A traffic dashboard cannot tell the difference between demand disappearing and a click disappearing. Those are separate events with separate causes, and for most startups only the second is happening.
Ahrefs measured a 58% lower average clickthrough rate for the top-ranking page in its December 2025 re-run when an AI Overview sits above it, and Ahrefs found 38% of those overview citations rank in the organic top 10 as of March 2026, so most named sources are not the pages ranking beneath them. The visit was never the thing you wanted. Being considered was.
So the operating answer is short: when the answer resolves the question, presence inside it becomes the earlier and more decisive event, and the click is optional.
6sense looked at the traffic B2B sites are losing to LLMs: it says the missing visits are likely not in-market buyers but the professionally curious and future buyers. Gartner projected search engine volume would fall 25% by 2026 because of chatbots, a projection and not a measurement, while Similarweb's tracking shows ChatGPT homepage referrals rising sharply after the May 2026 update.
What are you actually losing when the click disappears?
You are losing an observation, not necessarily a buyer. The click was a free measurement artifact: someone considered you, and the browser wrote it down. When an assistant answers in place, the consideration still happens and the log entry does not.
That distinction decides where the budget goes. If demand moved, you have a product and positioning problem. If only the record moved, you have a measurement problem plus a presence problem, both cheaper to fix. In Profound's 7.5 million-conversation en-US sample, the share of ChatGPT conversations classified commercial rose from 13.9% to 19.2% in a year, so buying questions are being asked, just not on your property. Profound's "more than doubled" volume figure is an extrapolation from estimated weekly users.
Commercial conversation volume is not a purchase queue. A corpus of 8,133 multi-turn human-LLM conversations found exploration markers in 50.3% of commercial conversations and 44.8% of PRISM conversations, so much of that volume is research inside a system you do not instrument.
The competitive shape changes too. G2 found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers start research with an AI chatbot more often than with Google, and a third of buyers said an AI chatbot led them to choose a vendor they hadn't considered. Read that as a doorway: a vendor the buyer wasn't considering can still enter the set.
Which questions still deserve a page from a small team?
Fewer than your content calendar assumes, and they are the ones a buyer asks immediately before choosing. Broad informational pages are the category the answer absorbs, so publishing more of them buys less every quarter.
Pick them by working backwards from the decision: what does someone compare you against, what disqualifies a vendor, what does implementation cost in time, what breaks. 6sense found nearly four out of five vendors are named to the shortlist on day one, and its 2025 buyer experience report found nearly all of them are vendors the buyer already has experience with. So the pages worth writing are the ones buyers use to narrow that list.
Resist the tactical shortcuts; the research does not support them. C-SEO Bench, published at NeurIPS 2025, found most rewrite methods largely ineffective and often harmful to citation rank, and its public code and setup tested conversational-SEO methods across two tasks and six domains. Rewriting a page in an AI-friendly voice is mostly a no-op; choosing better questions is the work that pays.
One partial counterexample: the GEO study at KDD 2024 found adding statistics, quotations, and citations raised visibility most in its benchmark, while fluency and plain-language rewrites also gained, by less, and keyword stuffing did nothing. The methods that did best added evidence, though C-SEO Bench's retest found even those rarely moved citation rank.
What kind of evidence survives extraction?
Evidence that can be lifted out of your page and still be true. An assistant does not quote your argument, it quotes a sentence, so a claim that needs three surrounding paragraphs to stay accurate gets dropped or mangled.
Start with the mechanics: nothing downstream matters if the page cannot be read. Vercel's crawler research with MERJ documented that OpenAI's and Anthropic's crawlers do not execute JavaScript, while Google's Gemini leverages Googlebot's infrastructure, enabling full JavaScript rendering. Cloudflare's agent-readiness study across the 200,000 most visited domains measured how widely sites have adopted agent standards: 78% had a robots.txt, but only 4% declared AI usage preferences and 3.9% served Markdown to agents that asked for it. Fetch your pages with scripting disabled first.
Then make each claim self-contained: number, source, date, conditions. AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited in the final answer.
Classic retrievability still matters, and the surface decides how much. Seer's sample of 500 SearchGPT citations found 87% matched Bing's top results, so on that engine ranking sits upstream of citation. On Google AI Overviews the Ahrefs top-10 figure above shows a looser relationship that one number would hide.
Skip the persuasion reflexes. Scarcity and exclusivity framing measurably reduces how often a model recommends a product, tested on three sets of 10 fictitious products so no brand equity could carry the result. Much of the answer sits off your domain: more than 4 in 10 AI citations point to sources brands do not control.
How do you measure presence when sessions stop reporting it?
You measure appearance rates on a fixed set of questions, repeated over time, and the unit of measurement is a distribution. A single answer tells you almost nothing.
SparkToro and Gumshoe found AI engines are highly inconsistent when recommending brands, with the same prompt producing different vendor lists across runs, and Search Engine Land's write-up of the same study reported the same instability. A 2026 variance-components study found run-to-run noise large enough to swamp real differences in small samples, and work on quantifying uncertainty in AI visibility argues for confidence intervals.
Separate the rungs instead of averaging them. The IAB's AI-era visibility guidance breaks the problem into Presence, Prominence, Portrayal, and Persuasion, and its guidelines sort the work into 4 groups for brands. A mention is not a citation, a citation is not a recommendation, and a recommendation is not a shortlist slot. Collapsing those into one score destroys the only signal you had.
Keep the site-side numbers you still trust, but demote them. GA4's key event rate is sessions with a key event divided by total sessions, which stays meaningful for the people who arrive and silent about those who never needed to.
What can you still count when nobody clicks?
Two machine-side surfaces still emit numbers after the click stops: search impressions and crawler requests in your logs. Google defines an impression as your link appearing in a result the user sees, with position and clicks counted separately. Impressions cover only Google's surfaces, and since August 31, 2026 they separate its AI features too: Search Console's Generative AI performance report (Search Console Help, read October 9, 2026) counts how often links to your pages were shown in AI Overviews and AI Mode, by page, country, device and date. It shows no clicks or queries, and Google says a site without enough impressions in AI features usually won't see it. So the report tells you a link to you was shown in an answer, not what the answer said or whether anyone clicked.
Your own click data will move the same way. Pew found users who saw an AI summary clicked a traditional result in 8% of visits against 15% without one, from metered browsing by 900 US adults. Those figures come from observed browsing, so they support an association and stop there.
Server logs are the easier surface to overread. Cloudflare's bot report put AI training at 52% of crawler requests in that sample, which measures access rather than attention, and a crawl is not an audience. Outside indexes read the same way: Semrush's AI Visibility Index reports at category level.
What do you do about the part of the funnel you cannot see?
State it as an unobservable region with a stated size, and stop filling it with invented attribution. The worst outcome of the click decline is not lost traffic, it is a board deck crediting whichever channel still emits a parameter.
The honest reporting move is one line item: early research now happens inside assistants we do not instrument; here is measured presence on the questions that matter, its run-to-run variance, and the direct and self-reported attribution we treat as its downstream shadow. Back-fitting pipeline to a visibility score is not defensible.
Ask sales the question no dashboard answers: what did the buyer already believe on the first call, and where did they get it.
How does Trovance help when the answer replaces the visit?
Trovance treats presence in the answer as the measured event. You define the buyer questions your market asks, and it sweeps them across AI engines every week as tracked questions, preserving each answer run as a snapshot: who was mentioned, who was cited, who was recommended, and which sources carried the answer. Share of answer over time is the reading.
A fixed question set is a proxy. Real buyers arrive mid-conversation with context you did not write, so treat the panel as a repeatable instrument and not a transcript of demand. Its value is that it holds still while the engines move.
That record makes the click decline interpretable. Because every snapshot keeps its citations, you can see whether a competitor won on their own pages or on a third-party source you never touched, and whether your absence is a retrieval or an evidence failure. The Brain keeps the claims you are entitled to make with the proof behind each, and recommended actions arrive as a ranked queue of content worth making, backed by answer evidence. Drafts are written in your voice from your author voice guidance and research sources, and a person reviews and approves everything before it publishes.
What Trovance will not promise is a citation, a ranking, or a recommendation. The engines are probabilistic, the variance is documented above, and any product guaranteeing an outcome inside someone else's model is selling against the evidence. There is no single visibility score here either: collapsing presence, citation, and recommendation into one number hides the rung you are losing. Past the answer, a cookieless first-party pixel records identifiable AI referrals and the conversions you define, and the Stripe revenue webhook, once set up, adds payments. Nothing is modeled, and we will say plainly which parts of that chain remain unobservable.
What it does close is the loop. After an asset ships, the next analysis cycle reruns the same questions and compares the new answers against the preserved ones, so you can verify whether the answer moved instead of assuming a traffic chart explained it.
What should you do this week?
Write down ten questions a real buyer asks in the two weeks before choosing. Ask each one in two or three phrasings on two engines, once a week for a few weeks, and record mentions, citations, recommendations, and shortlist slots separately. That baseline takes an afternoon and gives next quarter a fixed reference point.
Then fix retrieval: fetch your key pages with scripting off and see what an engine reads. Put one defensible number, with its source and date, on each page that answers a decision question. Check which third-party sources your engines cite, then add the unobservable-funnel line to your reporting before someone else fills it with a made-up model.
Do not expect this to move in days. In our experience retrieval fixes land first, then evidence, then earned third-party coverage. Nobody has published reliable intervals, so treat that as an order of operations. If you want the baseline without the manual runs, start a free Trovance trial and watch your buyer questions.
Measure what replaced the click
How to attribute AI traffic in GA4 - what your analytics can and cannot show.
How to measure AI search visibility without one score - why the rungs stay separate.
Do AI search visitors convert - what that referral stream is worth per session.
Does AI search visibility drive leads or revenue - the evidence for and against the pipeline claim.
Earn presence inside the answer
What gets cited by AI - the page traits that track with being cited.
How to earn third-party AI citations - most of the answer sits off your domain.
AI agents need evidence, not more content - why volume stopped working first.
A visibility gap is not a content brief - when the right response is no page.
