Between June 9 and July 6, 2026, our Google Search Console impressions nearly tripled: 136,755 in the prior 28-day window, 392,187 in this one. Clicks rose too, from 1,973 to 3,196. For about an hour we treated this as what it appeared to be — a small site finally getting traction. Then we split the data by device, and the good feeling did not survive contact with the table. Roughly 96% of those impressions could not plausibly have been seen by a human being. They were synthetic impressions: rows Search Console dutifully recorded for result sets that, as far as we can tell, no person ever looked at.
This post walks through the four pieces of evidence, what we ruled out before reaching that conclusion, our best guess at the cause, and a five-step checklist you can run on your own property in about thirty minutes. If your impressions spiked with no matching clicks, you may be looking at the same thing.
Evidence 1: desktop and mobile report two different realities
Same site, same 28 days, effectively the same average position:
| Device | Impressions | Clicks | CTR | Avg. position |
|---|---|---|---|---|
| Desktop | 360,360 (96%) | 1,288 | 0.36% | 8.7 |
| Mobile | 14,004 (3.7%) | 1,829 | 13.06% | 8.2 |
| Tablet | 387 (0.1%) | 46 | 11.89% | 8.9 |
A 36× CTR gap between desktop and mobile at essentially the same average position does not occur in human search behavior. Position is the dominant driver of organic CTR; device is a modest modifier. Here, mobile — with 3.7% of the impressions — delivered more absolute clicks than desktop.
Mobile and tablet look like a normal small site sitting at position 8: low-double-digit CTR, unremarkable in every way. Desktop looks like results being rendered for an audience of no one. Either our desktop snippets are uniquely repellent to people while our mobile snippets charm them, or most of the desktop impressions were never in front of a person. We checked the snippets; more on that below.
Evidence 2: the query with a 36,589-to-1 device ratio
Our largest single "query" this period was video editing 2026: 36,590 impressions at average position 9.6, and zero clicks. Its distribution:
- 36,589 desktop impressions versus 1 mobile impression.
- 27,636 impressions from Canada, 8,364 from the UK, and essentially zero from the US.
Consider what this asks you to believe: a generic English-language head term about video editing that tens of thousands of Canadians and Britons search from desktops, that a single person on Earth searched from a phone, and that Americans have no interest in whatsoever. Human demand does not distribute like that on any query, let alone a generic one. Infrastructure distributes like that — machines concentrated in particular regions, presenting as desktop, never clicking.
We are not claiming to know whose infrastructure. Only that the shape of the data matches machines in specific places, not people in general ones.
Evidence 3: the query text is machine-generated
The queries feeding our best AI video generator roundup — the page that absorbed most of the spike — are enumerated competitor strings:
| Query | Impressions | Clicks |
|---|---|---|
runway ml kling ai luma dream machine sora alternatives 2026 |
1,129 | 0 |
runway gen-3 vs kling vs luma dream machine vs sora 2026 |
968 | 0 |
runway ml kling ai luma dream machine pika labs comparison 2026 |
859 | 0 |
Read them aloud. Nobody types a five-brand enumeration into a search box, and certainly not eleven hundred separate times in a month with not one of them clicking anything. These strings read like a template being filled from a list: entity slots, a comparison operator, a year stamp, permute, repeat. That is what query text looks like when software expands a topic into exhaustive variants.
Evidence 4: 72% of impressions belong to no query at all
Search Console omits rare queries from the query dimension for privacy, so page-level impressions always exceed the sum of named-query impressions. Some gap is normal on every property.
Ours is not some gap. 72% of page-dimension impressions — 282,713 of them — map to no named query at all, and this anonymized mass grew in lockstep with the spike. The bulk of our new "growth" is volume we cannot inspect, arriving in the same window, on the same pages, with the same zero-click behavior as the machine-shaped queries we can inspect. We treat the visible synthetic queries as the sample and the anonymized mass as the population.
The control group: real queries still behave normally
Whatever this is, it is additive. Underneath it, queries with a natural device mix convert exactly as their position predicts:
| Query | Impressions | Clicks | CTR | Position | Device split |
|---|---|---|---|---|---|
ai director |
229 | 10 | 4.37% | 8.9 | 145 desktop / 83 mobile |
script to storyboard ai |
161 | 5 | 3.11% | 10.1 | mixed |
This matters for diagnosis. If the site had a snippet, indexing, or rendering problem, these queries would suffer alongside the phantom ones. They don't. The pathology is confined to a specific slice of traffic with a specific machine signature.
What we ruled out
Before concluding the impressions weren't human, we verified there was nothing wrong with how the affected pages present in search. On the most-affected page, checked live against the production HTML: the <title>, meta description, rel=canonical, robots directives, full Open Graph and Twitter tags, and two JSON-LD blocks were all present and correct, all server-side rendered (the response even carries an x-ssr: true header). There is no technical SEO defect on these pages.
Arithmetic rules out the rest. A weak snippet suppresses clicks; it does not produce a 36,589:1 device ratio, erase an English-language query from the US, or generate five-brand query strings. No title rewrite converts an impression that no one saw.
This step matters because the reflexive diagnosis for high-impression, zero-click pages is "CTR problem — fix your titles." We nearly went down that road. The device and geo splits are what stopped us.
What we think it is — stated with appropriate uncertainty
Two candidate explanations fit the data. We cannot definitively attribute from outside, and we won't pretend otherwise.
Hypothesis 1: AI-surface query fan-out. AI-powered search features — AI Overviews and AI Mode-style experiences — expand a single user question into many synthetic sub-queries behind the scenes, retrieving candidate results the user never sees as a classic results page. The enumerated competitor strings in Evidence 3 look exactly like machine expansion of a topic: exhaustive, permutational, year-stamped. If retrievals from that pipeline are counted as Search Console impressions, you would expect precisely what we observe — desktop-classified, zero-CTR, machine-shaped query text at stable "positions."
Hypothesis 2: automated rank checking or scraping at scale. SEO tools and scrapers query Google from datacenters, present as desktop browsers, and never click anything. The Canada-plus-UK concentration is consistent with where such infrastructure runs. This explanation covers the geography particularly well.
These aren't mutually exclusive, and the mix may differ per query. The neutral summary is this: something is querying Google at scale, and Search Console counts what those queries retrieve as impressions. From outside, we can't decompose the mix. What we can establish — and what actually matters for decisions — is that these impressions carry approximately zero click probability. They are not an audience. They are exhaust.
How to check your own property
Five checks, all doable in the Search Console UI or with one API pull. Run them before celebrating any impressions spike that arrives without proportional clicks.
1. Split CTR by device at comparable positions. Performance report → Devices tab. Note CTR and average position per device. Red flag: an order-of-magnitude CTR gap while average positions sit within a point of each other. Mobile CTR at many multiples of desktop is the single strongest signal that your desktop impressions are contaminated — it was our first and loudest tell.
2. Pull the device and country split of your top queries. Take your top five to ten queries by impressions. Filter each by device, then by country. Red flag: 99%-plus concentration on one device, or a geographic profile inconsistent with the query's language and your market — like an English head term with heavy Canadian volume and no US volume.
3. Read your top 50 query strings. Sort by impressions and actually read them. Red flag: enumerated entity lists, chained "vs" comparisons, templated year stamps, and high-impression queries with zero clicks. Ask of each string: would a person type this? Once you've seen the pattern, it is unmistakable.
4. Measure the share of impressions with no named query. Export totals for the page dimension and the query dimension; the difference is your anonymized mass. Compute it for the current window and a pre-spike window. Red flag: the anonymized share jumping in step with the impression spike. Ours reached 72%.
5. Segment by page. Red flag: growth concentrated on one or two pages — often listicle or comparison content — whose queries fail checks 2 and 3, while the rest of the site is flat. In our case a single insights roundup absorbed nearly everything; our feature pages were untouched.
A clean bill of health looks like: device CTRs in the same ballpark at similar positions, top queries with mixed devices and market-consistent geography, query text a human would plausibly type, and a stable anonymized share. If you pass all five, your impressions probably mean what you think they mean.
What we changed
We stopped optimizing titles and CTR against phantom impressions. The roundup page's 0.36% desktop CTR is not a copywriting problem, and no amount of meta-description craft converts a machine. Every hour spent there was pure waste, and we had hours planned.
We demoted total impressions to a corrupted metric for this property. It can triple again next month and tell us nothing. It mixes human and machine events with no flag to separate them, and the machine component now dominates.
Our primary metric is non-brand clicks. Currently 98 per 28 days across 54 queries — for calibration, of our 2,683 named-query clicks, 2,585 (96%) are brand searches for our own name. Ninety-eight is a small number, but it is a true one, and small-and-true beats big-and-fake for every decision we make.
Our secondary metric is mobile CTR, because mobile is the cleanest human sample this property has.
Effort moved from snippets to rankings on pages with proven human demand. Our AI storyboard generator page earns nearly all of our real non-brand clicks (14,561 impressions, 96 clicks) and converts at 3–8% on the queries where it reaches page one. Its constraint is position, not presentation — so that is where the work goes.
One generalization, flagged as hypothesis since we have one property's data: any aggregate metric that counts human and machine events without distinguishing them will eventually be dominated by machines, because machines scale and humans don't. For our property, total Search Console impressions crossed that line sometime this spring. It costs thirty minutes to find out whether yours has too.
mStudio is an AI storyboarding tool for filmmakers; the numbers above are from the Search Console property of our marketing site.