I opened my own search figures on Monday expecting a quiet five minutes, and got the best-looking number I have seen all year. Impressions up fourteen thousand on the month before. A cluster of searches where I had arrived on the first page of Google. If a client had sent me that I would have written back to say congratulations.
Then I looked at the column next to it, which said nobody had clicked. Not a few. None.
Every one of us has been trained to read a rising number as good news, and for most of the history of search that training was sound. A link in front of a person and a person seeing a link were near enough the same event that nobody had to separate them.
It cost me a third of a penny to find out I was wrong.
I nearly got away with it.
I knew what the culprit was. Everybody does in 2026 — it is the AI Overview, the answer Google now writes across the top of the page, which reads your work, summarises it, and leaves the reader with no particular reason to visit you. I had four search results saved from a few days earlier. Every single one had an AI Overview sitting above my page. That is four out of four, and I could have written the whole thing up on the spot.
So I bought one more. Not one of the four that agreed with me — the opposite one, a search on the same page that clicks perfectly well, about nine times in a hundred. It cost roughly a third of a penny.
It has an AI Overview too. Same box, same position underneath it, and it converts fine.
Which means the thing I was about to publish was wrong, and four pieces of evidence had agreed with it, and the only reason I know is that I went looking for the one that wouldn’t.
I have been thinking since about how ordinary that is. Four confirmations feel like proof. They are not proof of anything at all if you only ever asked the cases you already believed.
What was really going on is stranger and I still cannot fully name it.
The searches producing nothing are long — six or seven words, written out like sentences, several ending in a question mark. The ones that click are short: three or four words, the way you or I would actually type them. That on its own proves little, so I controlled for position and compared like with like, which is where it got interesting.
Sitting at the same height on the same page, in the same month, the long searches converted about eight times worse than the short ones beside them. One band further down, four thousand impressions produced no clicks whatsoever, while the short searches at those same positions were still clicking away.
And then the number that stopped me. Those long searches ran 338 desktop impressions for every single one on a phone. The rest of my site runs under seven to one.
I do not know what they are. I can tell you what they are not: they are not a few dozen people a day reading a sentence-shaped question into their phone, because the phones are not there. Google’s reporting has no field that tells you whether an impression was ever served to a human being, so anyone who says otherwise with confidence — me included — is telling you something they cannot check.
What I can say is that a number I have reported on for years, and that somebody is probably reporting to you, went up by a factor of twenty-nine in a month while my audience did not change at all.
This matters to you commercially and quite specifically. If you pay somebody for visibility work and the report leads with impressions and average position, that report can improve — honestly, verifiably, nobody lying anywhere — in a month when not one additional human being encountered your business.
It is the same shape as why a website gets visitors but not enquiries: the traffic is real, the failure is quiet, and the figure you are looking at is fine right up until you check the layer underneath it. It is the dashboard is a decoy one level further up — every layer reports on itself, honestly, and the honest report is not an answer to the question you were asking.
The check takes about ten minutes and I have written it out properly, with the numbers, in this week’s piece.
The short version: open Search Console, sort your searches by impressions, and read the top twenty out loud. If they sound like sentences rather than things a person would type, look at the click column beside them. Then add the Devices tab and see whether your desktop-to-mobile split changes shape from one search to the next, because a real audience does not do that.
None of it needs a specialist and none of it costs anything. What it needs is somebody willing to look hard at a month where the headline number went up, which is the single hardest time to persuade anybody that something is worth checking.
There is one more thing in the piece I have not mentioned here. The advertising trade solved exactly this problem in 1914, and the fix was not a better number.
