- Position 1 with an AI Overview read 1.6% at Ahrefs and 21.4% at Advanced Web Ranking for the same month and device. [1][2]
- Google says impressions, CTR, and average position were misreported from May 13, 2025 to April 27, 2026. Clicks were not affected. [6]
- A randomized test found 39.8% fewer organic clicks on queries with an AI Overview, on desktop. [9]
- Applying a published haircut to a 2026 baseline double-counts, because today’s clicks already include today’s AI Overviews.
- In our modeled sample, clicks land between +2.4% and -21.8%. One input, how often informational queries show an AI Overview, drives most of that range.
- A surviving click must be worth 2x at a 50% loss for quality to make you whole. The best published figure is 1.53x, from a vendor. [17]
Every budget season, an SEO lead opens a spreadsheet and multiplies search volume by a click curve. So give that spreadsheet one keyword, 100,000 searches a month, and a page in position 1, and ask how many visits it gets. The answer depends entirely on whose curve you paste in.
Those bars come from First Page Sage [3], Sistrix [4], Backlinko [5], AWR [1], and Ahrefs [2]. That’s a 25-fold spread, and it isn’t only age. AWR’s own pages even label that baseline two ways, as all results pages and as pages without an AI Overview. For December 2025, on desktop and from Search Console data both times, Ahrefs reports 1.6% and Advanced Web Ranking reports 21.4%. [2][1] Inside one method, AWR’s own position 1 with an AI Overview went from 25.58% in January 2026 to 1.96% in April. [1] First Page Sage tells readers its blended column “is the one to use for traffic forecasting.” [3] That column is 7.1%, and it’s exactly what you get by assuming an AI answer on 81.6% of queries. At Pew’s 18%, from real browsing in March 2025, the same arithmetic gives 19.2%. [11]
Why they disagree
None of these firms is lying. They’re measuring different things and calling all of them position 1.
| Driver | What differs |
|---|---|
| Sample | Millions of US results pages at AWR, 300,000 chosen keywords at Ahrefs, 1,184 client and panel domains at First Page Sage, 53 brands at Seer, where one account is 47% of a segment [1][2][3][12] |
| Query mix | Ahrefs compares AI Overview keywords with informational ones. AWR includes brand, where position 1 reads 27.3% against 19.4% for non-brand [2][1] |
| Position | Search Console records the topmost link, and every link in an AI Overview shares the AI Overview’s position. So a page ranking sixth but cited can log position 1 at a low CTR, by our reading of Google’s rules [15] |
| Date | The same method moved 13-fold in 12 months at AWR, and First Page Sage’s with-AI position 1 differs 3.2-fold between two of its reports five weeks apart [1][3][17] |
| A broken denominator | See below |
The data break nobody footnotes
This is the part that should change your plan this week. Google’s data anomalies page says a logging error prevented Search Console from accurately reporting impressions from May 13, 2025 until April 27, 2026, and that CTR and average position were affected along with them. Clicks weren’t. [6] John Mueller said the old data won’t be fixed, according to Search Engine Roundtable. [7]
Search Console keeps 16 months. [8] So today, every retained day before April 28, 2026 sits inside the error, and you won’t have a clean year of CTR until May 2027. The Ahrefs reading from December 2025 sits inside the window, while its December 2023 comparison sits outside it. Seer, AWR, and First Page Sage all published after Google’s notice, and the study pages we read from them don’t mention it.
You also have a second broken window. Nozzle’s work on poisoned results pages suggests AI Overview labels collected by automated tools since September 13 can’t be taken at face value, and its share of affected pages climbed again on October 21. [13] That leaves April 28 to September 12, 2026 as the period when both your CTR and your labels are fairly trustworthy: 138 days, with no seasonality to speak of. Clicks remain the one series Google vouches for across the break, so use clicks for year-over-year work.
The anchor: about 40%, not 58%
When vendor data is this compromised, look for evidence that doesn’t depend on it. Agarwal and Sen ran a randomized field experiment with 1,065 US participants on desktop Chrome, in January and February 2026, hiding AI Overviews for some. Organic clicks per search on triggered queries were 0.37 with an AI Overview and 0.62 without, a reduction the authors put at 39.8%. Across all searches, the reduction is about 18.5%. [9][10] It’s an unreviewed working paper, on desktop only, and the effect was significant only on informational queries.
Other reads land nearby, though they aren’t like-for-like: Pew’s 9% of visits with any click against 15% [11], and Seer’s February 2026 gap of 38%. [12] Position 1 numbers run higher, which fits the experiment’s finding that the effect concentrates in the top positions. A second, preregistered experiment shows how fragile this kind of test is: when its researchers tried to hide AI Overviews, a change in Google’s page code broke the treatment partway through, and the intent-to-treat effect wasn’t significant. [18] So hold 40% as a working anchor rather than a law.
The mistake in most 2027 models
Here’s the double count. People take a 2026 baseline and subtract 40%, or 58%, from the part exposed to AI Overviews. But your 2026 clicks already contain your 2026 AI Overviews. What you’re forecasting is the change from today’s exposure to next year’s, so you have to divide the old effect out before you apply the new one.
Clicks next year = clicks today × (1 - next year’s share × next year’s suppression) / (1 - today’s share × today’s suppression)
Share is how often the segment’s queries show an AI Overview. Suppression is how many clicks you lose when one appears. Run it by segment, then add up.
In our sample, subtracting 58% from today’s informational clicks gives 252,000. The formula’s base case gives 537,805 from the same 600,000. The inputs below are modeled, and each cell comes from a named study or is flagged under the table. Treat the range as a disagreement between studies. It isn’t a probability distribution.
| Segment (share × suppression) | Today | Low exposure | Base | High exposure |
|---|---|---|---|---|
| Informational | 36% × 50% | 36% × 39.8% | 53% × 50% | 85.9% × 58% |
| Commercial | 8% × 50% | 8% × 39.8% | 11.8% × 50% | 23.4% × 70.5% |
| Transactional | 5% × 24.9% | 5% × 0% | 6% × 24.9% | 10% × 48% |
| Brand and navigational | 15% × 0% | 15% × 0% | 15% × 0% | 15% × 39.8% |
| Clicks, sample site | 1,350,000 | +2.4% | -5.1% | -21.8% |
| Revenue, sample site | $5,760,000 | +0.9% | -1.2% | -8.9% |
Shares come from Seer’s accounts [12], the Agarwal and Sen sample [10], and First Page Sage [17]. Suppression comes from the experiment [9], AWR’s planning rule and its July 2026 readings [1], Ahrefs [2], and Seer [12]. The base commercial share is our derivation, the transactional high of 10% is Seer’s watch line rather than a measurement, and the 0% cells reflect effects that weren’t significant in the experiment rather than a measured zero. The sample site’s traffic and revenue are invented.
Now try your own numbers. This version moves only the informational segment, which in our sample produces most of the range: against the base case, presence alone moves total clicks by +4.9% to -9.4%. Everything else moved the total by under 1.5%, apart from informational suppression at +3.1% to -2.4%.
Your clicks next year, as informational exposure changes
The defaults are the sample site’s base case. Replace them with your own Search Console figures.
Modeled, not measured. It holds demand, position mix, and every other segment flat, so rank gains go in as a separate line using your own CTR by position.
Calibrate it from your own data
Export, and use clicks for trends
Bulk export holds everything except anonymized queries. [19] Use CTR only from April 28, 2026, and keep the anonymized gap as its own line.
Segment with one rule of your own
Take brand from Google’s filter, then classify intent yourself. Never borrow a presence rate from a source using a different classifier. Seer labels intent by results-page features, and in its informational set “price, cost, buy” queries triggered AI Overviews 83.4% of the time. [12]
Label AI Overviews repeatedly
Record a share of observations per query rather than a yes or no. Seer’s own label is a static snapshot applied backward across all months, and it says so. [12] Discount reads after September 12. [13]
Measure suppression on queries that flipped
Compare queries that gained an AI Overview after April 28 with those that didn’t. Comparing all queries with and without one is biased, because Seer notes AI Overviews tend to surface where clicks were already lower. [20]
Backtest and publish the error
Fit on the second quarter, predict the third, and show the miss. Seer’s own model missed organic CTR by 0.6 to 2.6 percentage points on rates of a few percent. [12] Your bands should be that honest.
Search Console’s generative AI performance report helps less than its name suggests. It reports impressions by page, country, device, and date, with no clicks and no query dimension, and only for your own links. [14] Use it to see which pages get shown. Don’t use it as a measure of AI Overview presence, and don’t add it to Web impressions, because it’s drawn from the same data.
Better clicks won’t save the forecast
Google describes the clicks that remain as higher quality, by which it means people don’t quickly click back, and it hasn’t published the data behind that. [16] We told readers the same thing in our July piece, more flatly than the evidence now allows. The randomized experiment found no significant difference in returning to search, bounces, or time on page, and its authors call this suggestive. [10]
The arithmetic is unkind anyway. To offset a loss, a surviving click must be worth 1 / (1 - suppression) times as much: 1.66x at 39.8%, 2x at 50%, 2.38x at 58%. The only with-and-without conversion figure we found comes from First Page Sage, at 2.6% against 1.7%, or 1.53x. It isn’t randomized, and the firm’s own net lead figure falls about 38%. [17]
Watch the ratio as well. In our sample, revenue per click rises 16.4% in the high scenario while revenue falls 8.9%, with no click getting better, simply because informational clicks convert least. Report revenue. Revenue per visit can hide the loss.
The page to bring
- Today’s clicks by segment, with the anonymized gap shown
- Low, base, and high scenarios, each assumption with its source next to it
- Revenue beside clicks, using your own conversion rate by landing page
- What’s unusable and why: CTR before April 28, 2026, and tool labels after September 12 without the vendor’s error rate
- The triggers for re-forecasting: informational AI Overview share passing your base value, transactional share above 10%, or clicks outside the range for two months
A finance lead will trust a wide range with its assumptions in view more than a precise number that came from a blog’s table. And if a single click curve is the only thing behind your plan, the question to ask is which of the eight bars above you’ve been using.
Sources
- Advanced Web Ranking · Google organic CTR by position: 2026 studymonthly position 1 with and without AI Overviews, US, Search Console data, through July 2026
- Ahrefs · Update: AI Overviews reduce clicks by 58%December 2025 data, published February 2026, desktop, 300,000 keywords, correlational
- First Page Sage · Google click-through rates by ranking positionMarch to August 2026, 1,184 domains, agency data
- Sistrix · Why almost everything you knew about Google CTR is no longer validmobile only, July 2020
- Backlinko · We analyzed 4 million Google search resultsabout 4 million results, page last updated April 2025, data date not stated
- Google Search Console Help · Data anomalies in Search Consolethe logging error from May 13, 2025 to April 27, 2026, and what it touched
- Search Engine Roundtable · Google Search Console fixes 50 week data logging issueJohn Mueller on old data not being fixed, read through trade press
- Google Analytics Help · Google organic search traffic reportSearch Console keeps data for the last 16 months
- Agarwal and Sen · The impact of Google AI Overviews on publisher traffic and user experience, abstractrandomized field experiment, working paper
- Agarwal and Sen · Full draft, June 17, 2026third-party copy of the working paper, with the 18.5% figure, sample, trigger rates, and engagement results
- Pew Research Center · Do people click on links in Google AI summaries?900 US adults, March 2025
- Seer Interactive · AIO impact on Google CTR, 2026 update53 brands, presence by intent, its own forecast miss, and method caveats
- Nozzle · Poisoned Google SERPsvendor study, updated through October 21, 2026
- Google Search Console Help · Generative AI performance reportimpressions only, no query dimension, all sites from August 31, 2026
- Google Search Console Help · What are impressions, position, and clicks?how AI Overview links are positioned and counted
- Google · AI in Search: Driving more queries and higher quality clicksGoogle’s own claim, with no data published
- First Page Sage · AI impact on website traffic report218 client sites, not randomized, with the only conversion comparison we found
- Wang and colleagues · AI in Search reduces publisher referrals without improving user experiencepreregistered preprint, 1,100 US participants, March 2026, showing that hiding AI Overviews is fragile as a treatment
- Google Search Console Help · About bulk data export of Search Console data to BigQueryeverything except anonymized queries
- Seer Interactive · AIO impact on Google CTR, September 2025 update3,119 informational terms across 42 organizations, and the selection point
Frequently asked questions
Why can’t I use a CTR by position curve for my 2027 SEO forecast?
Because the published curves no longer agree closely enough to pick one. For December 2025, on desktop and from Search Console data in both cases, Ahrefs reports 1.6% for position 1 on its AI Overview keywords and Advanced Web Ranking reports 21.4%. Build the curve from your own data instead, split by segment and by whether an AI Overview appears.
How much do AI Overviews really reduce clicks?
The most direct randomized field experiment we found, by Agarwal and Sen, reports 39.8% fewer outbound organic clicks on queries that trigger an AI Overview, and about 18.5% fewer across all searches. It covers 1,065 US desktop Chrome users in early 2026 and is an unreviewed working paper. Vendor figures for position 1 run higher, at 52% and 58%.
Is the Ahrefs 58% figure from December 2025 or February 2026?
Both, for different reasons. The data is from December 2025 and Ahrefs published it on February 4, 2026. It is desktop only, covers position 1, and is correlational. Its comparison group of informational keywords with no AI Overview also fell 48.6% over the same two years.
Can I trust my Search Console CTR from 2025?
Not for trends. Google says a logging error prevented accurate impression reporting from May 13, 2025 until April 27, 2026, and that CTR and average position were affected while clicks were not. A Google representative said the old data will not be fixed, according to Search Engine Roundtable. Search Console keeps 16 months, so clean CTR for a full year arrives only in May 2027.
Do AI impressions in Search Console help me forecast?
Partly. The generative AI performance report shows impressions by page, country, device, and date, and it shows which pages appear. It has no clicks and no query dimension, and it counts only your own links, so an AI Overview that does not cite you is invisible to it. It is not a measure of how often AI Overviews appear on your keywords.
Are the clicks that remain after AI Overviews better?
The evidence is weak. The randomized experiment found no significant difference in engagement, and its authors call that result suggestive. The one with-and-without conversion figure we found comes from First Page Sage, at 2.6% against 1.7%, a 1.53x difference. At a 50% click loss, a surviving click would need to be worth 2x to make you whole.
What should I show the CFO instead of a single traffic number?
One page: today’s clicks by segment, a low, base, and high scenario with every assumption and its source visible, revenue next to clicks, and the events that would trigger a re-forecast. Say plainly that the range is modeled and that it is a disagreement between studies rather than a probability.
Mikołaj Salecki
Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
