The click loss from AI Overviews is not distributed evenly across your keyword set, and treating it as one sitewide number is how teams cut pages that were still earning. Ahrefs' April 2025 study of 300,000 keywords found that an AI Overview correlated with a 34.5% lower average clickthrough rate for the top-ranking page on informational keywords, and a December 2025 re-run of the same study put it at 58%. That is a precise finding about one query shape. Queries carrying a buyer's constraints are a different population, and no public study tells you what happened to yours.
The short version: as a working hypothesis, an AI Overview resolves a broad question more readily than a narrow one loaded with conditions, and no public study measures where the surviving clicks land. Let query shape drive the plan: group your queries by how many constraints they carry, measure clickthrough inside each group, and move effort toward the groups that still send a person to a page.
The commercial case is stronger than the traffic case. G2's research found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers now start research with an AI chatbot more often than with Google, and Gartner predicted in February 2024 that traditional search volume would fall 25% by 2026. The constrained queries are where the remaining clicks and the actual buying decisions overlap.
Which of your queries are most exposed?
The queries most exposed are the ones a single paragraph can finish. A definitional question has one settled answer many pages already agree on, the easiest case for a summary to assemble and for a reader to accept without clicking. Add a constraint and the summary has to be right about a budget or an existing stack, and a wrong constraint is obvious to whoever typed it.
Read the Ahrefs figures as a description of that easy case: informational keywords, the top-ranking page, and a correlation rather than a causal estimate for any one site. Applying 58% to a keyword set full of comparison and specification queries imports a condition the study never tested.
Ahrefs found 38% of AI Overview citations rank in the organic top 10 in March 2026, so a Google ranking and an AI citation are different outcomes.
Why does a narrow query resist a one-paragraph answer?
Because every added constraint is a condition the answer has to satisfy at the same time, and retrieval degrades as conditions stack. A 2026 study of 1,000 multi-condition queries found that even its strongest configuration located a document satisfying every condition for 81.1% of queries while still placing partial matches ahead of complete ones. That was long-document retrieval, so it describes a mechanism; the public web is a different setting.
The same pressure shows up in indexing. An experiment across three retrieval pipelines and six datasets found question-matched indexing reached 71.5% average claim recall against 53.6% for naive retrieval, with average context precision of 63.5% against a 42.3% baseline. The same work found noise sensitivity got worse, and it tested neither public websites nor AI citations; what an engine publishes is a further step.
Buyers also add constraints gradually, which makes the narrow question the real one. A 2026 analysis of 8,133 multi-turn human-LLM conversations covered 670 English commercial multi-turn conversations plus a separate 7,463 PRISM conversations from 1,389 participants. A request dimension stated earlier remained only in the conversation history for 50.3% of the commercial conversations and 44.8% of the PRISM ones, and the analysis measured neither citations nor purchases.
Do narrow buyer questions matter more commercially?
Yes, and the reason has little to do with search. A constraint-heavy question states a buyer's segment, budget, and alternatives in one line, which is late-stage information you normally pay a form to collect.
The shortlist arithmetic makes those questions expensive to lose. 6sense found the average buying group member says they know 75% of the vendors on their shortlist. Set against the five-vendor buying group the same report describes, that leaves roughly one seat for a vendor the group has no prior experience with, which is this article's arithmetic and a step 6sense never took. The constrained comparison query is where a stranger competes for that position.
AI research makes that position reachable. In G2's research on how AI search is changing software buying, one in three buyers said an AI chatbot led them to choose a vendor they hadn't considered. Profound's 7.5 million-conversation sample saw the commercial share of that sample rise across a year inside a proprietary corpus, which is not a count of buyers.
Where do you find the long questions buyers ask an assistant?
In your own record of buyer language, because what a person brings to an assistant is a situation, phrased the way they live it. They state the constraint and the existing stack in one line; a keyword export flattens that to a head term. Sales notes and support tickets hold that phrasing, and your Search Console query report shows the long Google strings that already reach you, read with Google's own definitions of impressions, position and clicks. Public corpora such as a working paper drawing on 1.1 million sampled conversations describe a population rather than your buyers, and nothing public reports the prompts your buyers type into an assistant, so keep each question in the words it arrived in.
Then plan for the shorter questions the long one becomes. Google states that AI Overviews and AI Mode use a query fan-out technique and that pages must be indexed and snippet-eligible, and ChatGPT search turns a prompt into one or more targeted queries. The one-line situation is taken apart before anything is retrieved, so each piece has to be answerable from text on the page.
How should this change what you write?
Write to the constrained question in the buyer's own words, and stop rewriting broad pages to sound quotable. C-SEO Bench (NeurIPS 2025) tested conversational-SEO rewrite methods across two tasks and six domains and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them.
The strongest case left is for evidence, and it serves the reader either way. The GEO study (Aggarwal et al., KDD 2024) found that adding statistics, quotations, and citations lifted citation visibility in its benchmark, while keyword stuffing did nothing in the same error-bar table; C-SEO Bench's retest found those evidence methods rarely moved citation rank. Persuasion language can cost you outright: scarcity and exclusivity framing measurably reduced how often a model recommended a product across three sets of 10 fictitious products.
In practice the page unit becomes the constraint set, with the constraint in the heading and answered in the first two sentences. Put the price band, the integration name, and the honest exclusion into liftable text; a page written for one buyer situation carries that evidence at full strength.
Then confirm the page is readable at all. Vercel's crawler research with MERJ documented that GPTBot and its peers read raw HTML without running JavaScript, so a specification table injected client-side is an empty shell. Sequencing it with an existing search program is covered in adding GEO to an SEO program.
Why does a sitewide traffic average hide the split?
Because the average mixes two populations moving in opposite directions. If broad-query clicks fall while constrained-query clicks hold, one organic line drops and tells you nothing about which pages to keep. The decision you need is per query shape; the reporting default is per property.
Analytics makes that harder. GA4 names only 5 assistant examples for its separate assistant channel, so expect AI Overview and AI Mode sessions inside Organic Search beside ordinary organic ones. Segment by query group in Search Console first, then check what those sessions did, remembering that GA4 computes engagement rate as sessions with a key event divided by total sessions, a visit and not a decision.
A loss's composition matters more than its size. 6sense's read on B2B sites losing traffic to LLM interfaces is that the visitors going missing are likely the professionally curious and future buyers, while in-market buyers still interact with vendors at the same rate. A decline made of the first group is a different business event from one made of the second.
Answers need their own sampling discipline. SparkToro and Gumshoe.ai found AI recommendation lists are highly inconsistent when the same prompt is asked again, and a 2026 variance-components study reports that a single answer is an unreliable estimate of where a brand ranks. Size your panel before you draw a trend from it.
How does Trovance track the split by query shape?
Trovance makes the buyer question the unit of record instead of the keyword. You define the market questions your buyers ask, at the specificity they ask them, and Trovance runs those tracked questions every week across engines. Every answer run is preserved as a snapshot with its context intact: who was mentioned, who was cited, who was recommended, and which sources carried the answer.
Because each question is stored with its constraints, the record keeps the two populations in this article apart. A broad category question and a constrained comparison question are separate tracked questions with their own share of answer over time, so you can observe the broad one resolving inside the answer while the narrow one still routes a reader to a page. That is the split a sitewide traffic average erases.
The output is meant to be worked. The Brain keeps the claims you are entitled to make with the proof behind each, and recommended actions name what the record shows is missing and why. A missing page in the record is a candidate for a brief, not a brief on its own: a person decides whether the gap is worth closing before anything is drafted. Drafts use your author voice guidance and research sources, a person reviews and approves before anything publishes, and the next analysis cycle reruns the same questions so you can compare new answers against the preserved ones.
What Trovance will not promise is a citation, a ranking, or a recommendation on any query. The engines are probabilistic, the published record says most rewrite tactics do not move citation rank, and one answer is a sample rather than a verdict. Trovance is building toward a tighter join between query shape and observed click behavior; today it observes and preserves answers, and click data still comes from your own analytics.
What should you do this week?
Start by grouping, because everything downstream depends on it. Export your Search Console queries, tag each by how many constraints it carries, and compute clickthrough one group at a time.
Then act on the shape of the loss. For groups still clicking, write or sharpen one page per constraint set, with the evidence in text a model can lift. For groups that have gone quiet, the measurement shifts from traffic to answers: ask each question in a few phrasings, week after week, and record whether you were mentioned, cited, or recommended, though a single-turn prompt, however often you run it, is a lossy stand-in for a buyer adding constraints across turns.
Be honest about the timeline. How fast a rendering fix or an evidence change reaches an answer depends on re-crawl cadence, and nobody has published a reliable number for that. Anyone selling a guaranteed AI Overview placement is selling against the published record.
If you want that question-level view without building the harness, start a free Trovance trial with the ten buyer questions your narrow pages are supposed to answer.
Where the click actually went
How to get featured in Google AI Overviews - what is controllable on the surface that took the click.
What gets cited by AI - which page traits keep showing up inside answers.
Do AI search visitors convert? - what a surviving click is worth after it lands.
Can likely buyer questions improve AI retrieval? - whether writing to buyer phrasing helps retrieval.
Plan and measure by query shape
GEO vs AEO vs SEO - which program owns which part of this problem.
GA4 AI traffic attribution - pulling assistant sessions out of the organic average.
A visibility gap is not a content brief - deciding which missing page is worth writing.
How to build a competitor comparison page - the narrowest query shape and the page that answers it.
