions

Why don’t rank trackers measure AI search correctly?

An AI answer has no position 1, 2, 3. It varies from one run to the next and can name a brand without a link. A rank tracker measures a list of links, not that.

What’s the difference between a brand mention and a URL citation?

A mention is your name spoken in the answer. A citation is a link to a source page. One can exist without the other: Tom Capper advises tracking them separately.

How do you build a realistic prompt sample?

Take your customers’ sentences: customer service, long Search Console queries, reviews, questions from your semantic cocoon. Sort them by intent and run them several times.

How do you connect a campaign to an AI’s recommendations?

List the sources cited per prompt, sort them by family, date your actions (article, partnership, listing) and check whether those sources change afterward. It’s indicative evidence, not absolute.

How do you measure the impact of ChatGPT or Claude on my sales?

Add the question « How did you hear about us? » as a free-text field in the post-purchase email. Compare the answers with your analytics: you’ll find the AI channel hiding in your direct traffic.

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