Attribution and What It Cannot Tell You

The gap is the story

Attribution promises a clean sentence: this much revenue came from this channel. Everything about how search actually works in 2026 makes that sentence less true than it was, and pretending otherwise leads to real money being moved in the wrong direction.

Start with the scale of what never becomes measurable at all. SparkToro's analysis of Similarweb clickstream data found that 68.01% of US Google searches ended without a click in January to April 2026. Two caveats travel with that figure and should never be separated from it: the panel excludes the Google mobile search application, covering mobile searches within the browser instead — and SparkToro notes zero-click behavior is likely even more common inside the app — and long-term comparisons are not like-for-like, because the multi-year trend combines different panel providers whose users and devices are not the same population.

Read carefully, that is a statement about a measured window rather than a trajectory, and it is enough on its own. Most search activity produces no session, so it produces no row in your analytics. The gap between what happened and what you can measure is not a reporting failure to be fixed with a better tool. It is the condition you are working in.

Where measurement actually breaks, in order

It is worth being specific about the failure points, because they are cumulative and each one is invisible in the report that follows it.

1. The search that never became a visit. Answered on the results page, in a feature, or in an AI Overview. Search Console records the impression; nothing else records anything.
2. The visit that was never counted. Consent controls and tracking prevention mean measured sessions understate real sessions. The size of the undercount varies by audience, region and browser mix, and it is not constant over time — which means year-on-year comparisons in analytics can move because measurement changed rather than because behavior did.
3. The visit that was counted and misfiled. Someone arrives from a search, leaves, returns later by typing your name, and converts. Last-click credits direct or brand search. The original search created the demand and receives none of the credit.
4. The conversion that happened somewhere else. A phone call, a store visit, a reply to a sales email, a purchase from a different device. Cross-device and offline paths break the chain entirely, and they are more common in considered purchases, which are usually the valuable ones.

Each stage is a filter. What lands in your report is what survived all four, and it is systematically biased toward short, single-device, recently-clicked journeys.

Last click is a filing convention, not a finding

Last-click attribution answers exactly one question: which touchpoint immediately preceded the conversion. That is a real question, and it is not the question anyone actually asks in a budget meeting.

Its consistent bias is to over-credit whatever sits closest to the transaction — brand search, direct, retargeting — and to under-credit whatever created the demand in the first place, which is usually organic content, informational queries, and everything upstream. Cut the under-credited channel and the over-credited one keeps performing for a while, on demand it did not create. By the time the effect shows up, several months of reporting have already blessed the decision.

Multi-touch models redistribute credit across touchpoints and are better at describing the path, but they inherit every measurement gap above: they can only model journeys they observed, and the unobserved ones are not random. A model built on the visible subset is a confident description of a biased sample.

The practical position: use whatever model you use consistently, name it in the report, and treat the outputs as a directional comparison over time rather than a statement of truth. What matters is that everyone in the room knows the number is a convention, so that nobody presents a change in the convention as a change in performance.

What AI surfaces added to the problem

Two specific things changed, and it helps to keep them separate from the general anxiety.

Measured click impact. Ahrefs compared informational keywords with and without an AI Overview present and found a 34.5% lower average click-through rate for the top-ranking page where one appeared. The methodology should travel with the figure: 300,000 keywords, half with an AI Overview and half without, aggregated desktop Search Console click-through data, comparing March 2024 with March 2025. The authors also note that multiple citations within one AI Overview dilute clicks to any single link, and that Google provides no way to isolate AI Overview performance.

Reporting. Since 3 June 2026, Search Console reports generative AI impressions, the pages that surfaced, countries, devices and dates — but no click data, and with AI Overviews and AI Mode combined into one category. Appearance is measurable; the value of the appearance is not.

Worth carrying alongside both, because it complicates the panic usefully: Semrush's study of more than ten million keywords found AI Overview trigger rates peaked and then fell — 6.49% in January 2025, 24.61% in July 2025, 15.69% in November 2025 — and, tracked before and after on the same keywords, the zero-click rate decreased slightly rather than rising. These measure different things than the SparkToro figure, and a report that carries both is more credible than one that carries only the alarming one.

The important discipline: none of these percentages is your percentage. They are population averages across query mixes that are nothing like yours. Use them to explain a direction, never to forecast your own numbers.

The instruments that still work, used together

The response to an unmeasurable gap is not a better attribution model. It is a small set of imperfect instruments that fail in different directions, read together.

• Branded search demand over time, from Search Console. If more people are searching your name, something upstream is working, even when nothing in your attribution report can say what. It is one of the few observable signals of demand creation you own.
• Self-reported attribution. A single optional field on the inquiry form asking how someone first heard of you. It is noisy, it is biased toward what people remember, and it routinely surfaces channels that analytics never credited. Used as a directional cross-check against the measured picture, it is one of the highest-value fields on any form.
• Matched-period comparison. Compare like windows, with dated annotations for releases, campaigns and confirmed updates. Crude, and it survives measurement changes that break session-level comparison.
• Holdouts, where they are practical. Deliberately withholding activity from a region, a segment or a set of templates, and comparing. It is the only method on this list that speaks to causation rather than correlation, and it costs something real, which is why it is rare and why it is worth the argument on decisions that are large enough.
• Search Console impressions by segment, which see the appearances that never became sessions and are therefore the closest available view of the part of the gap that analytics cannot reach.

What to refuse to believe

Some claims should be treated as evidence about the person making them.

AI traffic attribution. No vendor has click data for Google's AI features, because Google does not publish it. Anything reporting AI-driven traffic or conversions is a model presented as a measurement, and its confidence intervals are usually not shown because they would be embarrassing.

A single unified score. Any provider reporting one blended number for search performance is reporting their own tool's weighting of inputs, chosen by them. Two vendors will disagree on the same site.

Share-of-voice figures without their universe. These are computed against the vendor's own keyword set or prompt list. Change the set and the number changes. Ask what universe it was measured on before treating any movement in it as meaningful.

Precise forecasts of organic revenue. Forecasts built by multiplying search volume by an assumed click-through rate by an assumed conversion rate compound three uncertain numbers, and the first two are exactly the ones that zero-click behavior has made least stable. A range with stated assumptions is honest; a number is not.

Rank tracking deserves a specific note. It reports a position for a synthetic query from a chosen location and device, and results are personalized, localized and device-dependent. It also measures a diminishing thing: when most searches in a measured window end without a click, position one does not mean what it used to. Keep it as a diagnostic instrument, not an outcome metric.

Making decisions anyway

None of this is an argument for measuring less. It is an argument for making decisions in a way that survives the uncertainty, which is a different skill from producing a confident report.

Three habits do most of the work. Decide on direction and magnitude, not on decimal places. If the question is whether to continue a program, you need to know whether it is working and roughly how well, and that judgment can be made from imperfect instruments that agree with each other. Precision that does not change the decision is not worth buying.

Triangulate before acting on a surprise. When one instrument reports something dramatic, check whether the others agree. A collapse in analytics sessions with steady Search Console clicks is a tracking failure, not a business event — and that single check has saved more wasted weeks than any dashboard.

State the uncertainty in the report, in advance of needing it. One short paragraph naming what the data cannot see: the sessions that consent controls leave uncounted, the searches that never produced a click, the conversions that happened offline. Written in month one it is credibility. Written in month nine, after a bad quarter, it reads as an excuse — and the same true sentence is worth far less then.

The consultants and marketers who do well in this environment are the ones who were already honest about measurement before it got harder. The gap has widened; the discipline has not changed.

Frequently Asked Questions

Published 2026-08-11