Ranked #1, 57 Clicks: What 7.5M of Our Own Search Console Impressions Show

Last updated: September 1, 2026.

We exported 16 months of Google Search Console for cleanlist.ai and recomputed every figure from the raw file. The property took 7,458,554 impressions at a 0.5439% CTR from an impression-weighted average position of 9.40. Dividing clicks by the published expected CTR for each page's position, only 18.5% of that impression pool behaves like human search. Desktop carries 88.25% of impressions at 0.300% CTR while mobile carries 11.52% at 2.380%, despite mobile ranking worse. One query, rocketreach alternatives, returned 373,692 impressions and 57 clicks, a query-level count that no single URL of ours owns. We hold the #1 organic result for it. The AI Overview above us cites seven domains and none of them is ours.

The export is published below as a CSV under CC BY 4.0, five withheld rows aside.

Where does Cleanlist lose in this study?

Badly, and in the place that matters most. On September 1, 2026 at 06:02 UTC we pulled the live US SERP for rocketreach alternatives and kept the result set as JSON. cleanlist.ai/alternatives/rocketreach is the number one organic result in it. Directly above it sits an AI Overview citing seven sources: aisdr.com, leadmagic.io, waalaxy.com, youtube.com, enrich.so, artisan.co and uplead.com. Cleanlist is not among them. The same pattern holds on zoominfo pricing, where an AI Overview cites six domains, none ours, and no Cleanlist URL appears anywhere in the ten organic results we retained, despite a 16-month average position of 7.48. We won the ranking and lost the answer.

One caveat on the 373,692 figure, because it is the number most likely to be quoted. Search Console reports it at query level, and a query's impressions are not any one page's. Two of our URLs can serve rocketreach alternatives, and at page level they hold 282,025 and 162,929 impressions. So the 57 clicks and the 0.0153% CTR are the query's, not the ranking page's, and the page-level and query-level exports of the same 16 months do not reconcile by design. The alternatives page economy study works the same inconsistency through from the page side.

What did we actually measure?

One Search Console domain property, cleanlist.ai, Web search type, exported September 1, 2026 and covering April 30, 2025 through August 29, 2026. That is 70 weekly rows, 827 pages, the top 1,000 queries, 238 countries and 3 device classes. Three page rows and two query rows are withheld from the published dataset for the reason set out in the methodology box, so every page-level and query-level figure below is computed on the 824 page rows and 998 query rows we publish, and the page and the CSV agree exactly. We also exported the Search Generative AI Features report, which covers 561 pages and only begins on May 18, 2026 because that is when Google started reporting it. Three independent breakdowns in the export (devices, countries and the weekly chart) each sum to exactly 7,458,554 impressions, which is how we know the property total. Everything in this post was recomputed from those files rather than read off a Search Console screen.

Why does the page-level export not match the property total?

Because Google counts impressions differently at the two levels, and it says so. Summing the full 827-row Pages export gives 7,725,528 impressions, which is 266,974 more than the property total, a 3.58% overstatement. Google's documentation states that when several URLs from one site appear in the same result set, "only one impression is counted for the entire card" at property level, while grouping by page gives each URL its own impression. So the 3.58% gap is the share of result sets where two or more of our URLs appeared together. This matters for anyone reproducing the work: a page-level analysis and a property-level analysis will not reconcile, and the difference is a real signal rather than an error. All per-page figures in this study use the page-level denominator. All site-wide figures use the property total.

The same effect runs through the AI features export, and it is worth stating rather than leaving for a reader to find. Its Pages file sums to 657,591 impressions while its weekly Chart file sums to 610,018, a 7.8% page-level overstatement of exactly the same kind. The ai_feature_impressions column of our published CSV sums to 657,172, the difference being the withheld rows. Page-level AI shares below use the 657,591 basis because they are page rows; the site-level AI rate uses the chart. The two are not interchangeable.

One more accounting note for anyone recomputing from our CSV. The published file holds 824 of the 827 page rows and 998 of the 1,000 query rows. The withheld rows carry 12,935 page impressions with 18 clicks, and 1,128 query impressions with 4 clicks, so the published page pool is 7,712,593 impressions and 40,745 clicks. That is the denominator for every page-level percentage on this page.

What is the headline number?

For each row we computed implied real impressions = clicks divided by the expected CTR at that row's average position, then compared it to the reported impressions. Across the 824 published pages the implied total is 1,426,727 against 7,712,593 reported, so 18.5% of the impression pool behaves like human search. Summing the rounded per-row columns in the published CSV rather than recomputing from the raw export gives 1,426,728.5, which is the same 18.5%. The remaining 6.3 million impressions produced click behaviour indistinguishable from zero. That 18.5% is the conservative reading, capping each row at 100% because an impression pool cannot be more than fully real. Without the cap, over-performing pages offset dead ones and the figure rises to 39.9%. Applying an AI Overview adjustment to the curve raises it further, to 21.4% capped and 60.9% uncapped. The honest statement is a band: somewhere between 18.5% and 60.9%, most plausibly near the bottom.

What CTR curve did we use, and why that one?

First Page Sage's "Google Click-Through Rates by Ranking Position", last updated May 28, 2025: positions one through ten at 39.8%, 18.7%, 10.2%, 7.2%, 5.1%, 4.4%, 3.0%, 2.1%, 1.9% and 1.6%. We picked it because it publishes exact integers for every position rather than a chart image, and because it is explicitly measured on result pages without other SERP elements, which is the right baseline for the question "what would a normal SERP have produced". For fractional positions between 1 and 10 we interpolate linearly in log-log space. Beyond position 10 we fit a power law to those ten points by ordinary least squares on log10, giving CTR(p) = 46.63 times p to the power of minus 1.4201, with an R-squared of 0.9879, and anchor the tail at the published position-10 value: CTR(p) = 1.6 times (p/10) to the power of minus 1.4201.

Is the curve wrong, or is the traffic wrong?

The obvious objection is that we picked a generous curve. Our own data answers it. Take the 56 queries in the export whose average position falls between 1.00 and 3.00 inclusive. They returned 19.115% CTR on 19,775 impressions, comfortably inside the 10.2% to 39.8% band the curve predicts for that range. We report the inclusive band on purpose. Cutting it at strictly under 3.00 drops the three queries sitting at exactly 3.00 and gives 53 queries, 19,570 impressions and 19.167%, which flatters the defence by five hundredths of a point and would be a boundary chosen after seeing the answer. Both readings pass. Worth naming before someone else does: rocketreach alternatives, the anomaly this whole study is built on, sits at position 3.01 and therefore falls outside this curve-defence band on either rule, by one hundredth of a position. cleanlist ai sits at average position 1.0 and returns 56.67%, well above the curve's 39.8%. reverse email lookup free sits at 2.87 and returns 17.40% against an expected 10.90%. Our best queries beat the published curve rather than falling short of it. So the curve is not systematically too optimistic for this property. The pages and queries that fail it fail it by two orders of magnitude, not by a calibration error.

What does the desktop versus mobile split show?

This is the single most discriminating fact in the export, and it is one line of a CSV. Desktop: 6,582,463 impressions, 0.300% CTR, average position 9.07. Mobile: 859,370 impressions, 2.380% CTR, average position 11.93. Desktop carries 88.25% of all impressions, ranks nearly three positions better, and converts 7.9 times worse. Run the same arithmetic per device and desktop's implied real impressions are 1,050,881, or 15.96% of its reported pool, while mobile's implied total exceeds its reported pool outright and caps at 100%. Same site, same pages, same titles, same snippets, same 16 months. Whatever is inflating the impression count is overwhelmingly a desktop phenomenon, and no explanation that treats both devices alike can survive that.

Where does the dead impression pool actually sit?

Not in the long tail, which is the assumption almost everyone starts with. Bucketing the 824 published pages by average position:

The graveyard is positions 5 to 8, holding 55.1% of every impression the site received at a 0.116% CTR. The deepest bucket, positions 21 and beyond, is the healthiest on this measure at 56.1%, and it is only 6.9% of impressions. The 15 pages that average a top-3 position return 0.128% CTR and are 0.9% human-like. Ranking better made things worse.

What happened to the query rocketreach alternatives?

It is 373,692 reported impressions, 57 clicks, 0.0153% CTR, average position 3.01. At position 3.01 the curve expects 10.16%, so the implied real impression count is 561. The reported pool is 666 times larger than the behaviour in it. We then checked the term's actual search demand with the Google Ads Keyword Planner via DataForSEO on September 1, 2026: 210 searches per month in the US, 590 per month worldwide in English, with the 12 reported months summing to 6,980 worldwide searches. Extrapolated across the export's 16 months that is roughly 3,360 US searches and 9,440 worldwide. Our reported impressions exceed worldwide measured demand by about 40 times, and US demand by 111 times.

Is that gap uniform across queries?

No, and the study would be dishonest if it implied otherwise. zoominfo pricing shows 104,276 impressions against a Keyword Planner volume of 1,600 per month in the US and 3,600 worldwide. Over 16 months that is 25,600 US and 57,600 worldwide searches, so the reported impressions exceed worldwide demand by only 1.8 times, which is well inside what query-variant bucketing and non-English traffic could explain on their own. The anomaly is extremely concentrated. Those two queries alone carry 477,968 impressions, 43.7% of the published query pool and 6.4% of every impression the site received, and between them they returned 59 clicks. Any explanation has to account for that concentration, not just for an average.

Do AI Overviews explain the gap?

They cannot explain most of it, on Google's own numbers. Google confirms that AI feature appearances "are included in the overall search traffic in Search Console" inside the Web search type, so AI Overview source links are counted as ordinary impressions at ordinary positions and are not separable in the main report. But the Search Generative AI Features export does give a bound. Over its window, May 18 to August 29, 2026, it reports 610,018 AI feature impressions. Matching that to the search chart takes one decision worth showing. The AI export's first row covers May 18 to 23, six days, while the search chart's overlapping row starts on May 17. Summing the search chart strictly from May 18 gives 14 weekly rows and 5,245,886 impressions, so the AI share is 11.63% over a window holding 70.3% of the site's impressions. Joining the AI export to the full search week that contains it gives 5,479,561 and 11.13% over 73.5%. We publish the strict same-day figure of 11.63% as the headline and the aligned-week figure as the variant, because comparing a six-day numerator to a seven-day denominator is the kind of join that turns into a correction later. Even if every AI feature impression were a dead one, it accounts for under 12% of the impressions in that window. Ahrefs, measuring 300,000 keywords, found AI Overviews correlate with a 34.5% lower CTR for the top-ranking page, not a 99% lower one.

Page by page, the AI Overview effect is real, large on specific pages, and absent on the biggest dead pools. /blog/2026-05-08-rocketreach-alternatives-2026 shows 98,543 AI feature impressions, 34.9% of its 282,025 search impressions. But the single largest impression pool on the site, /blog/2026-04-12-data-normalization-complete-guide with 502,592 impressions and 46 clicks at position 6.26, shows just 2,080 AI feature impressions, 0.4%. /solutions/find-founder-email, 356,480 impressions and 113 clicks, shows 1,334, also 0.4%. One caveat we cannot remove: the AI numerator covers the last 15 weekly rows of the window and the search denominator covers all 70, so these page-level percentages are floors and the true in-window shares are higher. The ordering across pages still holds, and the biggest dead pools are the least AI-touched.

Do other SERP features explain it?

No, and this is the cleanest refutation in the whole export because it needs no external source. The Search Appearance breakdown lists every labelled feature the property earned across 16 months: Translated results at 5,776 impressions, Review snippet at 6,981, Product snippets at 30. That is 12,787 impressions in total, 0.1714% of the pool. Featured snippets, sitelinks, video and image results, product grids and local packs are either absent or unlabelled here. Whatever is generating six million behaviourally-dead impressions, it is not carrying a SERP feature label in Search Console. Anyone running this analysis on their own property should check this file first, because it takes 10 seconds and eliminates an entire class of explanation.

Could it be rank trackers and scrapers?

This is the hypothesis the evidence fits best, and the one we cannot prove. Supporting it: 88.25% of impressions are desktop, which is where rank-tracking tools run; the United States is 74.92% of impressions but only 50.63% of clicks; the impression mass sits precisely on the commercially tracked head terms, competitor pricing pages and "best tools" listicles that every SEO platform in the category monitors; and one query's impressions exceed its entire measured worldwide demand by 40 times. Against it: the flood arrived as a step change, from 4,645 impressions in the week of February 1, 2026 to 51,034 in the week of March 1, and Google states it filters known bots. Search Console exposes no user agent, no IP and no session, so this cannot be resolved from this data.

Is a long tail dragging the average position?

No. This explanation predicts that the dead impressions sit at deep positions where CTR is legitimately near zero, and the position bucket table refutes it directly. Positions 21 and deeper are the most human-like bucket at 56.1%, and they hold only 6.9% of impressions. The dead pool is concentrated at positions 5 to 8, where the published curve expects between 5.1% and 2.1% CTR and we observe 0.116%. Separately, 402 of the 824 published pages recorded zero clicks across 16 months, which sounds catastrophic until you notice they carry just 179,048 impressions, 2.3% of the page-level pool. The dead impressions are not spread thinly across hundreds of forgotten URLs. They are stacked on a few dozen well-ranking commercial pages.

Could it just be bad titles and meta descriptions?

Partly, but it cannot be the mechanism. A weak title depresses CTR by a factor, typically tens of percent. It does not depress it by 100 times. More decisively, a page serves the same title and the same meta description to desktop and mobile, and our desktop CTR is 0.300% while our mobile CTR is 2.380% on a worse average position. If titles were the cause, both devices would suffer together. Snippet mismatch also cannot explain why reported impressions exceed a keyword's entire measured global search volume by 40 times, because no snippet decision changes how many humans typed the query. We take the criticism seriously on the pages that under-perform mildly. It does not reach the pages that under-perform by two orders of magnitude.

So what does the evidence actually support?

That impressions have decoupled from human demand, and a position-1 ranking no longer implies clicks. Three findings carry that conclusion: the desktop-mobile inversion, where the better-ranking device converts 7.9 times worse; the concentration of dead impressions at positions 5 to 8 rather than in the tail; and one query reporting more impressions than the term has worldwide searches. AI Overviews are a real and measurable part of it, bounded by Google's own AI features export at 11.63% of impressions in the only window where it exists. Labelled SERP features are 0.17% and can be discarded. What remains is a large desktop-weighted impression surplus whose origin we can characterise but not identify. Calling it "AI Overviews destroyed our CTR" would overstate what this data can show.

What can we not resolve from this data?

Four things, stated plainly. First, whether the desktop surplus is automated measurement traffic or genuine human impressions that never converted: Search Console publishes no user agent, IP or session, and Google gives no figure for how much bot traffic it filters. Second, the AI features export begins May 18, 2026, while the impression flood began the week of February 8, 2026, so the first 15 weeks of the anomaly are untestable for AI features. Third, Search Console reports the topmost position averaged across all queries for a page, so a page at #1 on one term and #40 on another shows a misleading mean. Fourth, we cannot separate query-variant bucketing in Keyword Planner from a genuine impression surplus at the 1.8 times level, only at the 40 times level.

How do you run this on your own property in ten minutes?

Export Performance, Web search type, last 16 months, and take Pages, Queries, Devices and Search Appearance. For every row compute implied = clicks / expectedCTR(position) using the curve in the methodology box, then ratio = implied / reported impressions. A healthy property lands near 1.0. Then run three checks. Devices: if desktop CTR is a small fraction of mobile CTR while desktop ranks the same or better, your desktop pool is inflated. Search Appearance: if labelled features are a rounding error, stop blaming SERP features. Volume: take your five biggest queries by impressions and compare against Keyword Planner volume for the same period. Any query reporting more impressions than the term has global searches is not a ranking problem.

What should a GTM leader change on Monday?

Stop reporting impressions and average position as pipeline-adjacent metrics, because on this property they moved opposite to clicks for six straight months. Report clicks, the implied-real ratio, and conversion. Second, re-read your content bets against the ratio rather than the impression count: on our site the pages holding the most impressions are the ones producing the fewest clicks, and the 15 pages averaging a top-3 position return a 0.128% CTR. Third, watch where the answer is being given. We rank #1 for rocketreach alternatives and are cited in none of the seven sources in the AI Overview above us. Ranking and being named have become two different objectives, and only one of them still delivers a visitor.

What is genuinely working, and what does it tell you?

One page on the site shows a normal curve, and its identity is the finding. /tools/reverse-email-lookup, a free single-lookup utility, took 496,283 impressions and 23,038 clicks at 4.64% CTR from average position 10.67, against a curve expectation of 1.46%. It beats the published curve by 3.2 times and it alone produces 56.5% of the site's clicks from 6.4% of its impressions. Six pages produce 80.4% of all clicks. Under a rule of "actual CTR at least half of expected", 138 of 824 pages qualify; they hold 16.3% of impressions and deliver 84.9% of clicks. The pattern: a page a human wants to use still gets clicked. A page that answers a question is increasingly read without being visited.

What did we change after the first internal review?

This page went through an adversarial recheck before it shipped, run by someone whose job was to break it. Five things moved, and the honest thing is to show them rather than to quietly publish the second version.

The curve-defence band. We first reported 53 queries at 19.167% CTR on 19,570 impressions for positions 1 to 3. That band was cut at strictly under position 3.00, which excluded three queries sitting at exactly 3.00. Inclusive of them the band is 56 queries, 19,775 impressions and 19.115%. It changes nothing about whether the curve is too generous, and choosing the cut that returned the higher number would have been a small piece of dishonesty on a page whose entire argument is that we are showing our worst data.

The AI feature share. We first reported 11.13%, which divided a 15-row AI export by a 15-week search sum whose first week starts a day before the AI window opens. On the strict same-day window the figure is 11.63% over 14 weeks, and that is now the headline. It moves the bound on how much AI Overviews can explain from a ninth of the pool to a little more than a ninth, which strengthens the AI explanation slightly and still leaves it nowhere near sufficient.

Three live-SERP claims. The first draft said the rocketreach alternatives pull covered desktop and mobile, that it ran to depth 20, and that our page was absent from the top 12 organic results on zoominfo pricing. The artifact we actually retained is one result set of ten organic positions per query with no device field. All three claims are now scoped to what the file supports. Nothing in the finding changed; the evidence for it was narrower than the sentence describing it.

The 373,692 impressions. That is a query-level count and the first draft attached it to one page's ranking. Two of our URLs can serve the query, holding 282,025 and 162,929 page-level impressions. The clicks and the CTR are unchanged and so is the anomaly.

The published row count. The dataset holds 824 page rows and 998 query rows rather than the export's 827 and 1,000, for the reason given in the methodology box. Every page-level figure was recomputed on that basis. The headline moved from 1,427,580 implied on 7,725,528 reported to 1,426,727 on 7,712,593, which is 18.5% either way. The uncapped figure moved from 39.8% to 39.9% and the AI-adjusted uncapped figure from 60.8% to 60.9%, so the published band is now 18.5% to 60.9%. Two rows of the position table moved by one or two pages and by up to three tenths of a point. No conclusion on this page depends on any of it, which is the only reason the exclusion was acceptable.

Where is the raw data?

The row-level dataset is published at /data/ai-overview-ctr-study-2026-09.csv under a Creative Commons Attribution 4.0 licence. Use it, check it, and correct us if the arithmetic is wrong. It contains 1,822 rows, one per page and one per query: 824 pages and 998 queries.

Sources

• First Page Sage, Google Click-Through Rates (CTRs) by Ranking Position, last updated May 28, 2025. The expected-CTR curve.
• Ahrefs, AI Overviews Reduce Clicks by 34.5%, April 2025, 300,000 keywords. The AI Overview adjustment factor.
• Google, Search Console: impressions, position, and clicks. Property-level versus page-level impression counting, and the topmost-position rule.
• Google, Google Search's AI features and your website. AI feature appearances are reported inside the Web search type.
• Google Ads Keyword Planner volumes and live SERP results retrieved through the DataForSEO API, September 1, 2026.