A recommendation that never appears in the report
A marketing lead opens the dashboard on a Monday morning. Sessions are flat. The channel report shows nothing unusual. Then a support ticket mentions a competitor that never appears anywhere in the funnel: three weeks earlier, an assistant recommended it while answering a question the brand had spent a year trying to rank for. No click arrived. Nothing in the analytics stack recorded the recommendation. The brand was visible in the place where the decision was forming, and completely invisible in the place where results are counted.
This is not an exotic edge case. Branch surveyed 300 enterprise marketing and growth leaders and found that 89% said AI search improved marketing performance in 2025 (Branch enterprise benchmark, reported by Business of Apps, 2026). The same survey found that 26% could not trace the path from AI discovery to conversion, and 24% said their analytics tools did not yet support AI attribution. Confidence and measurement are moving in opposite directions: the channel is widely credited, and almost nobody can put a number on it.
What each signal actually sees
| Signal | What it records | Where it goes blind | What it can still tell a team |
|---|---|---|---|
| Sessions | Visits that brought a click | About 70% of AI traffic lands in Direct | The volume that survived the click |
| Referral source | The domain that sent the visit | AI answers rarely leave a clean referrer | Which third parties mention the brand |
| Share of AI Voice | Mentions across a sampled prompt set | Probabilistic: resample and it shifts | The direction of visibility over time |
| Brand search volume | Queries for the brand name | Blends many causes | Whether exposure and demand move together |
| Conversion attribution | The path from discovery to sale | 26% of teams cannot trace it | Honest bounds on what is known |
A metric built for a moving target
Share of AI Voice, or SOAV, filled the gap: a brand's mentions divided by all tracked mentions, multiplied by 100 (GEO Playbooks glossary). It is simple enough to explain in one sentence and fragile enough to mislead a board. AI answers are generated by sampling against a prompt, not retrieved from a fixed index. Ask the same question twice and the two answers may name different brands. A traditional ranking is deterministic — the same query returns the same order. SOAV is one reading from a distribution, and a single reading is a thin basis for a quarterly target.
The tooling market grew around this anxiety. Meticulous Research put the GEO software market at $210 million in 2025 and $370 million in 2026, with a projection of $4.12 billion by 2036 and a compound annual growth rate of 27.3% (Meticulous Research, via GII Research, 2026). The report makes a blunter point than the forecasts: no answer engine offers a visibility interface comparable to Search Console. There is no place to log in and see which answers included a brand. That absence, more than any ranking algorithm, is what the new tools are selling.
The click data explains why. Semrush and Machine Relations examined nearly 69 million US desktop Google sessions and found that only 6–8% of AI Mode sessions continued to an external domain, leaving 92–94% with no outbound click; about 70% of the AI traffic that did arrive was filed under Direct in GA4 (Semrush / Machine Relations, 2026). Zero-click rates differ widely across studies, devices and definitions of a visit, so the figure worth keeping is this dataset's 6–8% rather than a universal constant. The practical consequence is stranger than a low number: a visit caused by an AI answer often arrives with no referrer, so it is counted as if someone typed the address.
Vendor-reported figures add to the sense of a market sprinting ahead of its measurement. Industry tallies put AI visibility tooling funding above $300 million between the summer of 2025 and the spring of 2026, with Profound alone raising $155 million at a $1 billion valuation. Those numbers come from vendor announcements carried by trade press, not audited statements, and they belong in the argument as a signal of demand — a caveat that matters when the same vendors sell the dashboards.
The temptation to fake precision
The real temptation is to dress SOAV in the clothes of an exact KPI. A dashboard reporting 4.2% to one decimal place looks rigorous while resting on a prompt list that somebody chose by hand. Resample the prompts next week and the number moves, because the answers underneath were sampled too. Teams that set a quarterly SOAV target borrow the authority of SEO metrics without the stability underneath. The defensible use is directional: hold the prompt set fixed, watch the band rather than the point, and pair it with brand surveys asking where people first encountered the brand. Correlation is the claim available here — exposure and demand moving together — not that a mention caused a sale. Accepting that blur is more rigorous than a false decimal.
That is the case for building observability into the catalog instead of bolting it on afterward. An open catalog that publishes a canonical, machine-readable entry for each mini-program gives the measurement problem a fixed object: the same description can be sampled by an assistant, cited inside an answer, and counted in a SOAV prompt set. CrossMiniApp's shared runtime adds what a page cannot — a governed identity and a record of which host surface actually executed the program — so the visibility signal and the execution signal refer to the same thing. Precision will not improve by pretending the click exists. It improves by making the recommended thing legible and by being honest about how much of the path stays unseen.