Zero-Click Search Study: 906,663 Impressions Show Ranking Better Now Means Fewer Clicks

Every SEO model rests on one assumption: rank higher, get more clicks. We pulled 90 days of Google Search Console data for a B2B software domain and found that assumption no longer holds for conversational queries. On the longest, most specific questions, this site ranked best and was clicked least.

This is our own data, published in full ratio form, because we could not find a study that cut click-through rate by query length rather than by position alone.

Methodology

• Source: Google Search Console, Search Analytics API, a single verified sc-domain property for a B2B software site.
• Window: 90 days, 25 May to 22 August 2026.
• Scale: 3,939,868 total impressions and 18,077 total clicks across all queries.
• Filtering: all branded queries were removed, leaving 24,254 non-branded queries, 906,663 impressions and 2,568 clicks. Branded search behaves completely differently and would have masked the effect.
• Cuts: click-through rate by position bucket, and click-through rate by query word count.
• Limits: one site, one industry, English-language, and Search Console reports position as an average across impressions rather than per impression. Google does not expose an AI Overview filter on this property, so the mechanism is inferred from the pattern, not read from a labelled field.

Anyone with Search Console access can reproduce this cut on their own property, and we would encourage it. The interesting question is whether the shape repeats outside our vertical.

One note on why we are publishing our own numbers rather than an anonymised aggregate. Studies in this area usually report percentage changes without the underlying counts, which makes them impossible to check and easy to misread. Absolute impressions and clicks are included here so a reader can verify the arithmetic, see exactly how large each bucket is, and judge for themselves whether a bucket is big enough to draw a conclusion from. The 12-plus-word bucket returning a single click is a striking number, and it deserves to be checkable rather than taken on trust.

Finding 1: click-through rate is far below every published benchmark

Start with the conventional cut, position buckets.

Benchmark columns are the rough consensus from long-running curve studies such as Advanced Web Ranking's CTR study and Backlinko's aggregate CTR research.

The gap at the top is the striking part. Position 1-2 returned 0.849% where the historical curve says roughly 25%. That is not a degraded version of the old curve. It is a different curve.

The obvious objection is that this must be a title or snippet problem. It is not. No title is bad enough to take a first-position result from 25% to under 1%, and the next cut shows why.

Finding 2: the longer the query, the better the rank and the worse the click-through

This is the cut that matters, and we have not seen it published elsewhere.

Read the last two columns together. As queries get longer, average position improves from 19.9 to 5.5, and click-through rate collapses from 0.333% to 0.002%.

Sixty-one thousand impressions at an average position of 5.5 produced one click.

Under the old model this is impossible. Ranking fifth on a specific, high-intent, long-tail question should be among the most valuable positions in search. Long-tail queries have historically converted better than head terms precisely because specificity signals intent.

That relationship has inverted. The queries where this site ranks best are the queries where users never leave the results page.

Finding 3: most impressions now produce nothing

More than five thousand queries where this site ranks in the top three produced no clicks at all. For context on how unusual that is, a top-three position has historically been the most contested real estate in search precisely because it reliably converted impressions into visits.

Why this is happening

The pattern points at one explanation, and it lines up with what the rest of the industry has measured independently.

Long, conversational queries are exactly the queries that trigger AI-generated answers. Google's own documentation describes AI Overviews and AI Mode as surfaces that synthesize an answer from multiple sources, with links presented alongside rather than as the primary result. Google has said these surfaces are aimed at longer and more complex questions.

Search Console counts an impression when a page is used as a source in those surfaces. So a page cited inside an AI answer accrues impressions and a strong average position while the user's need is met without a click.

That gives the exact signature in our data: better positions on longer queries, near-zero clicks.

Independent research agrees on direction. Pew Research found users are markedly less likely to click a link when an AI summary is present. Ahrefs measured a substantial decline in click-through for queries with an AI Overview. Semrush's analysis found AI Overviews concentrated on longer, question-shaped queries. SparkToro's zero-click work established the broader trend well before AI answers arrived.

Our contribution is the query-length cut. The effect is not evenly spread across a site. It is concentrated almost entirely in conversational queries, and it is severe there.

What we ruled out

Title and meta quality. A snippet problem would depress click-through roughly evenly across query lengths. Instead the effect scales with length, and short queries on the same site still perform an order of magnitude better.

Ranking loss. Positions improved over the window. The site's average position moved from 11.6 to 11.0 while clicks stayed flat.

Tracking error. Search Console is the source and the totals reconcile with the property's reported figures.

Low-quality impressions. These are not position-90 phantom impressions. The worst-performing bucket has the best average position in the dataset.

What this changes

Click-through rate is no longer a clean quality signal. A page can be doing exactly what search is now for, answering a question well enough to be selected as a source, and register as a failure in every click-based report.

Rank tracking on conversational queries has limited value. Position 5 on a 12-word question and position 5 on a two-word head term are worth entirely different amounts. Averaging them produces a number that means nothing.

Impressions become the leading indicator. If a page is cited in AI answers, impressions accrue while clicks do not. Impressions plus citation tracking measure that surface. Clicks do not see it at all.

The measurable outcome moves off-site. When the answer is consumed in the results page, the return on content is being the source that gets cited, so generative engine optimization and citation share become the scoreboard. That is a different discipline from ranking, and it is why we track AI citations as a first-class metric alongside search performance.

Content strategy splits in two. Commercial-intent queries, pricing, alternatives and comparisons, still produce clicks and still convert. Definitional and conversational content increasingly earns citations rather than sessions. Both are worth doing. They should not be measured the same way.

Does the effect hold across content types?

We split the same non-branded set by the shape of the query rather than its length, because intent and length are correlated but not identical.

Transactional and comparison queries still behave normally

Queries containing pricing, cost, alternatives, versus or a competitor brand retained clicks at rates far closer to the historical curve. These are queries where an AI summary cannot finish the job. Someone comparing two vendors on price needs to reach a page, check the plan tiers, and often talk to somebody. The answer surface can summarize, but it cannot complete the task.

This is consistent with what we see downstream. Our commercial pages, pricing comparisons and competitor breakdowns like Gemini Enterprise pricing and ChatGPT Enterprise pricing, are the pages that still send people onward to talk to us. Definitional pages almost never do.

Definitional and how-to queries lost the most

Question-shaped queries beginning with what, how, why or which showed the steepest decline. These are precisely the queries an AI answer can resolve completely. If somebody asks how a technology works and receives four accurate paragraphs, the job is done.

That does not make the content worthless. It changes what the content is for. A definitional page that gets cited is doing brand work at the moment of research, in front of a buyer, at zero marginal cost. It is simply not doing traffic work, and reporting that treats it as a traffic asset will conclude it failed.

The awkward middle

Queries that are conversational in form but commercial in intent are the hardest to read. "What is the most affordable way to automate finance operations" is a question and a shopping trip at the same time. In our data these produced impressions with occasional clicks, and they are the bucket most worth watching, because they are where an AI answer either satisfies the buyer or sends them looking.

What we cannot conclude from this

Being explicit about the limits matters more than usual here, because the finding is easy to overstate.

We cannot prove AI Overviews are the cause

Search Console does not expose an AI Overview appearance filter on this property. We are inferring the mechanism from a pattern that matches it closely: better positions on longer queries, near-zero clicks, concentrated in question-shaped searches. That is strong circumstantial evidence and it aligns with independent research, but it is not a labelled measurement. A researcher with access to per-impression SERP feature data could settle it properly.

One site is not an industry

This is a single B2B software domain. The direction agrees with larger studies, but the magnitudes here should not be generalized. A publisher, a retailer and a local business will each see a different shape, and the transactional end of the spectrum should be far less affected.

Average position is a blunt instrument

Search Console reports position averaged across impressions. A query that ranks second in classic results and appears as a cited source in an AI answer produces a single averaged number that describes neither situation accurately.

We did not measure what happened after the non-click

A user who reads an AI answer citing a brand and searches that brand two days later shows up as branded traffic with no connection to the original query. Our branded impressions fell over this window, which does not support a strong version of that effect, but we cannot rule it out either. Attribution across that gap is genuinely hard and we have not solved it.

How to run this cut on your own site

1. Pull Search Console query data for 90 days via the Search Analytics API, with a row limit high enough to avoid truncation. The API documentation covers the query dimension.
2. Remove branded queries. Skipping this step will hide the effect entirely.
3. Bucket by word count, not just by position.
4. Compute click-through rate and average position per bucket.
5. Compare the longest bucket against the shortest.

If the pattern holds, average position will improve as word count rises while click-through rate falls. That divergence is the measurement. Reproducing it on other verticals is the obvious next step, and academic work on generative search and its effect on the web ecosystem is starting to model the same dynamic.

If clicks are gone, how do you make a page worth writing?

The practical question this leaves is what a page should optimize for when the click is not available. Our working answer, applied to our own content since running this analysis.

State the answer before you explain it

Models extracting an answer take the most quotable, self-contained statement available. A page that opens with three paragraphs of context and states its answer in the middle is harder to quote than one that answers in the first two sentences and elaborates afterwards. This is the inverted pyramid, and it turns out to matter more for machine readers than it ever did for human ones.

Publish numbers you generated

This is the single largest lever and the most underused. A model summarizing a topic will attribute a statistic to whoever produced it. Restating someone else's figure makes you a middleman with no citation value; the original source gets named instead. Original data, even from a modest dataset, makes you the thing that gets named.

That is most of why this study exists. We have Search Console data nobody outside the company can see, and the query-length cut had not been published anywhere we could find.

Make structure explicit

Question-shaped headings that mirror how people actually ask, real tables rather than pipe-separated text that looks like a table, and structured data on question sections. The goal is that a machine parsing the page can identify which passage answers which question without inference.

Be specific enough to be worth quoting

Vague claims are unquotable. "AI can reduce costs" attributes to nobody. "Queries of 12 or more words averaged position 5.5 and returned one click from 61,782 impressions" has an owner. Specificity is what makes a sentence survive summarization.

Say the honest thing about limitations

Counterintuitively, stating what your data cannot show appears to help. A page that acknowledges its own limits reads as a more reliable source, to human readers and, in our observation, to models weighing which sources to lean on. It also protects you when somebody checks.

Keep the commercial pages commercial

None of the above applies to a pricing page. Those queries still produce clicks, users still need to arrive to act, and optimizing them for citability instead of conversion would be a straightforward error. The split is the point: two content types, two scoreboards, two sets of tactics.

How to explain this to someone who reports on traffic

The hardest part of this shift is not technical. It is that content teams are measured on sessions, and sessions are now a partial measure of a job well done.

A few framings that have worked for us.

Separate the two portfolios in reporting before anything else. Report commercial-intent pages on clicks and conversions, and conversational pages on impressions and citations. A single blended traffic number now averages two things that behave in opposite directions, and the average describes neither.

Show the query-length cut rather than arguing about it. The table above is more persuasive than any explanation, because it is the reader's own data when they run it. Somebody who believes the content is underperforming will usually change their mind when they see position improve and clicks collapse in the same row.

Be careful not to over-claim in the other direction. "Clicks do not matter any more" is wrong and will get you correctly challenged. Clicks matter a great deal on the pages where clicks still happen. The accurate claim is narrower: for conversational queries, on this dataset, click-through no longer tracks how well the page did.

Set the expectation early that citation measurement is immature. Tracking tools for AI citations are new, sampled rather than exhaustive, and disagree with each other. Reporting them alongside search data is worth doing, and presenting them with the same confidence as Search Console figures would be misleading.

What we would want to see measured next

Several questions this dataset raises and cannot answer.

Does the inversion hold outside B2B software?

Our vertical is unusually well covered by AI answers because the subject matter is technical, well documented and non-controversial. Health, legal and financial queries carry different treatment. Retail and local search have a transactional floor that ours does not.

How much brand lift does a citation produce?

A citation puts a brand in front of a researcher at the exact moment of evaluation. That should be worth something, and nobody has a clean measurement of how much. Our branded impressions declined across this window, which argues against a large effect, but branded search is influenced by too many other things to isolate it.

Does citation frequency correlate with anything commercial?

We track citations as a metric because clicks stopped describing the content's performance. Whether citation share moves pipeline is a separate question, and we do not yet have enough history to say. Anyone claiming a settled answer on that today is guessing.

Is the effect stable or still moving?

These surfaces change constantly. A study run six months from now on the same property could show a materially different curve in either direction. Treat this as a measurement of one window, not a permanent law.

What we are doing about it

We stopped optimizing conversational pages for click-through, because the clicks are not being lost at the snippet and no title rewrite recovers them. We moved that content toward being citable instead: direct answers stated plainly near the top, structured data on question-shaped sections, original figures rather than restated ones, and clear attribution so a model quoting the page has something to attribute.

Commercial pages, where clicks still behave normally, we still optimize normally.

Coworker AI connects to 50+ tools, maintains organizational memory across them, and runs agents that act on what they find. If you want to talk through what this shift means for your own content operation, book a demo.

Frequently asked questions

What is a zero-click search?

A search where the user's need is met on the results page and no result is clicked. Featured snippets and knowledge panels produced this for years; AI-generated answers have widened it substantially, particularly for longer conversational questions.

Why would ranking higher produce fewer clicks?

Because the queries a site ranks highest on can be the queries most likely to trigger an AI-generated answer. In our data, queries of 12 or more words averaged position 5.5 and returned one click from 61,782 impressions. The page is being used as a source rather than as a destination.

Does a high impression count with no clicks mean the page is failing?

Not necessarily, and this is the measurement trap. Search Console counts an impression when a page is used as a source in an AI surface. High impressions with near-zero clicks on conversational queries is closer to evidence the page is being cited than evidence it is failing.

Can better titles and meta descriptions fix this?

No, and our data rules it out. A snippet problem would depress click-through evenly across query types. Instead the effect scales with query length, and short queries on the same site perform an order of magnitude better with the same title conventions.

How large was the dataset?

906,663 non-branded impressions and 2,568 non-branded clicks across 24,254 queries, over 90 days, from a single verified Search Console property. Site-wide totals including branded queries were 3,939,868 impressions and 18,077 clicks.

Does this apply to every website?

Unknown, and we would not claim it does. This is one B2B software domain in one language. The effect should be strongest for informational and conversational content, and weakest for transactional queries where users still need to reach a destination to act.

What should replace click-through rate as a measure?

For conversational content, citation share in AI answers plus impression growth. For commercial content, clicks and conversions still work. The mistake is applying one scoreboard to both.

Related reading

• What Is Organizational Memory?
• Agent Workflows Explained
• LLM Agent Architecture
• Best AI Agent Builders for Enterprise
• ChatGPT Enterprise Pricing
• Gemini Enterprise Pricing
• AI Workflow Automation Examples
• What Is MCP (Model Context Protocol)?
• What Is AI Agent Memory?
• What Is an LLM Gateway?
• Mem0 vs Zep vs Letta Compared
• AI Agent Orchestration Platforms
• Enterprise AI Agents
• 16 Ways to Use AI For Knowledge Management
• Knowledge Management Trends
• Agent Performance Metrics
• Enterprise AI Pricing Comparison