Two studies worth more than the rest of the literature combined
Almost everything written about AI and the marketing funnel rests on vendor analytics. Two exceptions exist, and they are the backbone of this guide because they are causal rather than observational.
A preregistered randomised controlled trial (Wang et al., August 2026) assigned participants to search conditions. Under Google's AI Mode, external clicks fell 18.8 percentage points against control, significant at p<0.001. In the condition where AI Overviews were removed, external clicks rose 8.8 points. Daily sessions declined. Self-reported trust, satisfaction and sense of agency were all significantly lower, and use of competing search engines rose 11.2 points.
A causal clickstream study (Shi, Zhu & Gu, Bocconi, July 2026) used the staggered rollout of ChatGPT Search across more than 45,000 US households. Traditional search queries fell 9.4% immediately on access, growing to 17.0% below baseline by week twenty.
Insight
A ChatGPT conversation session produces an outbound click 5.2% of the time. A Google search query produces one 31.1% of the time. The study counts sessions for one and queries for the other, so the two figures are close to, but not exactly, a like-for-like comparison. The same study found the damage is wildly uneven by category: academic sites down 32.8%, reference and knowledge down 26.5%, developer and technical down 15.1%, news down 13.4%, marketplace and entertainment roughly flat.
Read those two together and the mechanism is clear. AI search does not redistribute clicks. It retains attention. The user gets the answer and the session ends.
The zero-click trend, attributed correctly
68.01% of US Google searches ended without a click in the first four months of 2026, up from 60.45% in 2024 and about 49% in 2019 — the fastest acceleration on record, and an undercount, since the panel excludes Google's mobile app where the effect is stronger.
Two clarifications most coverage skips. That number is about Google, not about assistants: it measures searches ending on the results page, which is evidence the results page increasingly answers the question itself — a different phenomenon from users leaving Google for ChatGPT.
The in-SERP effect is separately measured. Pew's analysis of roughly 900 US adults' browsing found traditional results were clicked on 8% of visits when an AI Overview was present against 15% when it was not. Links inside the summary itself were clicked on 1% of visits. Session abandonment ran 26% with an Overview against 16% without.
So there are two distinct forces, and conflating them produces bad strategy. Google's own surfaces are absorbing clicks that used to leave. Assistants are absorbing queries that used to happen on Google. The first is larger today. The second is growing faster.
The numbers that do not fit the headline
An article that only quotes the alarming figures is the article everyone else has already written. Here is the other half, and it changes what you should do.
And the conversion-quality story is unstable. Adobe’s April 2026 report put AI-referred retail traffic at 42% better converting than non-AI traffic in March 2026, against 38% worse in March 2025. A reversal that large in twelve months, from a vendor that also sells AI-optimisation tooling (Adobe LLM Optimizer, named in the same report), is not a settled fact about buyer behaviour. Other published multiples range from 1.2 to 4.4 times, with samples as small as fifteen client websites, and the highest figure has no disclosed sample, conversion definition or time period at all.
The defensible reading: AI-referred traffic is plausibly later-funnel and higher-intent, because it arrives after the research happened elsewhere. The specific multiple is rhetoric.
What genuinely breaks: attribution
This is where the practical damage lands, and it is under-discussed because there is no number to put on it.
Assistant traffic frequently arrives with no referrer header — mobile app sessions in particular — and lands in analytics as direct. ChatGPT appends a source parameter inconsistently: reliably on citation links, not on search-result links. Any analytics channel that depends on referrer matching therefore reports a floor on AI traffic, not a count of it.
Now combine that with the clickstream finding. If 94.8% of assistant sessions produce no outbound click at all, then the visitor who does eventually arrive — typed your URL, searched your brand name, clicked an email — was influenced by a conversation you will never see, and lands in your reporting as direct or branded search. The research happened. You just were not in the room, and your analytics has no field for it.
Insight
There is no fix, only detection heuristics, and none of them is validated at scale. Which leads to the uncomfortable operational point: if assistant influence is invisible in your analytics, the only way to know your position is to measure the assistants directly — asking them the questions your buyers ask and recording what they say. In this channel that is not a reporting workaround. It is the primary instrument.
What actually changes about the funnel
The classic funnel assumed the buyer traverses your surfaces: they find a page, read it, compare, convert. Each stage was observable because each stage was a session.
Evidence that consideration is moving into the conversation is real but thinner than the confident charts suggest. The strongest primary source is G2's April 2026 survey of 1,076 B2B decision-makers: 51% now start software research with an AI chatbot, up from 29% a year earlier. That is a self-report survey from a company with its own stake in AI visibility — but the sample is disclosed and the doubling in twelve months is hard to dismiss. Wider figures circulating in this space trace back to aggregations of aggregations and should not be repeated.
Awareness and consideration partly stop being yours to instrument. They happen in a conversation, and the first evidence you get is a visitor who already knows things.
The shortlist forms before the first visit. If the answer to “best tools for X” names three companies and you are not one, you did not lose the deal — you were never in it, and unlike page two of Google there is no scrolling to find you.
And traffic becomes a worse proxy for demand. Flat or declining sessions are now compatible with growing brand influence, and growing sessions are compatible with declining influence. The correlation that made traffic a usable KPI is weakening.
The strategic response
Do not act on the biggest numbers. The RCT and the clickstream study are worth more than every vendor chart combined, and both describe an effect concentrated in specific surfaces and specific content categories. If you publish reference or technical content, the measured declines are severe. If you sell into a category where marketplace behaviour dominates, they are close to zero. Find out which you are before reorganising anything.
Measure the channel directly, because analytics cannot. Fixed question set, multiple engines, re-fielded on a schedule. It is the only instrument that sees the part of the funnel that stopped producing sessions.
Change what you count. Sessions and rankings degrade as proxies. Mention rate on buying-intent questions, citation share by domain, and factual accuracy of what engines say about you are the measures that correspond to what is now happening.
And do not abandon what works. Google still drives roughly two orders of magnitude more traffic than ChatGPT, and 95% of assistant users still search. The strategic error available right now is not moving too slowly — it is reallocating away from a channel that still produces almost all your revenue on the strength of vendor statistics that reversed sign within a year.
The funnel is not destroyed. The observable part of it got smaller, and the part that decides the shortlist moved somewhere you have to go and look.
Wang, Gleason, Bart, Wilson & Metaxa (arXiv:2608.18352, 18 Aug 2026) · Shi, Zhu & Gu, Bocconi (arXiv:2607.07652, Jul 2026) · SparkToro with Similarweb (9 Jun 2026) · Pew Research Center via Search Engine Land (24 Jul 2025) · TechnologyChecker, Cloudflare Radar data (Aug 2026 data; page updated 1 Oct 2026) · Similarweb via Search Engine Journal (4 Sep 2026) · Alphabet Q2 2026 earnings release (22 Jul 2026) · Adobe Digital Insights (16 Apr 2026) · G2 B2B buyer research, n=1,076 (15 Apr 2026)