For the past two years, a dozen studies have reached the same conclusion: AI Overviews are destroying organic click-through rates. -34.5% in position 1 according to Ahrefs, 8% of users clicking compared with 15% when no AI summary appears according to Pew Research, -15.5% on average according to Amsive, and -61% according to Seer Interactive. They agree on the direction, but the magnitude varies fourfold.
The most widely quoted figure in that list, -61%, was later revised by its own authors. In their 2026 update, covering 53 brands and 5.47 million tracked queries, organic CTR for queries with an AI Overview rose from 1.3% in December 2025 to 2.4% in February 2026. The gap with queries that had no AI Overview was no longer 61%, but 38%. Seer also notes two awkward facts for the original claim: its control group moved too, and AI Overviews appear mainly on informational queries, which already attracted fewer clicks. This is a selection effect, not just a causal effect.
I do not have a thirteenth study to add. My in-house tool has detected AI Overviews since July 2026. Of the screenshots it has actually inspected since then, 34.1% contain an AI Overview. The phenomenon is therefore widespread in French SERPs. What is missing is not scale, but enough data to measure its effect. Yet the method already exists and fits into a Search Console export. I explain it at the end of this article.
This article explains what a CTR curve really measures and how to turn it into a metric you can use to prioritize the roadmap produced by an SEO audit. If any technical terms are unfamiliar, keep the SEO glossary open in another tab, old-school style.
In 30 seconds
The short answer to the question in the title: no published study, including mine, can answer it for French SERPs. AI Overviews are widespread, appearing on 34% of the result pages I track, but the figures in circulation combine three distinct effects without isolating any of them. The setup that separates them requires only a Search Console export, and you will find it at the end of this article.
Three things to remember if you are in a hurry:
- There is no single CTR for position 1. Published studies range from 9% to 37.8%, and the difference comes from what they measure, not from their quality: organic rank or the rank users actually see on screen, median or weighted average.
- Three quarters of the CTR gap between two positions comes not from position, but from the kinds of keywords found at each position.
- To prioritize a roadmap, combine the CTR for the visual rank with search volume. The result is an order of magnitude for comparing keywords, not a traffic promise.
A brief aside on this summer’s SEO studies: Microsoft Clarity has added an “AI Scrape-to-Referral Ratio”, which compares the number of AI bot scrapes with the referral traffic they send back. Microsoft gives an example of roughly 6,000 scrapes for a single returned visit. This is a real issue, but not the one discussed here. An AI Overview capturing a click on google.fr and a crawler ingesting your pages to feed a chatbot are two different losses, measured with two different tools. This article covers only the first.
What published studies agree on
Here are the four most prominent references, side by side with measurements from Cadrank, Pied Web’s in-house tool. To make the “Scale” column reasonably comparable, I have expressed it as a number of keywords throughout. Three of the four sources publish that number. AWR mentions “millions of keywords across thousands of sites” without ever giving an exact figure, so that cell is my interpretation.
| Source | Position 1 CTR | Scale (keywords) | Period | Position measured | Aggregation |
|---|---|---|---|---|---|
| Backlinko | 27.6% | 12.2M queries | not specified | Search Console | not specified |
| seoClarity mobile | 9.0% | 11.8M | two Mondays one year apart | rank tracker (organic) | Σclicks / Σimpressions |
| AWR | 33.9% commercial, 23.4% informational | millions, across thousands of sites | rolling, updated monthly | Search Console | weighted, by SERP feature combination |
| Cadrank | 37.8% | 5,364 retained (keyword, site) pairs | rolling 12 months | visual rank, in pixels | median |
A word about that column, because the lack of comparability is already a symptom. Each source counts in the unit that suits it: seoClarity reports 750M impressions and 32M clicks, Backlinko 4M results, and I count (keyword, site) pairs. These units cannot be converted into one another. Even the only roughly shared figure, the number of keywords, is not published by all four sources. It ranges from millions to a few thousand, while one source remains vague. The main point is that Cadrank is three orders of magnitude smaller than the others. That is its principal limitation: the analysis uses data from around ten Search Console properties.
A fourfold difference, from 9% to 37.8%, for the same position on the same search engine. None of these measurements is wrong. They answer four different questions. Four variables change from one study to another, and none is trivial.
- The position measured: organic rank, average Search Console position, or the actual pixel position on the page.
- The denominator: Search Console impressions or search volume.
- The aggregation: a median across individual keywords, or total clicks divided by total impressions.
- The device: mobile and desktop have neither the same level nor the same slope.
The next two sections cover the two variables in this ratio. A third looks at the claims made from it. Together, they explain much of the noise surrounding AI Overviews.
The weighted average
The intuitive way to calculate average CTR for a position is total clicks divided by total impressions. It is the right answer to one question, and the wrong answer to the question in this article.
In Cadrank’s data, the two calculations differ by 32 percentage points in position 1: 70.3% weighted versus a 37.8% median. A handful of branded queries, with two orders of magnitude more impressions than the rest, causes the gap. Worse, the weighted column does not even decrease consistently. Rank 5 shows 11.4%, above rank 3 at 10.3%, simply because one high-volume keyword happens to sit there.
The flaw is inherent to the method and appears wherever it is used. Consider the mobile B2B segment in the seoClarity study, which contains around 40,000 keywords:
| Mobile B2B | April 2024 | April 2025 |
|---|---|---|
| Position 1 | 21.9% | 3.3% |
| Position 2 | 2.3% | 16.1% |
Read one row at a time, CTR in position 1 fell by a factor of 6.6 in a year. Read both rows, and it simply moved down one row. With Σclicks / Σimpressions, a single high-volume query moving from position 1 to position 2 is enough to produce exactly this table. The same artifact puts position 2 above position 1 in four of their segments. The most striking is desktop real estate: 15.8% in position 1 versus 31.6% in position 2, a twofold inversion. Next come desktop B2B (14.2% versus 21.6%), desktop retail (5.6% versus 9.0%), and mobile finance (7.4% versus 10.6%). No human behavior produces that pattern.
The study concludes that the largest proportional decline occurs in position 3, suggesting that even top results are losing visibility, probably because of AI Overviews, ads, and other SERP features. That is possible. It is also exactly what weighted aggregation produces in narrow segments. This table cannot distinguish between the two explanations.
Σclicks / Σimpressions answers, “How much traffic does this portfolio receive?” It is the only column whose total reproduces actual traffic. The median answers, “How much traffic does a keyword in this position receive?” That second question is the one asked by a traffic-potential calculation performed keyword by keyword.
Aggregation
To estimate the value of a position, the median of keyword-level CTRs is more appropriate. Total clicks divided by total impressions describes an entire portfolio, and a handful of branded queries can drive it.
Users see a different position from your rank tracker
The second variable is the most interesting. There are three ways to define “position 1”:
- organic rank: the first of the ten blue links, as reported by a rank tracker;
- average Search Console position: an average over the period, combining all devices and page layouts;
- visual rank: the height, in pixels, at which the result actually appears on the page.
Using exactly the same clicks and the same 5,364 (query, site) pairs produces three different curves:
| Rank | Visual rank (pixels) | Search Console position | Organic rank |
|---|---|---|---|
| 1 | 37.8% | 46.2% | 25.2% |
| 2 | 14.8% | 20.8% | 7.4% |
| 3 | 9.3% | 9.0% | 3.4% |
| 5 | 5.7% | 3.1% | 1.9% |
| 10 | 2.4% | 1.2% | 1.4% |
| 20 | 1.3% | 0.2% | 0.2% |
Fitted to the power law CTR = a × rank⁻ᵇ, they give:
visual rank -> 0.378 × rank^-1.21 R² 0.994
Search Console -> 0.472 × rank^-1.59 R² 0.950
organic rank -> 0.155 × rank^-1.11 R² 0.851
A quick explanation of this notation, because it appears throughout the rest of the article. A power law describes a decline whose rate does not depend on where you start: CTR is divided by the same factor each time the position doubles. With b = 1.21, moving from first to second divides CTR by 2.3, and moving from fifth to tenth divides it by 2.3 again. The coefficient a is the level, meaning CTR in first place. The exponent b is the slope, meaning how sharply the curve falls.
Organic rank does not fit well. R² is 0.851, and the model misses its own position 1 by ten percentage points: 15.5% predicted versus 25.2% measured. The reason is simple once you see it. Organic position 1 is not a place, it is a mixture. Depending on what Google stacks above it, such as ads, an AI Overview, People Also Ask, or a local carousel, the same “first result” may appear on the first screen or three scrolls lower. A mixture of power laws is not itself a power law. The relationship becomes more realistic when it uses the position the user actually sees.
Average Search Console position produces a much steeper curve, 1.59 versus 1.21, and a position 1 that is nine points higher. This makes sense: it is an average over time and is calculated only for queries where the result received an impression. It reads high at the top and collapses at the bottom.
Horizontal axis
Measure CTR against the position the user sees, not against organic rank. This is the only horizontal axis on which the relationship becomes clear, because organic position 1 is not a place. It is a mixture of visual ranks determined by what Google stacks above it.
One curve, two constants
One final methodological point surprised me more than any other when comparing sources. All these published tables follow power laws, yet none says so. Fit a × rank⁻ᵇ to the ten seoClarity rows and you get an R² of 0.94 on desktop and 0.98 on mobile, with similar results for their 2024 columns. Everyone publishes ten figures per segment. Nobody publishes two constants.
Here is Cadrank’s curve. The line is median CTR. The shaded area around it contains the middle half of the keywords, and matters just as much as the line.
The same figures, including dispersion:
| Visual rank | p25 | Median | p75 | Model fit |
|---|---|---|---|---|
| 1 | 24.6% | 37.8% | 61.2% | 37.7% |
| 2 | 7.5% | 14.8% | 24.8% | 16.3% |
| 3 | 4.1% | 9.3% | 17.4% | 10.0% |
| 4 | 3.2% | 7.7% | 17.0% | 7.1% |
| 5 | 2.5% | 5.7% | 13.3% | 5.4% |
| 6 | 2.1% | 4.8% | 10.8% | 4.3% |
| 10 | 1.0% | 2.4% | 5.2% | 2.3% |
| 15 | 0.7% | 1.3% | 2.7% | 1.4% |
| 20 | 0.4% | 1.3% | 2.4% | 1.0% |
CTR = 0.378 × rank⁻¹·²¹, R² 0.994 across 20 positions. The largest error is 1.6 percentage points at position 2. Every other error is below one point.
Focus on the shape and order of magnitude rather than the third decimal place, for a measured reason rather than mere caution. Depending on the sample cut, a falls between 0.348 and 0.419, and b between 1.18 and 1.25. The difference between two equally defensible cuts is larger than the difference between the reported values and either cut.
Two constants replace a table, and more importantly, they can be compared. This is the only clean way to describe how two curves differ: by their level, coefficient a, or their slope, exponent b. Cadrank has a higher level and a flatter slope than seoClarity mobile, 1.21 versus 1.43. This is consistent: because Cadrank uses visual rank, its position absorbs what an organic curve must absorb into its slope.
Reading between the lines, the segmentation that truly separates the curves is keyword length, and it changes only the level.
| Keyword length | Level a | Slope b |
|---|---|---|
| 1 to 2 words | 0.266 | 1.13 |
| 3 to 4 words | 0.400 | 1.27 |
| 5 words or more | 0.471 | 1.21 |
In first position, a five-word keyword receives about 1.8 times as much as a one or two-word keyword. The ratio narrows lower down the rankings, to around 1.5 at rank 10, because the slopes are not exactly identical, 1.13 versus 1.21. The curve’s shape therefore changes slightly, but far less than its level. This is useful in practice: one shared exponent and one coefficient per segment describe all three cases, instead of three complete tables. The model is lighter to publish and remains readable even when a segment becomes small, because that segment no longer has to support its own slope by itself.
Quartiles
In position 1, the middle half of tracked keywords ranges from 24.6% to 61.2%, a factor of 2.5. In position 10, it ranges from 1.0% to 5.2%, a factor of 5. A position predicts an order of magnitude, not a number, and uncertainty widens as you move down. A claim such as “CTR in position 1 changed from X to Y” is difficult to interpret.
Reading a descriptive curve as a promise of gains
Every published CTR curve, including Cadrank’s, is cross-sectional: it compares different keywords. Keyword A ranks first and receives 38%. Keyword B ranks seventh and receives 3.5%. The natural conclusion is that moving B from seventh to first would produce roughly ten times as many clicks. But that is not quite what the data says.
Search Console retains around sixteen months of history. Importing it produces a monthly series for each (keyword, site) pair. We can then compare a keyword with itself before and after a position change, using fixed effects for each keyword and month. Volume, intent, brand awareness, and the competitive environment are held constant. Seasonality can no longer be mistaken for a ranking effect.
| CTR decay exponent | across keywords | a keyword compared with itself |
|---|---|---|
| by Search Console position | 1.78 | 0.46 (95% CI: 0.43 to 0.50) |
| by Cadrank visual rank | 1.18 | 0.22 (95% CI: 0.19 to 0.26) |
The cross-sectional values in this table are refitted to the panel sample, which explains the difference from 1.21 in the curve above.
Between 74% and 81% of the slope described by the conventional curve does not come from position. It comes from composition: the keywords for which a site ranks first are not the same kind as those for which it ranks tenth. The keyword-length table above shows the scale of that difference. At the same rank, a five-word keyword receives 1.8 times as much as a two-word keyword.
The obvious objection is measurement error, because a noisy regressor flattens the exponent toward zero. The test is to retain only movements of several positions, and it does not support that objection: the exponents remain 0.46 and 0.22 whether the required movement is one position or four.
What I cannot do is choose between 0.22 and 0.46. The two axes disagree, so the causal exponent is bounded only within a factor of two, and neither sample is clean enough to settle the question.
The two curves are not competing. They answer two different questions.
- The cross-sectional curve answers: how much traffic does a result in this position receive? The mix of keywords is a feature here, not a flaw. It is the right curve for estimating the weight of a portfolio or assessing a competitor whose Search Console data you will never have.
- The within-keyword curve answers: how much would this keyword gain by moving up? The keyword is its own control, so the mixture disappears.
Using the first curve to answer the second question, for example by saying “move this page to position 3 and you will gain X visits,” overestimates the gain severalfold. Yet that is what almost everyone does, including me until this month.
Interpretation
A conventional curve answers “how much does a result placed here receive?” To answer “how much will I gain by moving up?”, each keyword must be compared with itself before and after its position changes. The real effect is two to four times smaller. Spoiler: I never go into this much detail.
The real cost of an AI Overview
From the curve’s perspective, an AI Overview first does something very concrete: it pushes the first organic result down the page. What does that cost? The question requires a causal exponent, not a descriptive one. That is exactly the distinction established in the previous section.
With the within-keyword exponent, falling one position costs between 14% and 27% of CTR, depending on the axis used. Falling two positions costs between 21% and 40%. For comparison, the cross-sectional curve predicted -57% and -73% for the same movements. The gap between these readings is exactly what the panel measures.
The published declines are -34.5% from Ahrefs, -38% in Seer’s latest update, and -47% from Pew, comparing 15% with 8%. They reach the top of the causal range, and Pew exceeds it. Simply moving lower on the page is therefore no longer enough to explain everything.
However, these studies are themselves cross-sectional. They compare a group of queries with AI Overviews with a group without them, or the same position from one year to the next. They therefore contain exactly the composition bias that the panel has just quantified in Cadrank’s data. Seer names this issue in its update without correcting for it.
I do not have enough data to provide a statistically robust figure. The conclusion here is more modest and more troublesome: before concluding that clicking behavior has collapsed, three effects unrelated to behavior must be removed.
- Stacking. The AI Overview pushes the result lower on the page. A study that measures CTR against organic rank cannot separate “fewer users click this result” from “this result is three scrolls lower.”
- Selection. AI Overviews appear preferentially on informational queries, whose CTR was already low before AI Overviews existed. Seer says so explicitly, and Amsive finds the same pattern: on branded queries, the presence of an AI Overview is associated with an 18.7% increase in CTR, compared with -20% on non-branded queries.
- Composition. Comparing one group of queries with another first measures the difference between those two groups. In the Cadrank corpus, this component accounts for between two thirds and five sixths of the apparent gap between two positions.
Once these three effects are removed, a distinct effect may remain, and it may be substantial. The right setup tracks the same keywords before and after an AI Overview appears on their SERP, with each keyword serving as its own control. This requires no special tool: your Search Console data is enough, together with a record of your site’s pixel position. Cadrank has had the necessary tracking in place for a few days, and a publication will follow.
From CTR to traffic potential
traffic potential = CTR(visual rank of the first organic result) × average monthly volume
The first half comes from the SERP: the pixel position of the first organic result and the curve above. The second comes from Google Ads Keyword Planner. I examine the respective value of these two data families in my article about search volume versus Search Console impressions.
There are four precautions to take, because a metric you cannot interpret is worse than having no metric at all.
Smooth the data. Cadrank divides the page into 185-pixel bands to convert height into rank. A first result that oscillates around a boundary because an ad block is 30 pixels taller this week changes rank and sees its traffic cut in half, even though nothing meaningful changed in the SERP. Cadrank therefore smooths the last four screenshots, weighted 4/3/2/1 by recency. Today’s screenshot retains 40% of the weight, allowing a real change, such as an AI Overview that remains in place, to pass through within a week rather than a month.
Remember that the two halves of the calculation have different denominators. CTR is measured in clicks per Search Console impression, while volume counts Keyword Planner searches. These counters do not cover the same thing, and their ratio deteriorates as you move down the page. Their product is therefore an upper bound, and the lower the position, the looser that bound becomes.
A Keyword Planner “keyword” is a basket of close variants. The displayed volume therefore exceeds the actual demand for any single query, and the larger the basket, the larger the overstatement. This can be measured: when only high-volume keywords are retained, the ratio of actual impressions to reported volume falls monotonically. I described this search-volume flaw in 2025.
Be wary of the bottom of the scale, where the metric starts working against you. Keyword Planner floors low values just as it inflates high ones. 0 is the most frequent volume in my entire database. Among keywords rated at 10 or less, 31% actually receive more impressions than their stated volume, sometimes by three orders of magnitude. In my corpus, keywords with volume at or below 10 account for nearly one third of actual clicks, even though calculated traffic potential classifies all of them as “ignore.”
Treat traffic potential as a comparative metric. It helps rank keywords within a roadmap, compare two competitors over the same scope, and decide whether a topic deserves a new article or a page overhaul. It cannot predict the number of SEO visits, and especially cannot quantify what a ranking improvement would deliver. That question belongs to the much flatter within-keyword curve. I quantify it in another metric, Adjusted Potential.
Measure your own curve
You can repeat the exercise with your own data, without any special tool. Export queries from Search Console, under Performance, Queries, including the Position column. For more than 1,000 rows, use the API. Then:
- round position to the nearest integer;
- for each position, calculate the median query-level CTR, never total clicks divided by total impressions;
- set a minimum number of impressions to exclude noise, around 50 over twelve months is sufficient;
- fit
log(CTR) = log(a) - b × log(rank)with linear regression, a spreadsheet is enough; - compare your two constants with Cadrank’s.
There is one major caveat: you will measure against average Search Console position, not visual rank. This corresponds to the 0.472 × rank⁻¹·⁵⁹ curve in the table above, the steepest of the three. Its slope is partly an artifact of how Google averages position. That is exactly why Cadrank measures position itself, screenshot by screenshot.
If you want visual rank without building your own extraction, Monitorank is one of the few services that reports pixel position. If you track rankings there, joining its pixel position from the API with your Search Console export gives you the correct horizontal axis, the one on which the relationship becomes clear.
The before-and-after study your data already supports
The same export can answer the question in the title for your website. Comparing a keyword with itself is exactly what the panel above does, and the change being compared does not have to be a position change. A change in the SERP works just as well.
- Retrieve twelve months of query history, using one API call per month.
- Retain only queries whose average position changed by no more than half a rank over the period.
- Split them into two groups: those with an AI Overview today and those without one. This is the only information Search Console will not provide. You must obtain it from a rank-tracking tool or collect it manually for a sample.
- Plot median CTR month by month for each group.
If the group with AI Overviews declines while the control group remains flat, you have an answer for your own queries without depending on anyone else’s sample.
A few points to note:
The control group is not optional. Without it, you measure the general drift in CTR over twelve months. That drift is not zero, and AI Overviews are not its only cause.
A constant Search Console position does not mean a constant visual position. That is precisely what makes the setup interesting, since an AI Overview pushes you down the screen without changing your organic rank. However, a SERP that gains two ads produces exactly the same signal.
AI Overviews today, in five findings
For French SERPs, I have no figure ready to publish, and I have not seen a robust one elsewhere. What I do know is:
- much of the published data relies on an aggregation controlled by a handful of queries, to the point that position 2 outperforms position 1 in the authors' own tables;
- measuring CTR against organic rank mixes the behavioral effect with the stacking effect, even though stacking is the first thing an AI Overview changes;
- in Cadrank’s data, between 74% and 81% of the CTR gap between two positions comes from the nature of the keywords found there, not from position;
- the causal effect of falling one position is therefore two to four times smaller than the conventional curve suggests, and slightly less so for two positions. It no longer explains the published declines on its own;
- the most widely quoted figure lost one third of its magnitude between two publications by the same author, yet almost nobody repeated the correction.
The limitation of Cadrank’s measurement is that its 37.8% comes from ten French websites, not a representative sample of the web. The curve applies to those ten websites. It is not a universal benchmark.
At the current pace of AI Overview rollout, a few more months of French screenshots will be needed to settle the question. Until then, the recommended approach remains unchanged: record the position users see, and read medians with their quartiles rather than averages. As soon as the question becomes causal, “how much will I gain by moving up?”, switch curves and compare each keyword with itself. Traffic potential from the first curve remains an order of magnitude, which is exactly what is needed to prioritize a roadmap.
If the question is urgent, do not wait for me. The before-and-after method described above works with your own queries and will always be more relevant to your website than the average from a US corpus.
I will republish this curve when there is enough AI Overview data.
AI, SEO and Web strategy session
Are your SEO numbers telling the right story?
Read your Search Console data in context to separate a real trend from measurement bias and identify an actionable priority.
€350 excl. VAT, up to 90 minutes
Continue reading: