A leadership team asks a simple question about AI search: what is it doing for revenue? The common answer is one number. Revenue per mention, a share-of-voice gap multiplied by market size, or a single ROI figure that adds referral sales to an estimated value of being mentioned.
Each of those numbers is built by multiplying two kinds of evidence that were never measured on the same thing. One kind is what ChatGPT, Gemini, Perplexity, and AI Overviews say: which brands the answer mentions and which pages it cites. The other is what people do on the brand's site: visits, leads, and sales.
This guide calls them the answer layer and the site layer. Both belong in the report to leadership, in separate columns. No figure should be computed across them.
Key takeaways
- AI search leaves evidence on two layers: the answers themselves, and the visits, leads, and sales on the brand's own site.
- In a Zumi export of 50,336 citations from AI answers, 91.6% went to third-party pages and 1.6% to the tracked brand's own pages, so most cited pages could never send the brand a visit that analytics credits to an AI engine.
- Any figure whose inputs mix an answer rate with visits, leads, or money is an estimate built on an assumption, not a measurement.
- The page is the one fair join: a cited page and the AI visits landing on it can sit in the same table row, with no ratio computed between them.
- When both layers move the same way, they are consistent with each other. Neither proves the other, and only a controlled test shows what the AI work caused.
What can the answer layer tell leadership on its own?
The answer layer is read before any click exists. It shows whether a brand gets mentioned for the questions buyers ask, how the answer describes it, which competitors appear instead, and which pages the engine cites as its sources. Zumi reports it as four signals: mention rate, how often answers mention the brand; share of voice, how often it appears relative to tracked competitors; average position, where it lands among the brands an answer mentions; and citation share, how often its own pages are cited as sources.
That is decision-grade evidence without any revenue figure attached. An answer that describes the product wrongly, or mentions three competitors and not the brand, points to a specific page to fix or a specific source to earn coverage in.
The answer layer has limits of its own. A mention rate describes the tracked questions, not every buyer. Answers also vary from one run to the next.
What can the site layer tell leadership?
The site layer is the familiar one. Analytics can group visits from chatgpt.com, perplexity.ai, gemini.google.com, and similar sources into an AI channel and follow them to leads and orders.
Three more sources fill in part of what referral data misses. Many AI visits arrive without any record of the site that sent them, and analytics files them as direct traffic. A "how did you hear about us" field with AI assistants as an option catches buyers who never clicked, and searches for the brand's name can be watched for change.
All of this measures people who reached the brand. None of it reads an answer.
A rise in searches for the brand's name after a month of AI work is consistent with that work. It can also come from a campaign, a launch, or press coverage, so the report names the causes it ruled out.
Why does one combined number fail?
The two layers count different things. A mention is counted per answer, whether or not anyone clicks. A visit is counted per person who arrived, whatever role an AI answer played in the decision.
Ownership breaks the link further. In a Zumi tracked-answer export on the AI visibility tools category, dated August 8, 2026, 46,118 of 50,336 citations, or 91.6%, pointed to third-party pages such as best-of lists and blog articles.
Competitors' pages took 6.8%, and the tracked brand's own pages took 1.6%. That is one category and one tracked brand, and the split will differ elsewhere.
A click on a third-party citation lands on that third party's site. Any visit that follows reaches the brand from that site, not from the AI engine. Referral analytics therefore sees mainly the slice of citations that point to pages the brand owns, and a formula that turns citations into expected visits starts from the wrong base.
The sample doesn't scale either. A share of voice measured across a few hundred tracked prompts describes those prompts. Multiplying it by a market's total value assumes the tracked prompts stand in for every purchase in the category, which no prompt set can show.
How should the monthly report be laid out?
Two sections, one for each layer, each with its own scope stated: which engines, which prompts, which dates for the answer layer; which channels, which tracking setup, which dates for the site layer.
The answer section reports the four signals by topic, the competitors mentioned in the brand's place, and the pages cited most often. The site section reports AI referral visits by engine, the leads and orders that followed, and the self-reported source counts.
The two sections meet in one table, joined by page.
| Page | Answer layer | Site layer |
|---|---|---|
| A page on the brand's site | Share of tracked answers citing it, and for which topics | AI referral visits landing on it, and what those visitors did next |
| A third-party page | Share of tracked answers citing it, and for which topics | Not available: visits from this page land on the third party's site |
The table carries no third column. Visits per citation, or revenue per cited page, would rebuild the combined number one row at a time. The table answers a narrower question: which pages matter on each layer, and where they overlap.
The third-party rows are where much of the work sits. A page the engines cite often that the brand doesn't own is a coverage target, and its value shows up in the answer layer only. How to track AI citations covers how to build that list of cited pages.
How can a crossed figure be spotted in a vendor or agency report?
The test is a list of the inputs behind the number. If they include a mention rate, share of voice, or citation count on one side and visits, leads, pipeline, or money on the other, the figure is crossed.
Crossed figures show up in a few common forms:
- Revenue or pipeline per mention, or per citation.
- A share-of-voice gap multiplied by market size, or by revenue at risk.
- A share of pipeline listed as one of the visibility metrics.
- One ROI figure that adds referral revenue to an estimated value of being mentioned.
A crossed figure isn't necessarily wrong about direction. It is an assumption presented with the precision of a measurement. The useful response is a request for the two layers, reported separately.
What wording keeps the report honest?
When mention rate and AI referral visits rise in the same quarter, the right sentence is that the two are consistent with each other. When they split, the next step is to read the answers and the landing pages, not to adjust a formula.
Words such as "proves", "confirms", "drives", and "generated" claim a causal link that neither layer measured. Attributed revenue is revenue from visits the analytics tied to an AI source. Some of those buyers would have arrived anyway.
A causal figure needs a controlled test, such as changing a set of pages and holding a matched set back, with the outcome measured on the site layer. Without one, the report shows direction on both layers and says so.
When is this not the answer?
A store whose AI referral traffic is large enough to follow to checkout can report attributed sales from AI sources in detail. That is still a site-layer figure, and it sits in the site section. It doesn't become a value per mention.
A team that runs controlled tests, with a matched set of pages held back, can report the lift those tests measured. That figure describes the test. It sits beside the answer-layer signals, never inside them.
Zumi, an AI Search Intelligence Platform, reports the answer layer: the four signals for each engine in the plan, with the answers and cited pages behind every figure. Book a demo to see that layer for a brand's own category.
Sources
- Zumi tracked-answer export on the AI visibility tools category, dated August 8, 2026: 12,625 cited pages and 50,336 citations, classified by the cited page's relationship to the tracked brand (third-party, competitor, own). The tracked brand in this run was Zumi itself. An industry run with no client data. Unpublished; figures recorded in Zumi's claims register.