Half my impressions vanished. My clicks did not move.

Between June 2025 and June 2026, one of the sites I own lost half its impressions in Google. Twenty thousand a month down to ten. The clicks went from 322 to 318. Nothing happened. Zero-click search is real and it is reshaping how people find businesses, but the metric everybody is watching to measure it is the wrong one, and I can show you that with my own data rather than another opinion about it.

The site is openmindsresources.com, the agency I co-founded and still run operations for. I have full Search Console and analytics access, so these are my own numbers, not a case study I read somewhere. Every figure below was pulled on 25 July 2026 and none of it comes from a client account.

What the numbers say

June against June, because that is the only honest comparison. Same month, same seasonality, both complete.

Widen the lens and it holds. Across the five full months from April to August 2025 the site averaged roughly 21,300 impressions and 356 clicks a month. Across April to June 2026 it averaged roughly 11,800 impressions and 324 clicks. Impressions fell by nearly half. Clicks fell by about nine per cent. If I had only been watching the impressions graph I would have concluded the site was dying.

The impressions I lost were never worth anything

You can only see the reason at query level. Our brand contains two very common English words, so for years Google showed us to people searching “open mind” and “openmind” who wanted meditation apps, self-help books, anything except a Malaysian agency. We sat at position 34 for those queries. Almost nobody clicked, which is the correct behaviour from everyone involved.

Read the middle row slowly. We went from 1,042 impressions to 174, a drop of 83 per cent, and it cost us two clicks. Google stopped showing us to people who were never looking for us and started showing us near the top to the ones who were. My dashboard reported that improvement as a collapse.

An impression at position 34 is not an opportunity. It is a row in a report. I have sat in meetings where a client was upset about exactly this shape of graph, and the honest answer was that the traffic they lost had never really been theirs.

AI referrals are measurable now, and mine started in June

Google Analytics has begun reporting a channel called AI Assistant, and on this property it now has data in it. Nothing at all before June 2026, then twenty sessions in June and twenty-one in the first twenty-two days of July. I went looking for something else entirely and found this sitting in the default channel report.

Those are small numbers and I am not going to inflate them. They matter because half this industry is currently insisting that traffic from AI assistants cannot be tracked, which is being used to excuse a great deal of vague strategy. It can be tracked. On a mid-sized property in Kuala Lumpur it went from zero to a measurable trickle inside two months, with no special instrumentation. If you have not looked at your own, look this week, because you are watching a channel arrive in real time and the baseline only exists once.

Search Console will not help you here. It exposes only two search-appearance types on this property, both of them job listing formats, and Google publishes no AI Overview appearance type at all. You are therefore inferring AI presence from where the session lands, not from where Google served the impression. Know that gap exists before you promise a client an AI visibility report, because somebody will eventually ask you how the number was produced.

What should worry you instead

Look again at which queries survived. Almost all of them are our own name and its misspellings. Non-brand discovery on this property is thin, and once the junk impressions stopped there was nowhere left to hide that fact.

When most results get answered on the page, the searches that still send you a human being are mostly the ones where somebody already knew what to type. So the brand does the distributing now, and the keyword work only decides how tidily you get found by people who were already looking. That reordering is uncomfortable if you have spent a decade selling the other way round. I have.

The published research points the same way, even where I cannot verify the detail myself. SparkToro’s 2026 analysis puts it at under a third of Google searches still sending a click. Trackers measuring queries that carry an AI Overview report organic click-through rate down by around 61 per cent, while brands cited inside those answers pick up roughly 35 per cent more organic clicks than brands that are not. Being the source Google quotes is worth more than being the tenth link it offers.

What I changed on my own site because of this

I rebuilt randytoo.com this month with all of the above in mind, so the work is inspectable rather than theoretical.

I consolidated the structured data into one connected graph. The site was emitting a thin Organization block and a separate Person, which is two half-identities pointing at nothing much. Now there is a single canonical Person that acts as publisher, page subject and article author, with every reference repointed to it and no dangling pointers left behind. If a machine is going to attribute an answer to somebody, I want the identity it reads to be unambiguous.

I published an llms.txt file, wrote headings that answer the question they pose rather than tease it, and made sure every claim on the site is one somebody could check.

Then I cleaned up the crawling. Search Console was reporting 85 URLs as not indexed on a site with eleven real pages, which looks alarming until you read the reasons. Eighty-one of the 85 traced to a single pattern: a query string ending in a timestamp, first detected in August 2022, still being pinged by something that predates this entire rebuild. Google was correctly declining to index junk duplicates of my homepage and spending crawl budget doing it. The sitemap now lists eleven URLs and Google reports eleven discovered.

None of that is clever. It is closer to housekeeping. But most of the sites I audit fail on exactly this level, and the owners are shopping for AI visibility tactics instead.

What I report now instead of rank position

Three questions, in this order. Are we the source the answer is built from, and can I show where. Is non-brand demand growing, or are we only being found by people who already know us. Of the clicks we do get, how many turn into a conversation.

Position and impressions still appear in my reports, at the bottom, as diagnostics. They are useful for spotting that something broke. They are close to useless as a measure of whether the work is paying.

How to check whether zero-click search is actually costing you

Half an hour in two tools will tell you, and you do not need to buy anything to do it.

Open Search Console and go to the performance report. Change the date range to compare the same calendar month against the same month last year rather than accepting the rolling default, because the default window will straddle whatever happened to you and blur it. Turn on all four metrics at once. If impressions have fallen much faster than clicks, and your average position has improved at the same time, you are looking at a cleanup rather than a loss.

Then switch to the queries tab and sort by impressions rather than clicks. You are hunting for high-impression, high-position, near-zero-click rows: generic words, competitor names, anything sitting past position 20. Add up their lost impressions. That number is usually most of the fall, and it was never revenue.

In Google Analytics, open the traffic acquisition report and look down the channel list for a row labelled AI Assistant. If it is there, note today’s figure somewhere permanent, because you are early and the baseline is only available once.

Finally, sort your pages by lost clicks rather than lost impressions. That list is the real problem, and on most sites it is short enough to fix.

What I cannot prove

I do not know why the step change happened when it did. Something shifted around October 2025 on this property and my data cannot tell me whether it was a core update, the expansion of AI Overviews, or changes we made ourselves. Anyone who tells you they can separate those three from Search Console alone is guessing with more confidence than I have.

This is also one property, one market, sixteen months. It is not a study. It is what happened to a site I am responsible for, published in full so you can argue with it. I would rather be corrected with better data than be agreed with politely.

If your impressions chart is falling and you cannot tell whether that is a problem or a cleanup, that is a genuinely useful hour of work and it starts with the query-level view, not the summary. Happy to look at it with you: start a conversation. If you want the longer version of how I use AI in this work, I have written about what it actually does in my consultancy.