Measuring AI Overview Traffic Loss

Four measurable quantities, and three that are not

Before any of this becomes a damages input, be precise about what "AI Overview traffic loss" is asking. Four quantities can be measured with current data, each from a different source, each with its own foundation.

1. Whether and how often AI answer features appear for a defined query set, over time. Measured by dated capture of results pages for that query set, from fixed locations and devices.
2. Whether the claimant's own URLs appear inside AI surfaces. Measured from Google's generative AI performance reporting, at page level, for periods after that reporting existed.
3. The claimant's clicks and impressions by query cohort, before and after. Measured from first-party Search Console exports.
4. Market-level click behavior. Measured by published panel studies, which describe a market and not a party.

Three quantities that are routinely asserted cannot be measured from available data: an AI-feature click-through rate computed from Google's own reporting; the split between AI Overviews and AI Mode in a party's data; and query-level AI presence taken from Search Console rather than from independent capture. An opinion that states any of the three has produced a number the sources do not contain, and the number is the first thing the other side will trace.

The source inventory, and what each source can carry

Search Console. First-party, authoritative for the claimant's own impressions, clicks, click-through rate, and average position, with a query dimension and a page dimension. It is limited by a rolling sixteen-month retention window, by row sampling on large query sets, and by the fact that AI feature impressions are counted additively inside ordinary web-search totals rather than carved out of them. Subtracting one from the other double-counts, and the error is invisible in a summary chart.

Analytics. Sessions, engagement, and conversions, joined to landing pages. It cannot tell you which query produced the visit, and it therefore cannot support a query-cohort design on its own.

Server and CDN logs. Useful for crawl behavior and for confirming that a page was served, and often the only surviving record of an early period. They do not identify AI provenance: a user who clicks a link inside an AI answer arrives with an ordinary Google referrer, or with none at all under referrer-policy restrictions. Any assertion that logs distinguish AI-referred visits from ordinary organic visits should be tested closely, because in the general case they do not.

Results-page capture. Either the party's own dated screenshots and HTML captures or a vendor's tracking panel. This is the only source that establishes whether an AI answer appeared for a specific query on a specific date, and it carries the authentication questions that go with any web capture: who captured it, from what location, on what device, at what frequency, and whether the capture was contemporaneous or reconstructed after the fact.

Published studies. Corroboration for a market-level proposition, and nothing more.

Building the query cohorts

The strongest available design compares the claimant's own queries that trigger AI answers against its own queries that do not. Building those cohorts is the substance of the work.

Take the claimant's queries from a pre-period export, ranked by impressions, and cut the list where volume stops supporting a comparison. Capture the results page for each query on a schedule, holding location, device, and language constant, and record the capture date and method for every observation. Then classify each query as AI-triggering, non-triggering, or intermittent.

The intermittent bucket is where honest analysis lives, and it is larger than most reports admit. AI answer presence is not a fixed property of a query. It varies over time, it varies by location and device, it can vary between users, and a query that returned an AI answer in March may not in June. Three consequences follow. Classification must be time-stamped rather than treated as a label. A single capture per query is a snapshot, not a measurement of exposure. And any query whose classification changes mid-window either moves into the intermittent cohort or is excluded, with the rule stated.

Match the two cohorts on intent, pre-period impression volume, and topic, exactly as with any other control design, and publish the query lists. If the cohort membership cannot be reproduced from the report, the comparison cannot be tested.

The comparison, and precisely what it estimates

With cohorts defined, the comparison is a difference-in-differences: the change in clicks for the AI-exposed cohort from baseline to damage period, less the change in clicks for the matched unexposed cohort over the same window, on the same site.

State exactly what that estimates, because the gap between the estimate and the claim is where these opinions fail. It estimates the relative change in click volume for this site's AI-exposed queries against its own matched unexposed queries over a stated window. It does not estimate the causal effect of AI Overviews in general. It does not separate AI Overviews from AI Mode. It does not capture users who never searched at all because they asked an assistant outside Google, which is a substitution that leaves no trace in any data either party holds.

The confounds are concurrent and real. Ranking updates ran in essentially every quarter in which AI features expanded, so there is no clean control period — a stronger statement than saying the analysis is difficult. Result-page layout changed for reasons unrelated to AI. Query phrasing shifted as users adapted to generated answers, which changes which pages are eligible at all. The claimant's own migrations, redesigns, paywalls, and media spend moved in the same window. Each of these is either controlled for by the design, addressed separately, or disclosed as uncontrolled.

The published statistics, and the exact form each should take

Three measurements have documented methodologies. Everything else circulating on this subject should be treated as unsupported until the underlying study is produced, and at least one widely repeated click-loss figure larger than the one below has no reachable methodology at all. Do not use a number you cannot trace to a published method, however often you have seen it quoted.

Zero-click share. SparkToro's analysis of Similarweb clickstream panel data found that 68.01% of United States Google searches ended without a click between January and April 2026. Present it as a level for that stated window, not as a change over time: SparkToro states that comparisons against earlier years are not directly equivalent because the underlying data providers changed. The panel covers desktop and mobile browser sessions, weighted on an assumed two-to-one mobile-to-desktop ratio, with sample size undisclosed — and it excludes the Google mobile app, so SparkToro's own view is that the true rate is likely higher rather than lower.

Click-through impact. Ahrefs reports that AI Overviews are associated with a 34.5% lower click-through rate for the top-ranking organic result, from 300,000 informational-intent keywords split evenly between those with and without AI Overviews, using aggregated Search Console desktop click-through at position one and comparing March 2024 against March 2025, published 17 April 2025. Ahrefs states the analysis is correlational rather than causal, that Search Console does not expose AI Overview click-through directly, and that it expects the figure to be the highest the rate will be.

Independent corroboration. The Pew Research Center found users clicked a traditional result link in 8% of Google visits where an AI summary appeared, against 15% where none did, and clicked a link inside the summary in only 1% of visits. The panel was 900 United States adults sharing browsing activity through March 2025, covering 68,879 searches of which 12,593 produced an AI summary. Pew publishes its method. Google has publicly disputed the study, and a report that says so is stronger for it.

The 702(d) line, drawn in this specific analysis

Federal Rule of Evidence 702(d), as amended effective 1 December 2023, asks whether the opinion reflects a reliable application of the principles and methods to the facts of the case. It targets overstatement, and this subject invites overstatement more than any other in the field.

Here is where the line falls. Defensible: that AI answers appeared for a stated proportion of the claimant's query set on stated dates; that the exposed cohort's clicks fell by a stated proportion more than the matched unexposed cohort's over a stated window; that a stated share of the claimant's URLs appeared in AI surfaces once such reporting existed; and that market-level studies report declines of a stated magnitude with stated methods and limits.

Not defensible: a percentage of the claimant's total traffic loss attributed to AI Overviews specifically; an AI-feature click-through rate presented as measured; a market average multiplied by the party's revenue and offered as damages. That last one is arithmetic, not evidence about this party, and it is the most common error in the category.

Query-cohort attribution is genuinely hard. The exposure classification is unstable, the confounds are simultaneous, and the substitution channel is invisible. An expert who claims more precision than those conditions permit is making the same mistake this section spends four pages describing — and doing it on the subject where the mistake is easiest to demonstrate, because the opposing expert can re-capture the same queries and show the classification drifting.

The defense version, and the preservation clock

A defendant facing this claim has an unusually strong toolkit, because the measurement is fragile in ways that are demonstrable rather than argumentative.

Ask for the query lists and the raw capture files, with timestamps, locations, and devices. Ask whether the AI-presence classification was contemporaneous or reconstructed after the claim arose — reconstruction is not disqualifying, but it has to be disclosed and it changes what the classification proves. Re-capture a sample of the queries and show how many have changed cohort. Ask whether AI impressions were subtracted from web-search totals, since they are additive. Ask what the analysis did with the claimant's own site changes in the window. And ask the direct question: what portion of the decline does the analysis leave unexplained.

The preservation point cuts on both sides and it is urgent rather than theoretical. Search Console performance data is a rolling sixteen-month window; results-page appearance for a given query on a given date is not retained anywhere by anyone unless someone captured it; and analytics retention settings expire by default. The evidence in this category is perishable, and by the time a complaint is drafted the most probative period has often already passed out of the first-party record. Where a client's dispute is heading toward a claim of this kind, preservation is a step for counsel to take at the outset, not after the first case management conference.

Frequently Asked Questions

Published 2026-08-11