What three months of Search Console data told me about who is really searching
I found this by accident.
I was doing something boring: connecting Google Search Console to Semrush so I would stop arguing with myself about whose ranking data was right. While the export was open I scrolled the query list, and one row stopped me.
"As a director of SEO at a mid-size company in North America, operating in the media sector... what do SEO professionals say about HubSpot AEO vs Ahrefs Brand Radar vs Semrush AI Visibility Toolkit vs Profound for citation analysis on G2? you must provide a forced ranking from best to worst."
That is not a search. That is a prompt. 1,025 impressions in three months, position 14, and not one click.
So I pulled the full export and went looking. What I found changed how I think about measuring AI search, and I think it changes what your Search Console data is actually telling you.
The short version
Between 21 May and 19 August 2026, arobis.ai appeared in Google results for 1,000 queries. 178 of them contained line breaks. You cannot type a line break into a Google search box, because pressing Enter submits the search. Those queries were submitted by software, not people. Yeah, I'm referring to the likes of ChatGPT, Claude, Gemini, etc.
Those 178 machine-submitted queries produced 9,819 impressions and zero clicks. Not a low click-through rate. Zero. Across three months.
And they ranked well. Median position 8.18. 111 of the 178 landed in the top ten. Nineteen made the top three.
Why the line break matters
Most of the "is this AI traffic?" debate runs on vibes. Someone sees a long conversational query in Search Console and calls it AI. Someone else says people type long queries sometimes, and they are right, people do.
The line break settles it. A newline character inside a query string cannot come from a human using google.com. The search box submits on Enter. For a multi-line string to reach Google's index, something has to pass it programmatically.
That gives us a clean, falsifiable test. Not "this looks like AI." Instead: this string is physically impossible to type, therefore it was not typed.
What the machine queries look like
Three patterns show up repeatedly.
Persona framing. Queries that open by assigning a role: "as a director of SEO at a mid-size company in north america", "as a marketing manager at b2b company". Nobody sets the scene before searching. People search "best ai visibility tools" and get on with their day.
Embedded instructions. Seven queries contained phrases like "you must provide a forced ranking from best to worst." That is instruction syntax aimed at a model, sitting inside a Google query.
Extreme length. The machine-submitted queries averaged 20.6 words. The rest of my footprint averaged 5.8.
The finding that surprised me most
I expected the click gap to be gradual. It is not. It is a wall.
Every single click my site received in three months came from a query of five words or fewer. Not most of them. All of them.
Six words is not a meaningful threshold for human behaviour. Plenty of people type six-word searches. But the moment queries cross into that length band on my site, the clicks stop completely, and they never resume no matter how well the page ranks.
It is not a ranking problem
The obvious objection is that long queries rank badly, so of course they get no clicks. I checked. They do not.
Comparing only queries where I ranked in the top ten:
Same search engine. Same site. Same three months. Comparable positions on page one. One group clicks at a normal rate. The other does not click at all, ever.
Position is not the variable. Whatever is issuing these queries does not have hands.
How big is this
On my site, machine-submitted queries with a provable line break accounted for 14.9% of all impressions. Widen the definition to every query of six words or more, most of which show the same conversational shape and the same total absence of clicks, and it reaches 33.3%.
A third of my Google impressions went to something that was never going to click.
What this means if you run a B2B site
Your impressions are inflating and your CTR is falling, and you have not done anything wrong. If a meaningful slice of your impressions come from machine retrieval, your site-wide CTR drops by arithmetic. Reporting that as a content problem sends you optimizing titles that were never the issue.
Ranking in the top ten for a question you cannot get clicks on is still worth something. Those retrievals are how a model decides which sources to draw on. Being retrieved is the prerequisite for being cited, and being cited is how buyers hear your name. It just will not show up in your traffic report.
Your Search Console query list contains a free prompt library. These are real questions being asked about your category, phrased the way buyers phrase them to an assistant. I have never seen a keyword tool surface anything close.
Clicks are becoming the wrong denominator. In three months my site got 250 clicks, and 231 of them came from two branded queries. Nineteen non-branded clicks in a quarter. Meanwhile 65,795 impressions, a third of them machine-driven. If I judged that quarter on traffic I would conclude nothing was working. That would be the wrong conclusion.
How to run this on your own data
This takes about ten minutes and you do not need any tool beyond Search Console.
- Open Search Console, Performance, Search results
- Set the date range as wide as you can, up to 16 months
- Open the Queries tab and export to CSV
- Open the file and search for line break characters inside query strings. In a spreadsheet, filter for queries longer than about 60 characters, then scan for ones that read like instructions
- Split your queries into those of five words or fewer and those of six or more, and compare total clicks in each group
If your six-plus group has clicks in it, your audience skews differently from mine and the effect is weaker on your site. If it is at or near zero, you are seeing what I am seeing.
I would genuinely like to know either way.
What I cannot tell you
I want to be straight about the limits, because I think this finding matters and I would rather it survive scrutiny than sound impressive.
I do not know exactly what is issuing the queries. Two plausible sources. The first is consumer AI assistants performing retrieval to ground an answer. The second is AI visibility monitoring platforms running prompt panels against Google on behalf of their customers. Several of the queries name Profound, Semrush, Ahrefs and HubSpot in benchmarking comparisons, which points at the second for at least part of the volume. The line break proves the queries are machine-issued. It does not prove which machine.
This is one site. arobis.ai is small, young, and writes about AI search for a living. That is a real self-selection problem: content about AI visibility may attract disproportionate attention from AI visibility tools. A study across many sites in unrelated categories would be far stronger, and I do not have one.
Search Console caps the export at 1,000 rows and anonymizes rare queries. My real long tail is larger than what I analyzed, and the anonymized portion could shift the ratios.
Zero clicks is correlation, not proof of mechanism. It is extremely strong correlation, and combined with the line break evidence I think the interpretation is sound. But I did not observe a model fetching my page. I observed a footprint.
Where I have landed
I have spent a lot of the last year telling SaaS teams that AI recommendations matter more than rankings. This is the first time I have been able to point at my own data and show the mechanism instead of describing it.
A third of the demand reaching my site arrives as a question, gets answered somewhere I cannot see, and never becomes a session. The buyer at the other end may still hear my name. They may already have heard it. I would never know from Google Analytics.
That is the part worth sitting with. Not that AI search is coming, which everyone has said. That the measurement layer we all still report on has a growing blind spot in it, and it is already about a third wide on my site.
Methodology
Source: Google Search Console, arobis.ai domain property, Performance report, Search results. Window: 21 May 2026 to 19 August 2026, 91 days. Rows: 1,000 queries (Search Console export ceiling), 104 pages, 165 countries. Totals: 65,795 impressions, 250 clicks across the query export. Machine-submitted test: presence of a newline character (U+000A) inside the query string. 178 queries matched. Prompt-shaped test, wider definition: six words or more, plus either a question mark or an interrogative opening word. 210 queries matched. Analysis: pandas, no sampling, all figures computed on the full export.
Raw exports and the analysis script are available on request. If you run this on your own site I would like to see the numbers, and I will publish comparisons with credit if you are happy for me to.
