Measuring the business impact of AI search

Visibility in AI answers can be measured. Revenue caused by that visibility cannot be measured directly by anyone today. That gap is why AI search work often fails to get funded: the person approving the budget asks what it returned, and the honest answer is four indirect methods and a confidence statement. This page sets out the four, what each one misses, and how to combine them.

Why the causal link is missing from your analytics

A generated answer frequently resolves the question on the surface. When no click follows, no referral event is recorded anywhere in your stack. The reader may still act: they remember a brand name, and later search for it, type the domain, or arrive through another channel. By the time a session exists, it looks branded or direct. The answer that caused it left no trace.

Google's own reporting stops short of closing this gap. Google states that "sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console" and are "reported on in the Performance report, within the 'Web' search type", per its AI features documentation. The dedicated generative AI performance report, introduced in June 2026, reports impressions in AI Overviews and AI Mode broken down by page, country, device and date. It carries no clicks, no click-through rate, no position and no queries, and Google says it is still rolling out to a subset of properties. The surfaces themselves are covered in AI Overviews and AI Mode. First-party AI reporting tells you that you were shown. It does not tell you what that was worth.

Four routes, and the weakness of each

1. Referral identification in analytics

Segment sessions by referrer host and count what arrives from the assistants that pass one. It is the only method that produces a session-level number tied to an identifiable engine, which makes it the natural starting point.

Its weakness is scope. It captures only the click-through minority, so it undercounts the channel it is meant to measure. Referrer behavior also differs by engine and by client: mobile apps, in-app browsers and desktop clients do not behave alike, and no engine publishes a referrer specification. Every host list in circulation is maintained by observation and goes stale without warning.

2. Branded search lift

Track impressions and clicks for queries containing your brand name in Search Console, month over month. Of the four, this is the leading indicator most consistent with how AI discovery works: the reader learns a name inside an answer, then searches for it. When AI visibility rises and branded search rises behind it, the sequence is at least the right shape. See Search Console for the report itself.

Its weakness is competition for the same effect. Branded search rises for paid campaigns, press coverage, product launches and seasonality. Search Console query data is also filtered and capped, so long-tail brand variants and misspellings are missing from the total.

3. Direct traffic lift

Apply the same period comparison to the direct channel, on the theory that a reader who saw your name in an answer types the domain later.

This is the weakest of the four, because direct is a catch-all rather than a channel. Untagged email, QR codes, app opens, bookmarks, stripped referrers and misconfigured redirects all land in it. A rise in direct is consistent with AI discovery and equally consistent with a dozen other explanations, including a tracking change you made yourself.

4. Self-reported attribution

Add one field at the point of conversion asking how the customer heard about you, on the order form, the lead form or the first onboarding step. Keep it to a single question with a short list plus a free-text option.

Its weaknesses are ordinary survey weaknesses: recall is imperfect, people name the last thing they remember, response rates vary by form design, and free text has to be coded before it can be counted. Its strength is unique. It is the only one of the four that sees the no-click path, and the only one that asks the customer instead of inferring from a log.

Building a read you can defend

• Use all four. No single method carries the claim. Agreement between independent, differently biased methods is the evidence.
• Compare against a pre-period of equal length. Match the season where you can, and note whatever else changed in the window: campaigns, launches, press, pricing, migrations.
• Hold the visibility side constant. A fixed prompt set on a monthly cadence, as described in measuring AI search visibility honestly, is what makes the input comparable across periods. If the prompt set moved, the comparison is not one.
• Define the numbers you report. Visibility metrics are not standardized, so state which definition you used, as set out in the GEO metrics, defined.
• Expect correlation and say so. Write the confidence level into the report itself. "Branded search rose over a period in which AI visibility also rose, while two other campaigns were running" is a defensible sentence. The same movement reported as a result caused by AI search is not.

What not to do

• Do not attribute all direct traffic growth to AI. Direct is the bucket everything unexplained falls into, which is why it is available to be misread.
• Do not present correlation as causation. Two lines moving together is the starting point of an investigation, not its conclusion.
• Do not report a single month as a trend. Sampled visibility and self-reported attribution are both noisy at small volumes, and one month is inside the noise.
• Do not build revenue from a value-per-citation figure. No published source supports a per-citation value, and a number multiplied by an invented constant is an invented number.

The honest close

No vendor and no analytics product can currently prove that revenue was caused by an AI citation. The decisive event happens inside systems outsiders cannot observe, and Google states plainly in its AI features guide that "No third-party tool has access to our internal ranking or AI systems." Everything available from outside is inference from the residue.

That does not make the work unaccountable. It makes the standard of evidence explicit. If a product claims causal revenue attribution for AI search, ask three questions: what does it observe directly, what does it infer, and what error does it report on the inference. A method that cannot be described is not a method.