How to measure AI-driven traffic and bookings
AI assistants deliver guests your analytics cannot see. The measurement problem is solvable — with the same modeled-conversion science the major ad platforms run, plus receipts they can’t offer.
Most AI-driven visits are invisible to conventional analytics: the assistant's agent reads your pages first, the guest arrives later with the referrer stripped, and the session files under Direct. Measuring the channel takes three instruments working together: deterministic receipts where a referrer or tagged link survives, calibrated probabilistic attribution where it does not, and aggregate incrementality measurement — the same modeled-conversion approach Meta and Google adopted when iOS privacy changes broke deterministic tracking. Kismet implements all three under Landfall, an open measurement standard we publish, so every number on a report is labeled by how it was measured.
Why doesn't AI traffic show up in Google Analytics?
When an assistant recommends a property, its agent fetched your pages while composing the answer — but the human guest arrives later, often through an in-app browser or a copy-pasted link, with the referrer stripped at the HTTP level. Published analyses in 2026 put the loss at 35 to 70 percent of AI referrals, which land in analytics as Direct. Native AI channels in analytics products cover only some assistants and only the referrer-bearing minority of their traffic.
The result: the faster the AI channel grows, the more of your “Direct” traffic is actually AI-delivered — and a standard analytics setup cannot tell you which part. Measuring the channel means measuring around the broken referrer, not waiting for it to come back.
How can you attribute a visit that has no referrer?
By anchoring on the event you can observe: the fetch. An AI answer starts with the assistant's agent retrieving your pages — an event a first-party edge can log with the agent's identity, the exact URL, and a timestamp. When a guest lands on that same page shortly afterward, referrer-less, on a deep page, as a new session, the timing against that page's own baseline arrival rate carries real statistical evidence. Kismet scores every such landing against the page's measured null: how often do landings like this happen by chance on this page? A landing whose timing clears a two-sigma bar on its own page's baseline is classed attributed; a weaker but still significant one is classed likely; a session carrying a literal AI referrer or tagged link is verified. Known non-AI causes — an email campaign click, a paid-ad click ID — always outrank the inference.
What happens when AI fetches your pages constantly?
Per-visit timing evidence fades — by design. When an assistant fetches a page hourly, every landing is close to some fetch, and closeness stops meaning anything. A method that kept claiming individual visits at that density would be manufacturing certainty. The honest answer is the one ad platforms reached years ago for the same problem: measure the aggregate. If AI delivery is real, arrivals on fetched pages run measurably above the page's baseline rate — an arrival lift with a confidence interval, computed from the same event logs. Individual receipts still surface where they exist; the aggregate lift carries the channel's scale. Landfall formalizes this handoff so the methodology stays valid at any fetch volume.
Are modeled conversions legitimate, or made-up numbers?
They are the industry-standard answer to broken observation. When Apple’s App Tracking Transparency and related privacy changes broke deterministic conversion tracking, Meta and Google both moved to conversion modeling: statistical models estimate the conversions that can no longer be observed directly, and the modeled portion is blended into standard reporting columns. Google’s own documentation describes modeled conversions in Google Ads and behavioral modeling in GA4, included only at high confidence and suppressed when data is insufficient.
Landfall adopts the same discipline with two additions the platforms do not offer: the modeled total is capped by an independently measured incrementality bound, and the model is recalibrated monthly against deterministic ground truth — sessions where a receipt survived. The uncertainty is also printed on the report rather than hidden. Same readability, more honesty.
How do you know the model isn't overclaiming?
Because a second, independent measurement bounds it. For every page an AI agent fetches, the hours shortly after a fetch are exposure windows and all other hours are baseline. If AI delivers nobody, the arrival rates in the two are identical; the excess of in-window arrivals over the baseline rate is the channel’s incremental contribution, with a confidence interval from the arrival statistics. This baseline-and-lift design is long-established measurement science — the same family as the Bayesian structural time-series counterfactuals published by Google research and used across the industry for hard-to-attribute channels.
Landfall makes the bound normative: the sum of modeled per-session attributions must reconcile to the incrementality interval every reporting period. If it ever exceeds the bound, the model is recalibrated or capped, and the correction is disclosed. No sum of per-visit claims can outrun the aggregate evidence.
What does an AI performance report actually show?
Three kinds of numbers, each labeled by how it was measured. Delivered: visits and bookings with deterministic receipts — a surviving AI referrer, a tagged link, a minted token. Estimated: the modeled contribution, presented as a headline figure with its range and confidence printed in the report’s fine print. Lift: the incremental visits measured against baseline, as an interval.
Individual bookings appear as evidence-graded exhibits — verified, attributed, likely — so a manager can click from any line to the underlying guest journey. The report footer names the tracking version and links the full methodology, because a number you cannot audit is a number you cannot defend.
Read the standard
Landfall is the open specification behind every Kismet AI performance report: the three tiers, the reconciliation bound, the calibration requirement, and the reporting language. Published for review, implementation, and criticism.
