- AI answers compress awareness, research and shortlisting into one conversation — stages that used to generate three to ten site visits now generate zero.
- The click data is unambiguous: with an AI summary present, users click a traditional result on about 8% of visits vs. 15% without one,1 and Gartner projected a 25% drop in traditional search volume by 2026.2
- What still reaches you is smaller but better: late-funnel, pre-sold visitors who arrive with a shortlist already made — assistant referrals are widely observed to convert unusually well.
- The operational shift: measure presence in answers, not clicks, and treat the AI answer as your de facto landing page — because for most buyers, it is.
Run this experiment: ask ChatGPT or Perplexity to help you choose a project management tool for a 12-person agency. Watch what happens in the next four minutes. The assistant names candidates, compares pricing tiers, flags a limitation of one option, answers your follow-up about integrations, and produces a two-item shortlist.
Count the websites you visited. For most people, the answer is zero.
That four-minute conversation replaced what used to be an afternoon of funnel activity: a "best project management tools" listicle (awareness), three vendor feature pages (research), two pricing pages and a comparison post (shortlisting). Every one of those steps used to be a measurable touch — a session in your analytics, a retargeting cookie, a lead-scoring signal. Now they happen inside a chat window you cannot see, run by a model you don't control.
This is the zero-click buyer journey. It doesn't mean buyers stopped buying, or even that they stopped visiting websites. It means the selection happens before the visit. And it changes what marketing has to measure and optimize.
Three funnel stages, one conversation
The classic B2B or considered-purchase funnel assumed the buyer assembled their own picture from fragments: your blog post here, a review site there, a competitor's comparison page, a pricing page. Each fragment was a click, and each click was a chance to persuade, capture an email, or set a cookie.
A conversational assistant collapses that assembly work. The model has already read the fragments — or retrieves them at answer time — and hands back the synthesis. The buyer's job shrinks from "research the category" to "interrogate the synthesis": which of these handles EU data residency? which is cheaper at 50 seats? Follow-up questions that once meant more site visits are now just more turns in the same conversation.
What the click data actually shows
This isn't a thought experiment; the measurement is in. Pew Research Center tracked real browsing behavior from nearly 900 U.S. adults and found that when a Google search returned an AI summary, users clicked a traditional result link on about 8% of visits — versus 15% when no summary appeared. Clicks on sources inside the AI summary happened on roughly 1% of visits.1 The citation link is, behaviorally, a footnote almost nobody follows.
Ahrefs, studying keywords before and after AI Overviews arrived, put the cost to the winners at roughly a 34.5% lower click-through rate for the #1 organic position when an AI Overview is present.3 And that's just AI features inside classic search — it doesn't count the journeys that begin and end in ChatGPT, Claude or Perplexity and never touch a results page at all. Gartner's projection that traditional search engine volume would drop 25% by 2026 was, at the time, treated as aggressive; it now reads as a reasonable description of the direction of travel.2
Read the 1% figure carefully. It's tempting to conclude "citations don't matter." The opposite is true: if almost nobody clicks through to verify, then whatever the answer says about you is the entire brand experience for that buyer. The citation isn't a traffic channel — it's a trust signal to the model and the few humans who check.
What still reaches your site — and why it's good
Here's the part the doom-laden takes miss: the traffic that survives compression is disproportionately valuable.
Think about who still clicks through. Not the browser at the top of the funnel — the assistant satisfied them fully. The click-through is the buyer who made it through the synthesis, kept your brand on the shortlist, and now needs something only you can provide: a demo, a trial, exact pricing for their seat count, a procurement contact. They arrive late-funnel and pre-sold — the assistant has already made your pitch, answered the objections, and eliminated your weaker competitors.
Teams watching their analytics have widely observed the same pattern we see across MentionBeat customers: referral sessions from assistants are a small slice of total traffic but convert at multiples of the search average, with fewer pages per session and shorter time-to-signup. That's not a paradox. It's what you'd expect when the middle of the journey happens elsewhere: the visit isn't research anymore, it's execution.
- Fewer visits, higher intent. The assistant filtered out the tire-kickers for you.
- Shorter on-site journeys. Pre-sold visitors skip the education content and head for pricing, trial, or contact.
- Less persuasion needed — but zero tolerance for mismatch. If your site contradicts what the assistant said (price, feature, positioning), you break trust at the worst possible moment.
The dark funnel just got a lot darker
"Dark funnel" used to describe word-of-mouth, Slack communities, podcasts — influence you couldn't attribute. AI conversations are the same phenomenon at industrial scale, with a twist: the influencing agent is a system whose outputs you can actually sample.
The attribution symptoms are familiar by now. Direct traffic creeps up as buyers hear a name in a conversation and type the domain later. "How did you hear about us?" surveys start returning "ChatGPT told me" — arguably the most honest attribution data many teams now collect. Branded search grows while non-branded informational queries decay, because the informational intent got absorbed upstream. Last-click models credit the final navigation and miss the conversation that did all the work.
None of your existing tools see inside the conversation. Your analytics start at the click; the journey now starts three stages earlier. Which leads to the real strategic question: if you can't instrument the journey, what can you instrument?
Measure presence, not clicks
You can't put a pixel in ChatGPT. But you can do what the model does: ask the questions your buyers ask, at scale, and observe the answers. The measurable layer moves from behavior on your site to representation in the answer. The old dashboard maps to a new one:
| Search-era metric | Answer-era counterpart | What it tells you |
|---|---|---|
| Impressions | Mention rate across a buyer-prompt suite | How often you appear in answers at all |
| Ranking position | Recommendation rate and position within the answer | Whether you're endorsed or merely listed |
| Share of SERP | Share of voice vs. named competitors | Your slice of the shortlist across the category |
| Landing page quality score | Accuracy of representation | Whether the answer's claims about you are true |
| CTR | Citation rate in browsing engines | Whether your pages are the retrieved source |
Two properties make this harder than rank tracking. Answers are stochastic — the same prompt yields different brand sets run to run, so single samples are noise and you need repeated runs with confidence intervals. And they're engine-specific — ChatGPT, Gemini, Claude and Perplexity retrieve from different indexes and train on different corpora, so presence in one says little about the others. This is exactly the kind of repeated cross-engine sampling a platform like MentionBeat automates, but the methodology matters more than the tooling: suite of real buyer prompts, multiple runs, intervals, tracked over time.
MentionBeat runs real buyer prompts across ChatGPT, Claude, Gemini and Perplexity and shows your mention rate, share of voice and how accurately you're represented — the journey your analytics can't see.
Get a free visibility reportThe answer is the landing page now
If a buyer's first substantive encounter with your brand is three sentences inside an AI answer, those three sentences are doing the job your homepage used to do. You'd never ship a landing page with a wrong price or a discontinued product on it — but many brands are effectively shipping exactly that, every day, inside answers they've never read.
Optimizing for accurate representation is therefore not a nice-to-have; it's conversion rate optimization for a page you don't host. The levers:
- Audit what's being said, not just whether. Sample answers about your brand and score them for factual accuracy: pricing, feature claims, positioning, comparisons.
- Make your canonical facts effortless to retrieve. Current pricing, specs and positioning in clean HTML with schema markup — not locked in PDFs or JavaScript — so browsing engines quote the current truth, not a cached 2024 version.
- Fix the highest-damage errors first. A wrong price or "doesn't integrate with X" claim in answers costs more than a missed mention, because it disqualifies you silently.
- Align the post-click page with the answer. Pre-sold visitors land deep. Make sure pricing and claims match what assistants are currently saying, and clear the path to trial/contact.
- Re-sample after every meaningful change. Retrieval-based engines can pick up corrections in days; parametric knowledge shifts with model updates. Track both loops.
And keep some perspective on the losses. The clicks that disappeared were mostly the expensive ones: informational visitors you paid to educate, most of whom never bought. The zero-click journey outsources your top-of-funnel education to the assistant. Your job is to make sure the education is accurate and that you're in the shortlist when it ends.
Frequently asked questions
No — reframe what it's for. That content's audience is increasingly the model, not the human. Well-structured, well-sourced educational content is the raw material assistants synthesize from, which is how you get into the answer. What changes is the success metric: judge it by presence and accuracy in answers, not by sessions.
Triangulate. Combine self-reported attribution ("how did you hear about us?" with an explicit AI-assistant option), assistant referral traffic where it exists, movement in branded/direct traffic, and your measured mention rate over time. No single signal is airtight; together they're decision-grade. Treat mention rate and share of voice as the leading indicators they are.
If anything it bites harder in B2B, where the "research phase" was longest and most measurable. Technical buyers use assistants to build shortlists before any vendor knows they're in-market — a deal you never got to compete for. The upside: niche B2B categories have thin source material, so a well-documented brand can dominate the answers. See our B2B GEO playbook.
Sources & further reading
- Pew Research Center — "Google users are less likely to click on links when an AI summary appears in the results", July 2025.
- Gartner — "Gartner Predicts Search Engine Volume Will Drop 25% by 2026, Due to AI Chatbots and Other Virtual Agents", February 2024.
- Ahrefs — "AI Overviews Reduce Clicks by 34.5%", 2025.
- Aggarwal, P., et al. — "GEO: Generative Engine Optimization", KDD 2024 / arXiv:2311.09735 — on which content changes measurably move visibility inside generative answers.