Impressions up, clicks down is the single most common thing a client has shown me on a screen share since the middle of last year. The chart is always the same shape. Impressions climbing in a confident diagonal, clicks flat or sagging underneath it, and a conclusion already written in the room before anybody asks what either line is counting.
The conclusion is always AI. Sometimes it is said carefully, sometimes it is a slide titled the zero-click era. Either way the evidence is that gap, and the gap is calculated from two numbers that Google now says were not measured to the same standard.
On April 3, 2026, Google updated its Search Console data anomalies page to acknowledge a logging error that over-reported impressions. The error began on May 13, 2025. Correct logging resumed in late April 2026. Clicks were not affected. The historical data inside that window will not be rebuilt.
Fifty weeks of an inflated numerator sitting under an accurate one. That is not a footnote to the AI story. It is a measurement problem that sits underneath every click-through rate anybody calculated in that period, and it points the error in the opposite direction to the one the industry assumed.
What Google logged wrong for fifty weeks
Google's Search Console logging error over-reported impressions across properties from May 13, 2025 until correct logging resumed in late April 2026, a window of roughly fifty weeks that Google disclosed on its data anomalies page on April 3, 2026.
Four details in that disclosure decide how much of your reporting history survives, so they are worth separating out.
| DETAIL | WHAT GOOGLE STATED | WHAT IT MEANS FOR YOUR ACCOUNT |
|---|---|---|
| Affected metric | Impressions only, described as a data logging issue | Impression counts, and anything derived from them, are suspect inside the window |
| Unaffected metrics | Clicks and the other performance metrics were not affected | Clicks are the one continuous, trustworthy series across the whole period |
| Direction of the error | Impressions were over-reported, not under-reported | Real impressions were lower than shown, so real CTR was higher than shown |
| Window | May 13, 2025 to April 27, 2026, with logging correct again from late April | Year over year impression comparisons spanning those dates compare two different measurements |
| Retroactive fix | The historical data will not be reconstructed | You cannot repair the window, you can only label it and stop citing it |
| Magnitude | Not published | Nobody can state how wrong their own numbers were, including us |
The magnitude row is the one that constrains everything that follows. Google acknowledged the direction of the error without quantifying it, which rules out the obvious fix. You cannot apply a correction factor to the bad window, because no correction factor has been published. Any agency offering you a restated eighteen month CTR trend is inventing the restatement.
What you get instead is a boundary. Before May 13, 2025, impressions were logged correctly. After April 27, 2026, impressions are logged correctly. Between those dates they were logged high by an unstated amount, and the honest treatment of that span is an annotation, not an adjustment. Search Engine Land's write-up of the Search Console bug that inflated impression counts has the disclosure timeline if you need to put a source in front of a stakeholder.
Why impressions went up while clicks went down
Impressions rose and clicks fell for at least four independent reasons between 2025 and 2026, only one of which is AI answering queries above the organic results, and one of which was a logging error inside Google itself.
Treating that divergence as a single phenomenon is where the analysis usually goes wrong. The causes have different fixes, different owners, and different levels of evidence behind them, and a chart cannot separate them for you.
Those four causes are not competing theories. They are concurrent, they overlap, and they stack in the same direction on the same chart. That is exactly why the chart is a poor instrument: four additive causes with one visible symptom, and no way to apportion them without going query by query.
The practical consequence is a sequencing rule. Rule out the measurement causes first, because they are free to check and they contaminate everything downstream. Only then argue about the behavioural ones, which is where the interesting work is. We make the same argument about the new AI reporting surfaces in what the Search Console generative AI report actually measures, and about whether the multimodal report is new data or moved data. A new metric in a familiar interface is the easiest place in this job to mistake a definition change for a performance change.
The CTR collapse everyone measured had a broken denominator
Click-through rate is clicks divided by impressions, so over-reported impressions produce an under-reported CTR, which means the CTR decline measured during the bug window overstates the real decline by an amount Google has not disclosed.
The arithmetic is not subtle, and it is worth walking because the direction surprises people who have been reading about AI eating their clicks for a year.
Take a page reported at 100,000 impressions and 2,000 clicks. Reported CTR is 2.0%. Now suppose the true impression count was 80,000, because clicks were logged correctly and impressions were not. True CTR is 2.5%. The page performed 25% better than the dashboard said, and nothing about the page changed. Those figures are illustrative, chosen for round arithmetic rather than drawn from any disclosed inflation rate, because no inflation rate has been disclosed.
Now the uncomfortable half of this, because the argument gets abused in both directions. A broken denominator does not mean the click loss was imaginary. It means one specific measurement of it was unreliable, and the better evidence never came from Search Console in the first place.
Clickstream research is the place to look, because it observes behaviour rather than reported impressions. The Bocconi University paper by Qiaoni Shi, Kai Zhu and Kai Gu, published to arXiv on July 8, 2026, reconstructed browsing sessions from URL-level Comscore desktop records across more than 45,000 United States households. Wider access to ChatGPT Search cut traditional search queries by 9.4% on average, deepening to 17.0% after twenty weeks of exposure. Separately, work by Wang and colleagues in August 2026 found AI Mode reduced external clicks by 18.8 percentage points.
Hold those two findings next to the bug and the picture resolves. Real click loss, measured outside Google's reporting. A reported CTR decline that was partly an artifact of Google's reporting. Both true, and the second one is the reason so many teams over-attributed the first.
Decline in referral traffic by destination type after wider ChatGPT Search access, indexed to the largest fall at 32.8% (Bocconi University, Answering Without Referring, July 8, 2026)
The spread across those categories is the finding worth carrying into a planning meeting. Referral loss is not uniform, it tracks how answerable a page's job is. Pages that exist to state a fact lose most. Pages that exist to let somebody do something, compare something or buy something lose least, and a site-wide average hides which of those two you are.
Record query volume and a 9.4% decline can both be true
Google reporting record query volumes and a clickstream study finding a 9.4% fall in traditional queries are not contradictory claims, because the two are counting different events over different populations with different definitions of a query.
At its developer conference in May 2026, Google said query volumes hit a record and that people who use its AI features in Search use Search more. The company reported AI Overviews passing 2.5 billion monthly active users and AI Mode passing 1 billion in its first year. Google has not released the user-level data behind the query claim, has not stated a baseline, and has not defined what counts as a single query inside a conversational AI Mode exchange.
“A conversational exchange that fans out into eight retrieval steps can be counted as eight queries or as one. Until the counting rule is published, record query volume is a statement about an internal metric, not a number you can reconcile against your own.”
The Bocconi figure counts something narrower and better specified: traditional search engine queries issued by observed households, before and after wider ChatGPT Search availability. One number can rise while the other falls with no dishonesty anywhere, and the reconciling fact sits in a third study. Similarweb reported in May 2026 that 95% of ChatGPT users also use Google.
People did not migrate. They added a tool, moved some question-shaped work onto it, and kept Google for the rest. Total question volume up, traditional query volume down, and referral loss concentrated wherever a page's only job was to answer. The single most quoted number from the Bocconi work makes the mechanism plain: ChatGPT produced a clean outbound referral in 5.2% of conversation sessions, against 31.1% of Google queries. Search Engine Journal's read on this, that people are not leaving Google for AI but using both, is the right frame.
Which leaves the attribution problem, and it is not solved by any of this. Assistant referrals arrive thinly and often without a referrer header at all, so the visits that do come through are undercounted at the door. We went through the specific failure modes in the limits of the GA4 AI assistant channel, and the same caution applies to reading AI influence through paid signals, which we covered in AI traffic attribution and the smart bidding signal.
What to do when impressions are up and clicks are down
The correct first response to an impressions up, clicks down chart is to establish which part of it is measurement and which part is behaviour, before any strategy conclusion is drawn from the gap between the two lines.
That takes about a day of work in a normal account, and it is the cheapest analytical work available this quarter because it mostly consists of removing bad inputs rather than finding new ones.
The thread running through all four is a preference for metrics that count events over metrics that count opportunities. Impressions are a count of opportunities, which is why they are so easily inflated by both a logging error and a wider query mix, and why they were never a strong basis for the conclusion the industry built on them. Clicks, revenue and verifiable citations count things that happened. Setting that hierarchy up properly is most of what we do in reporting and analytics engagements, and it is the first thing we check in an SEO and GEO audit, because a measurement layer nobody trusts makes every argument above it unresolvable.
None of this is a reason for relief about AI and organic traffic. The independent research is clear that referral loss is real, concentrated and larger on exactly the informational pages most content programs are built from. The correction is narrower than that and more useful: one specific number was wrong in one specific direction for fifty weeks, and a great many conclusions were sized against it.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.