44 percent of impressions gone: why I optimised nothing

On 31 July my Search Console showed a 35. Thirty-five impressions for an entire day, after weeks of five-digit figures, and for two seconds I genuinely thought the domain had dropped out of the index overnight.

It had not. The most recent day in Search Console is generally incomplete and keeps filling up with trailing data over the following two to three days. A classic. An expensive classic, if someone builds a meeting on it.

The look at the preceding weeks was considerably less pleasant. There was a real break on 10 July, with an average of a good 21,000 impressions per day before and just under 12,000 after. At first that looked to me like a penalty.

Back in April I took apart the GSC impressions bug here and wrote quite plainly that teams make wrong decisions on the basis of broken numbers. Now I was sitting in front of such a curve myself. So I optimised nothing and measured first instead, and this article is the test scheme that came out of it.

I compared two windows of equal length, 15 days each, directly before and directly after the break. All the figures here come from a single property and are an observation, not general evidence.

The impressions almost halve, the clicks barely move and the CTR rises by more than 60 percent. In a real visibility loss, precisely that should not happen.

If the domain had genuinely slipped, fewer people would see the results and click accordingly, and both curves would have to run downwards in parallel. In my case they do not. That suggests something is disappearing from the count here that practically never got clicked before.

Google defines the impression in its official help remarkably briefly: “How many times a user saw a link to your site on Google.”

The decisive part sits in the word “saw”. It is entirely sufficient that your result was delivered on the search results page, and a script scraping that page produces exactly the same impression as an interested reader. The metric does not require any actual user contact.

A click works fundamentally differently, because someone actively decided. That makes the click curve the calmer and more honest figure.

Anyone who starts rewriting content and swapping titles the moment impressions fall ends up repairing a display. That costs weeks and changes nothing at all about the cause.

In Search Console you group the performance report by device, and after two clicks you can already see whether your drop affects both worlds at all.

Desktop loses almost half of its impressions, mobile loses little, and on both devices the clicks fall only in the single-digit to low double-digit range.

Rankings can certainly differ between desktop and mobile, but they rarely do so by this factor, and certainly not without taking the clicks along. A one-sided desktop drop with stable clicks points to the counting, not to the assessment.

One possible explanation is that part of the affected queries is rather desktop-heavy. If such a source falls away, the impression count drops precisely there, without a single person fewer reaching your site.

External CTR benchmarks by position are currently worth remarkably little, because the values fluctuate strongly by query, device and the presence of an AI answer. You do not need them either, because your own domain provides the better yardstick.

So I took all pages from the period before the break that sat at a similar average position. Same domain, same window, same measurement.

Three pages at positions 5.6 to 6.9 sit between 1.7 and 3.7 percent, while the fourth page at position 7.3 collected more than 27,000 impressions and produced all of three clicks from them. That corresponds to a CTR of 0.01 percent and is therefore roughly one hundred and fifty times below the weakest comparison page. Three comparison pages are, however, a small sample; this here is a plausibility check and not proof.

Even so, a gap like that is hard to explain by chance. For a page that sits stably on page one and receives practically no clicks, that points rather to these not being typical user impressions. That very page lost around 85 percent of its impressions after 10 July, at a position that stayed constantly around 8.

Position stable, impressions gone. That combination is compatible with a correction in the counting and, in my case, fits a classic ranking loss less well.

Almost every domain has at least one steady performer, and in my case that is an explainer on an abbreviation which has served the same demand consistently for months and therefore acts as my control group.

Over the same comparison window it rose from 37,844 to 39,824 impressions and improved its average position from 7.9 to 7.1. So it gained while the domain as a whole lost 44 percent.

That makes a domain-wide problem very unlikely. If domain-wide problems were the cause, I would have expected effects there as well.

The counter-test then shows where the vanished impressions actually sat.

According to Google’s status dashboard the June 2026 spam update started on 24 June 2026 and was done after a good two days, and my article about it shot up and fell away just as quickly afterwards. That need not be a penalty; it fits the end of a news cycle better.

Anyone who writes a lot about Google updates builds up an impression base that breaks away on schedule and is simply part of the format. How strongly such phases show up in the measurement tools is something I put into context in my guide to interpreting volatility sensors.

A single finding withstood all three checks. My foundational article on crawling and indexing slipped on 9 and 10 July from position 11 to 13 down to 15 to 18, and its impressions halved in the same move from 15,621 to 8,200. It had not recovered by the end of July.

In my case that points rather to real movement, because position and impressions fall together here, the effect persists over weeks, and it affects a page with a real share of clicks.

This is exactly where the work pays off. Not on the 27,000 impressions which, going by my pattern, look more like a measurement signal than a user signal, but on a single URL that genuinely lost four positions and is not getting them back.

That is the real yield of the exercise. A panicked “44 percent gone” turns into a concrete task on exactly one URL.

The obvious move would be to attribute the drop to the known logging error. That would be convenient and probably wrong.

Google’s data anomalies page describes the problem like this: “A logging error prevented Search Console from accurately reporting impressions from May 13, 2025 until April 27, 2026.” That window therefore ended on 27 April 2026, and my break sits two and a half months after it.

For July 2026 there is simply nothing to be found in either official source. The status dashboard lists the June 2026 spam update of 24 June as the most recent confirmed update, and the anomalies page likewise ends at that date.

What there was were unconfirmed reports. Several volatility trackers registered movement around 11 and 12, 18 and 19, and 23 and 24 July. Search Engine Roundtable noted: “Google Search showed heavy ranking movement over the weekend of July 18 and 19, 2026, and as of July 21 Google has not confirmed any core or spam update behind it.”

My break sits at the start of that window. That is temporal proximity and not proof.

Three events have shifted this figure in quick succession.

First, in mid-September 2025 Google switched off the num=100 parameter and thereby removed a convenient way to retrieve a hundred results in a single query. An exact date is not documented by Google; the trade press names 8 to 14 September. What that means for the data situation is in my article on the num=100 shutdown.

Second, the logging error. Google phrases this strikingly cautiously and says merely that the impressions were not recorded correctly, while the widespread reading of an inflation comes from the reporting and not from Google’s own text. According to Search Engine Land the following also applies: “John Mueller from Google confirmed on Bluesky that this is only fixed going forward and the old data will not be fixed.” Any year-on-year comparison touching that window therefore stays permanently unreliable.

Third, the structure of the search queries themselves is changing. In my own query reports, very long, sentence-like queries now appear that look more like machine generation than human typing behaviour. I cannot prove that, and it is an observation on a single property. It is striking nonetheless.

None of this amounts to a recommendation to ignore impressions in future. As a directional indicator of demand they still work, but as a target figure that budgets or evaluations hang on, they have simply become too unstable.

The impression tells you that something was delivered, and the click tells you that someone decided in favour of it. The larger the share of automatically generated queries becomes, the further these two figures drift apart.

That is why the right response to a collapsing impression curve is not activism. It is a measurement.

And sometimes the result stays unsatisfying. To this day I do not know what actually happened on 10 July, but I now know fairly precisely what did not happen, and that was by far the more expensive question.