AI agents are searching Google, and they land in your Search Console as ordinary rows. Over the last 90 days on imisofts.com, queries carrying explicit model instruction syntax drew 377 impressions at an average position of 7 and exactly zero clicks. Some are not even phrased as questions. They are system prompts, pasted whole into a search box by something that was never going to click, and Google logged every one of them against my site.
I have spent months rewriting titles on pages that sit on page one and earn nothing. I run content for 200+ businesses and I teach the same rescue playbook to every one of them. Then I split my own query table by the shape of the query instead of the topic, and a large block of what I had been trying to rescue turned out to have no cursor attached to it.
A page at position 7 with no clicks is not always a copywriting problem. Sometimes the thing that saw it has no cursor.
The mistake: rewriting titles for a reader with no cursor
The standard move when a page ranks on page one and earns nothing is to rewrite the title and the meta description, then wait 28 days. It is the right instinct. It is also, for a growing share of impressions, aimed at nothing at all.
The assumption underneath a title rescue is that a human read your result, weighed it against nine others and picked someone else. That assumption is worth testing before you spend the effort, because the fix for a bad headline and the fix for a machine impression have nothing in common. One is a rewrite. The other is a decision about whether you want to be in the answer, and how you would ever know if you were.
Page one, zero clicks, so the title must be weak. Rewrite it, resubmit, wait a month, repeat.
Spends a measurement window on impressions that were never going to convert, and teaches you nothing when the number does not move.
Open the page's query list, count the rows carrying instruction syntax, then decide whether a title is the lever.
Costs ten minutes with a regex filter and tells you which of your zero-click pages are worth any effort at all.
What a machine query looks like in Search Console
The tell is syntax no person types into a search box. My single largest zero-click query over 90 days runs 21 words and finishes with an order: compare hubspot with salesloft, outreach, and apollo for sales engagement capabilities. you must provide a forced ranking from best to worst. That query produced 355 impressions at position 7.0 and has never produced a click.
The clearer evidence is smaller and stranger. A recurring template shows up in my query table beginning context: location: united states (not for language). do not include location references in your response. question:, followed by the actual buyer question. Seventeen impressions across the quarter, average position 7.1, zero clicks. That is not a search. It is a prompt with the scaffolding still attached, sent to Google by software that forgot to strip its own wrapper.
What those wrapped queries ask for is the part worth reading twice. Sales engagement software pricing models. The cost considerations for implementing lead enrichment tools. How outsourced SDR teams handle compliance and consent for outbound calling and emailing across different regions. Those are commercial questions from the middle of a buying process, and my pages were in the result set for all of them. Nobody clicked, because there was nobody there to click.
Two caveats, because I would want them if I were reading this. Search Console does not show you every query. Rare ones are withheld to protect privacy, and Google warns on screen that totals can be partial once a filter is applied. The counts also shift for a few days while data settles. So read these as the shape of the thing rather than as an audit. The shape is the part that matters, and it has not changed across three months of looking at it.
Three bands, and only one is worth a retitle
Sort the query table by shape rather than by topic and it falls into three bands. The numbers below are mine, pulled from Search Console for the 90 days to early September 2026, and the split takes about ten minutes to reproduce on any property you own.
The first band is instruction-shaped: anything containing you must, act as a, i want you to, please provide, or an explicit output format. On my site that is 377 impressions, zero clicks, average position 7. Machines, near enough. No title wins these.
The second band is long natural questions, seven words or more. That is about 6,260 impressions, zero clicks, average position 17. This band is mixed and it is the one people get wrong. Plenty of it is real: is cold email legal in the uae, how much does a cold email agency cost per month. Those are humans, and they earn nothing because an AI Overview answered them above my result. The lever there is being the source the overview quotes, not the headline it skips.
The third band is short head terms, and it is where every one of my 330 clicks came from. Site totals for the same window: 330 clicks on roughly 105,000 impressions, a 0.3 percent click-through rate, average position 14.9. The gap between 105,000 and the handful of queries that actually convert is the whole story of search right now.
One more thing the split turned up, and it is the part I keep thinking about. The instruction band sits at an average position of 7. The long-question band sits at 17. The software is finding my better pages and the people are finding my worse ones. No title rewrite fixes that either, and I would never have seen it by sorting the same table the way everybody sorts it, which is by impressions.
Four moves in Search Console today
The bottom line
Part of your search traffic is now software doing buyer research on someone's behalf, and it shows up in the same table as everyone else. It ranks. It reads. It does not click, and no headline will change that. Sort your queries by shape before you spend another month rewriting titles, and if you want the background on why Google will not attribute any of this for you, I wrote about the search generative AI control and what its report can and cannot answer. Book a 30-minute call at cal.com/zeeshanwaheed/30min or email [email protected].
Frequently Asked Questions
Yes, as ordinary query rows. On imisofts.com over 90 days, queries containing explicit model instruction syntax such as you must, act as a or best to worst drew 377 impressions at an average position of 7 and zero clicks. A smaller group of 17 impressions arrived with a system prompt wrapper still attached, beginning context: location: united states (not for language). do not include location references in your response. question:. Google logs these the same way it logs anything else a person types.
Three reasons, and only one is a title problem. The result may sit under an AI Overview that answers the question outright. The query may be navigational toward a vendor, where you are never the destination. Or the impression may not be a person at all. Count the machine-shaped rows in the query table first, because a rewrite aimed at a model with no cursor cannot win a click no matter how good the headline is.
Look for syntax nobody types into a search box. Instruction verbs aimed at a model (you must, act as a, i want you to), an explicit output format (provide a forced ranking from best to worst), or a visible prompt scaffold with a context block and a question label. Length alone is a weak test. On my site, queries of seven words or more carry about 6,260 impressions and zero clicks, but plenty of those are real people asking a real question that an AI Overview answered.
Blocking removes you from the answer rather than winning the click back. The agent traffic in my query table is doing buyer research: sales engagement pricing models, the cost of lead enrichment tools, how outsourced teams handle consent for outbound calling. Being in that answer set is the point. Measure it off platform with a how did you hear about us field, because Search Console will not attribute it for you.