STAT's Tom Capper: Stop Using Citations as Your AI Search KPI

An AI answer can recommend your product without ever linking to your site — and Tom Capper, Director of Search Product Strategy at STAT, says that's the core measurement problem search teams now face. In the Search Engine Journal webinar "Tactical Solutions For The Biggest AI Search Measurement Challenges," Capper argued that zero-click search is an attribution challenge, not just a traffic problem, and that teams should measure visibility alongside business outcomes rather than expect clicks to describe the full customer journey.

A customer who sees a recommendation might search the brand directly, visit the site later, or buy offline. The purchase still matters even when analytics can't connect it to the earlier answer.

Why brand mentions and citations are different things

Capper illustrated the gap with an AI-generated car recommendation: Audi was the brand being recommended, while the cited links pointed to entirely different websites. A report tracking only whether Audi's domain earned a citation would miss the recommendation itself.

His advice: define the brand entities you want to track, including sub-brands, rather than assume your domain captures every valuable appearance — and apply the same discipline to competitor sets so comparisons stay meaningful.

He drew the line directly:

"So yeah, citations have URLs. That's still valuable tactical data, but I don't think it should be your KPI. Mentions have entities."

— Tom Capper, during the webinar (24:10–24:30).

Mentions are the outcome to understand; cited URLs help explain which sources shape it.

How do you measure visibility when AI answers vary?

Non-deterministic answers — different responses to the same prompt — don't make measurement impossible. Capper framed it as a sampling problem, comparing it to the location- and device-specific Google results SEOs have always tracked. A small prompt set gives an unstable picture, especially when the business already tracks a much larger keyword set.

He cautioned against dumping a keyword list into a prompt tracker. Instead:

• Start from People Also Ask questions as a realistic prompt source
• Consider the personas, locations, budgets, and needs behind each question
• Sample situations broadly enough to recognize patterns, not to reproduce any individual's conversation
• Collect answers from the web interface, not the API — Capper said STAT found API responses can differ from what users actually see

What are citations actually good for?

De-emphasizing citations as a headline KPI doesn't mean ignoring them. Capper described starting with pages cited when competitors get recommended. Those pages reveal third-party sources worth investigating — review sites, industry publications — where a team can then work to improve representation and monitor whether those sources change alongside brand recommendations.

That reframes the reporting question: which sources changed, what work did the team do, and how did tracked answers change afterward? Keep the prompt sample and collection context consistent when making that comparison.

Which metrics belong in the report?

Capper recommended leading indicators that support — not replace — traditional performance reporting:

• Presence: whether the brand appears in the relevant search result or AI answer at all
• Recommendation and context: whether the answer recommends the brand, merely names it, or ties it to a strength or problem
• SERP prominence: how far down the page a result sits and how much visible space it occupies

Position alone misses real differences. Capper showed examples where a lower-ranked result appeared above the fold while a higher-ranked one was pushed below it — same rankings, different visibility.

He also warned against trusting volatile sentiment scores. A chart swinging sharply between positive, neutral, and negative may be amplifying small differences in essentially neutral answers; the qualitative themes behind a mention are often more useful than a summary score.

Can surveys fix AI attribution?

During the Q&A, Loren Baker asked how teams can connect visibility to the customer journey when multi-touch attribution can't observe the earlier exposure. Capper's answer: combine evidence rather than expect one analytics model to explain everything.

Post-purchase surveys can surface customers who say an AI recommendation helped them discover or choose a brand. Campaign tracking can separately show whether that brand is being recommended more often. He suggested passing survey answers into analytics as a dimension, so self-reported discovery can be compared against the channel analytics credits — while warning that displaying survey options in the same order every time can bias respondents toward the first option.

That's a fuller account of influence, not a promise to reconstruct every buyer's path or reclassify all direct traffic as AI-driven.

Putting the plan into practice

Capper's webinar closed with concrete changes for the next reporting cycle: audit what the current dashboard misses (brand mentions and recommendations, not just visits and citations); document brand entities, competitors, prompt contexts, and collection environment before comparing over time; record cited pages investigated or influenced and track how recommendations shift; and compare visibility trends with revenue, conversions, and survey answers rather than presenting visibility as a substitute for commercial results.

Search Engine Journal's next webinar, "AI Shopping: Get Your Products Found in Google & ChatGPT," features Chris Sachs, VP of Value Enablement, and Aris Vrakas, Product Lead, at seoClarity, covering how AI agents already decide which products shoppers consider — worth monitoring for teams tracking how AI-driven discovery moves from answers to transactions.