Marketing Attribution Broke Again. 68% of Searches End Without a Click

68.01% of US Google searches ended without any click in the first four months of 2026, according to clickstream panel data. Buyers now research in AI answers, private channels and closed group chats, none of which write a row into your analytics. The dashboards did not go blank. They kept producing confident percentages against a shrinking share of the actual journey, and that confidence is the problem.

What dark discovery actually means for measurement

Attribution broke because the research now happens where tracking does not. Buyers read AI answers, ask peers in private channels, and arrive at your site with no referrer. The fix is not a better tracking script. It is a mix of self-reported attribution, holdout tests and modelled channel effects.

Dark discovery is the umbrella term for every moment of awareness that leaves no server-side trace. That includes a Slack message, a podcast, a WhatsApp forward, a colleague's recommendation and an answer generated inside a chat interface.

None of these are new categories of human behaviour. What changed is the proportion. A journey that used to leak a few untracked touches now leaks most of them, and the tracked remainder is no longer a representative sample of the whole.

That distinction matters more than the missing data itself. Analytics can survive missing rows. It cannot survive missing rows that are systematically different from the rows it kept.

Three numbers that describe the size of the gap

The argument only works if the gap is large and growing. Here is the evidence I would put in front of a board, with its limitations attached.

Search stopped sending clicks

SparkToro's 2026 analysis of Similarweb clickstream data put the US zero-click rate at 68.01% for January to April 2026, against 60.45% in 2024 and roughly 50% in 2019. Clicks to organic results, paid ads and Google properties fell by a combined 9.51 percentage points over those two years.

The same analysis reported that AI Overviews appear on more than 20% of searches, and that click-through rates drop by close to 60% when they do. AI Mode was a rounding error during the window at 0.34% of searches, which tells you the trend is not yet driven by the feature everyone is writing about.

The buyer arrives with a shortlist you did not see form

6sense surveyed nearly 4,000 buyers for its 2025 B2B Buyer Experience Report, at a median purchase value between $200,000 and $300,000. First vendor contact happened at 61% of the way through the journey, earlier than the 69% it recorded in 2024, on an average cycle of 10.1 months.

Buyers evaluated 5.1 vendors on average and bought from their day-one shortlist 95% of the time. 94% said they used large language models during the process. The composition of that shortlist is examined separately in the piece on how AI assistants build a B2B shortlist.

Read those two findings together and the conclusion is uncomfortable. The event your funnel treats as the beginning of the relationship is closer to the end of the decision.

Why the dashboard still reports a neat number

Missing data does not make a report empty. It makes the surviving data look more important than it is, and every default in the stack pushes in that direction.

Last touch always finds a last touch

A last-touch model cannot return "unknown". It assigns credit to whatever is nearest the conversion, because that is the only thing it can see. If 60% of the journey is invisible, the model does not report 60% uncertainty. It reports 100% certainty about the 40% it kept.

Multi-touch models are worse here, not better. They inherit the same visibility problem and then distribute fabricated fractional credit across it, which produces decimal places that look like precision. My position is that a decimal place on an attribution report is a warning sign rather than a feature.

Referrer stripping compounds it. Links opened from AI assistant mobile apps often reach your server with no referrer header, and those sessions land in the direct bucket alongside genuine bookmark traffic. A rising direct line is currently as likely to mean lost provenance as it is to mean brand strength.

The last-touch pattern also shapes what gets funded. Channels close to the form fill keep proving themselves, and channels that create demand months earlier keep failing to. That feedback loop is discussed in the analysis of what saturated AI content did to search visibility.

GA4's AI Assistant channel is real, and narrower than the announcement

Google added an AI Assistant channel to GA4's default channel group on 13 May 2026. It requires no configuration. When Analytics recognises an AI referrer, it sets the medium to ai-assistant and groups the session in default channel group reports.

This is a genuine improvement and it does not close the gap. Three limits are worth writing on the wall before anyone presents an AI channel chart.

First, coverage is a named list rather than a rule. Google's default channel group documentation defines the channel as arrivals "from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok". Google's own launch note described chatbots including Claude. Two Google pages, two different lists, and no published method for how a new assistant joins.

Second, the classification depends on a referrer arriving at all. Sessions that lost their referrer in an in-app browser handoff are still direct traffic, whatever channel groups exist downstream.

Third, it is not retroactive. GA4 processes data daily, so every AI session before mid-May 2026 stays classified the way it was classified then. Any year-on-year comparison of this channel measures the reporting change as much as the behaviour.

Search Console has the same shape of limitation. Its generative AI performance report covers AI Overviews and AI Mode together, and Google's help page states that it is rolling out "to a subset of website owners". The one split most teams want, AI Overviews against AI Mode, is not available. What the report can and cannot prove about citations is covered in the piece on AI citations against classic search rank.

Four methods that survive dark discovery

Everything below gives up individual-level truth. That is the trade. You stop asking which touch caused this deal and start asking whether the channel moves the aggregate.

1. Self-reported attribution

Add an open text field asking how the buyer heard about you, on the demo form and again in the sales call. It is the only instrument that reaches inside a private Slack channel, because the person who was in that channel is the one filling it in.

Keep it open text rather than a dropdown. A dropdown offers the answers you already believe, and a buyer picking the least wrong option is noise you cannot detect later.

2. Holdout and geo incrementality tests

Turn a channel off in a matched set of regions or accounts, leave it on elsewhere, and compare pipeline. This is the only method in the list that produces causal evidence rather than correlation.

Google's Meridian GeoX, announced at Google Marketing Live 2026, packages geo experiments as open source and feeds the results back into a mix model as calibration. The method predates the tool by decades. What is new is that a small team can now run it without buying a measurement platform.

3. Marketing mix modelling

Mix modelling regresses aggregate outcomes against aggregate spend and external factors, with no user-level tracking at all. That is precisely why it is resistant to cookie loss, referrer stripping and dark social.

Google open-sourced Meridian on 29 January 2025 after testing with hundreds of brands globally, and it accepts incrementality experiment results as priors. The honest caveat is data hunger. A model needs a long history of varied spend, which most B2B companies do not have.

4. Declared baselines and branded demand

Record what you believe before you spend, then check the aggregate signals that dark discovery still moves. Branded search volume, direct sessions to deep pages, and unaided mentions in sales calls all respond to awareness the funnel cannot see.

None of these prove causation on their own. Used as a pre-registered prediction rather than a post-hoc explanation, they are considerably harder to fool yourself with. That discipline is the same one applied to vendor claims in the piece on verifying outcomes in case studies.

What each method can and cannot answer

Choose by the question you actually need answered, not by which tool your stack already bills you for.

The first row is included because it is what most teams still report. It is not a recommendation. The bottom three are complements rather than alternatives, and the strongest setup runs a holdout to calibrate a model and a self-reported field to sanity check both.

Where this argument is weakest

Three objections are strong enough that I would raise them myself before a sceptical CFO did.

The first is nostalgia. Attribution was never as good as people now remember. Cross-device gaps, cookie deletion, offline conversations and word of mouth were breaking these models long before AI answers existed. Framing 2026 as the year measurement broke overstates a discontinuity in what is really a long decline.

The evidence base has real limits

The zero-click figure comes from one vendor's panel, and panels have coverage bias by construction. It also measures searches, not buyers, so a single research session with many refinements is counted many times. The 6sense figures come from a survey run by a company selling software to solve the problem the survey describes, which is worth holding in mind even though the sample is large and the method is disclosed.

Self-reported attribution has its own well-known failure. Buyers name the channel they remember, which favours recent and high-salience touches, and they sometimes name the answer that flatters them. It is directional evidence, not a ledger. The counter-case is that being roughly right about a whole journey beats being precisely wrong about a fragment of one, and I hold that view but cannot prove it from public data.

The last objection is the practical one. Mix modelling and holdout testing both need scale. A company with 40 deals a quarter cannot detect a 10% lift, and any model fitted to that history will fit noise. For those teams, self-reported attribution and honest qualitative work are not a stepping stone toward better measurement. They are the measurement.

What to stop reporting this quarter

Removing a misleading number is faster than building a better one, and it changes decisions immediately.

One more, and it is the least popular. Stop reporting a single number for marketing-sourced pipeline. The number is a modelling choice presented as a fact, and every board that receives it treats it as an audited figure. Replace it with a range and the assumption that produced each end. How much of the reported gap between measured and modelled revenue is real is examined in the analysis of AI referral traffic and its conversion multiple.

Frequently asked questions

Where to start this week

Two changes, both cheap enough that nobody needs to approve a budget.

Add an open text field asking how the buyer first heard about you, on your demo form and in the discovery call script. Do not offer a dropdown. Give it one quarter and compare what buyers say against what your last-touch report says, then bring both to the same meeting.

Then annotate your analytics on 13 May 2026 and mark every AI channel comparison that crosses it. It takes ten minutes and it stops one specific bad decision, which is treating a reporting change as growth.

Related reading