AI Search Has a Dashboard. Does It Have a Return?

Evidence-based marketing says test the outcome you care about. In AI search, that means following the buyer's question to a recommendation, a transaction and, where the data allows, incremental return.

If your AI visibility score rose this quarter, what would you need to see before increasing the budget? More people exposed to your brand? More AI assistants recommending it? More orders that would otherwise have gone elsewhere?

Those are different outcomes. A CMO who would demand evidence of incremental return from a media investment should ask the same of an AI search program funded to grow sales. The first question is what the vendor actually measured.

What evidence-based marketing asks of us

The Ehrenberg-Bass Institute's contribution to marketing was to make observed buying behavior the starting point. Look for patterns in what buyers do, then test whether those patterns recur across brands, categories, and markets. That is the principle I would apply to AI-mediated buying. If the outcome you care about is which brand gets recommended, observe the recommendation. If you care about sales, observe the transaction. If you want to know whether your spending caused additional sales, design a test that can answer that question. Ehrenberg-Bass Institute's research history · Frank Bass on empirical generalizations

A thousand first prompts are still first prompts

A single-prompt test sends a question and records one answer. It can tell you which brands appear, which pages are cited, and, if the question asks for a choice, which brand the assistant picks given the information in that prompt. Some shoppers stop there. If your research question is what an assistant says in its opening answer, this is the right test.

A multi-turn test follows the same conversation as the buyer supplies new information. Ask for a small-business bank account and an assistant might recommend a low-fee online bank. Add that the business takes cash at a weekend market and needs to deposit it every Monday. Does the recommendation survive? Putting every requirement in the first prompt can show what the assistant chooses when it knows the full brief. It takes a conversation to see whether the assistant asks about cash deposits, when its recommendation changes, and which competitor replaces it. Repeating isolated prompts cannot recover that sequence.

In AIVO Standard's Beyond Visibility study, we ran 1,427 structured probes across ten industries. In 87.3% of probes, a competitor had replaced the brand established at the opening turn by turn four (95% confidence interval, 85.5% to 88.9%). The figure describes the tested conversations with a brand established up front. A change is not automatically a loss. If the buyer adds a requirement the first brand cannot meet, switching is the right answer. That is why the conversations themselves matter: they let a team tell a warranted switch from one driven by missing or inconsistent evidence.

In a separate study, we ran identical prompts three times across 1,800 brand-level groups. The cited sources changed in 95.8% of groups, while the final brand choice changed in 14.6%. If you watched only citations, you would have seen far more movement than there was in the decision. AIVO's displacement study and methodology · AIVO's repeat-run findings

This is what AI Win-Rate measures: the share of completed test conversations in which an assistant or agent recommends the brand outright. We repeat realistic buying scenarios and report the results with confidence ranges. We also retain the conversations so a team can see who won instead and at which turn the recommendation changed. An assistant's stated reason gives the team something to investigate. A follow-up test can check whether changing the relevant evidence changes the outcome. Win-Rate is closer to the decision than a visibility score. It is still not return. That takes the next two steps. AIVO's AI Win-Rate methodology

Agentic commerce takes measurement past the recommendation

Visa's Agentic Commerce Benchmark put 22 models through 15 shopping scenarios. Across repeated attempts, 57.2% of runs passed. A pass required the model to uncover the shopper's relevant requirements and make the right purchase decision, including deciding to buy nothing when nothing fit. That benchmark asks whether the agent made a sound decision. A merchant also needs to know whether the agent found the right product in stock and completed the order. Visa Agentic Commerce Benchmark

Those stages need their own records. A retailer can link a product answer to a subsequent click, booking, or order through a reference that survives the handoff. In one small controlled AIVO Assist test, the system correctly credited all 14 bookings that followed an answer, with zero false credits. That is evidence that the tracking worked under the test conditions, not proof of accuracy at scale: with 14 bookings, the true rate could still be as low as about 77% at 95% confidence. To estimate incremental bookings, compare the outcome with a credible control, such as a randomly assigned holdout where feasible. To estimate ROI, compare the incremental contribution after variable costs with the cost of the program. Give the same scrutiny to AIVO's tools as you would to any other supplier's. AIVO's controlled attribution test · Google's distinction between attributed and incremental conversions

Why are CMOs still buying visibility as a growth measure?

Generative engine optimization (GEO) and answer engine optimization (AEO) platforms can help teams see where brands appear, which sources assistants cite, and where content needs work. A CMO can reasonably buy those capabilities to diagnose discoverability. The harder question is why some teams still award or renew a growth mandate when the evidence they ask for stops at mentions, citations, or share of voice.

One plausible reason is that the dashboard fits the way companies already buy search. The SEO team has the budget. The supplier can provide a baseline quickly. The dashboard can show movement by the next quarterly review. Measuring recommendation requires following buying conversations. Measuring commercial return takes cooperation from analytics, commerce, and sometimes retail partners. That organizational difficulty helps explain the appeal of a visibility score. It leaves the return on the spend unresolved.

Even vendors in the field are working on that gap. AirOps is developing a common measure of AI search exposure that could feed marketing mix models and incrementality tests. An estimate of how many people were exposed would be a step beyond counting brand appearances in a sample of prompts. It would still take outcome data and a credible test to estimate what the exposure contributed. AirOps on its AI search measurement initiative

Before signing or renewing, ask the supplier to take one important buying situation through to a decision. Have it show who appears in the first answer, who wins after follow-up questions, and what it can link to a booking or sale. If the business case promises incremental growth, agree on the comparison and the result that would justify expanding the spend.

And ask one more question before the contract is signed. If visibility goes up and recommendation, sales, and incremental profit stay where they are, would you renew?