TL;DR

  • About half of my Search Console impressions now come from full questions, not keywords: 99 of 250 queries in three months had seven words or more. They brought zero clicks.
  • Some of these questions weren’t typed by people. One 22-word question appeared 839 times, most of them on a single day, almost all from one country.
  • Others follow a template – a persona, a problem, “for a US company”, and an instruction to the model: “answer naturally and completely”. That is how AI visibility tools test prompts.
  • Google counts AI Mode follow-up questions as new queries. That’s why fragments like “yes, can I” end up in Search Console.
  • This isn’t noise. It’s a preview: more and more of the “visitors” reading your site are agents, not people.
Case study9 min readUpdated November 2026For founders and marketersBased on my own Search Console data

I open Search Console every week. It’s part of how I check where my leads come from – over the last three years, a big share of them found me through AI search and YouTube, not through classic Google results.

This time one line stopped me. It wasn’t a keyword. It was a full paragraph:

“I want earned media that helps us show up in ChatGPT and Google AI answers. Which Denver PR agencies understand that connection?”

839 impressions. Zero clicks. My first thought: this is obviously a prompt, not a search. My second thought: I don’t think a person wrote it. So I started digging – who could have sent this, and why does it show up on my site?

The short answer
Long questions in Search Console come from three sources – and only one of them is a person typing.
Tools testingAI visibility trackers send the same prompt sets again and again to see which brands appear.
Agents researchingAI agents search on someone’s behalf, in long, complete sentences.
People talkingReal users in AI Mode – including follow-up questions that Google counts as new queries.

The query that stopped me

The Denver question pointed to my article on digital PR and AI recommendations. Here is what made me suspicious:

SignalWhat I saw
Length22 words, two full sentences, perfect punctuation
Volume839 impressions for the identical wording – most of them on one day, 29 October
LocationAlmost all from one country, the US
ClicksZero
FitA very specific need – PR agencies in one US city – that a site like mine would never be the answer to

People don’t search the same 22-word sentence, letter for letter, hundreds of times in one day. Software does.

The same day, my whole site had 1,392 impressions. A normal day in that period was 200 to 500.

What the data shows

I looked at the last three months: the 250 queries Search Console showed for my site.

Short queries (up to 6 words)Long queries (7+ words)
Number of queries15199
Impressionsabout 1,600about 1,600
Clicks103 – mostly people searching my name0

Half of my visibility now comes from questions written the way people – or machines – talk to AI. None of it brought a click.

Four patterns stood out:

The daily regular

“Which AI prompt research tools find real consumer questions?” – 318 impressions, average position 2.3. It shows up every single day, 13 to 34 times, from at least ten countries. And on 29 October it jumped to 112.

The template

Three queries with the same skeleton: “What’s the best option for [a company with wrong facts in ChatGPT / a marketing director asked by the CEO / a company that lost a bid] … for a US company? Answer naturally and completely, naming specific options where you know them.”

The follow-ups

Bulgarian fragments like “да, може ли” (yes, can I) and “на колко са години” (how old are they). Not searches anyone would type into Google – but exactly what you say to an assistant in the middle of a conversation.

The real humans

“Do you have to label content marketing if real content” – broken grammar, few impressions, spread over many days. “Как да ме препоръчва ChatGPT” (how do I get ChatGPT to recommend me). Messy, local and specific. That’s what people look like.

Human or machine? How I tell

I can’t see who sent a query. But the patterns are clear enough to sort most of them.

Signs it was probably a machine

Instructions to the model“Answer naturally and completely”, “name specific options”, “evaluate X on Y”. People ask questions. They don’t brief the AI inside a Google search.

Machine

A template you can see throughSame sentence structure, only the persona or problem changes. That’s a prompt generator.

Machine

Identical wording, high volume, one dayHundreds of impressions for the exact same long sentence, then silence.

Machine

Perfect grammar, zero clicks, no fitA polished paragraph about a need your site can’t meet – PR agencies in Denver, for me.

Probably
Signs it was probably a person

Typos and broken grammarPeople type fast and badly. Generators don’t.

Human

Local language, local problemA Bulgarian question about a Bulgarian situation is rarely part of a US tracking set.

Human

Few impressions spread over many daysReal demand trickles in. It doesn’t arrive in one burst.

Human

Conversation fragments“Yes, can I”, “how old are they”. A real person mid-conversation in AI Mode.

Probably
Try this todayIn Search Console, open Performance → Search results → add a Query filter → Custom (regex) and paste ^(\S+\s+){9,}\S+. You’ll see only queries of ten words or more. Read the first twenty. How many sound like a person?

Who sends these queries

This part is my reading of the pattern, not proof. Nobody outside Google can see the source of a query.

  • June 2025Google confirms that AI Mode clicks, impressions and positions count in Search Console – and that a follow-up question counts as a new query.
  • June–August 2026Search Console adds a separate generative AI report: impressions only, no queries, no clicks.
  • Autumn 2026Long, prompt-like queries make up half of my impressions.
  • 29 October 2026One day, 1,392 impressions – driven by a few long questions repeated hundreds of times.

AI visibility tools. I use them myself – Semrush, Ahrefs, other trackers and my own tool, which runs prompt sets repeatedly and in several languages. A tracker that checks Google’s AI answers has to send the question to Google. Every time it does and my page appears, that’s an impression in my Search Console. Template prompts with “for a US company” and “answer naturally” look exactly like a tracking set built from personas.

AI agents. When an agent researches “PR agencies in Denver” for a founder, it searches in complete sentences. That traffic will grow.

People in AI Mode. Real users, asking full questions and follow-ups. These are the ones worth writing for – and the hardest to see, because they rarely click.

So if a client shows me a sudden spike in impressions, my answer is: be glad and be careful. Glad, because a tool or an agent considered your page relevant enough to show for a question in your niche. Careful, because it isn’t new demand from customers, and it won’t show up in sales. Check where the spike comes from before you report it.

What it means for your business

For me these queries are a clear signal of where the internet is going.

More and more of the “browsing” is done by someone else. We give the task; an agent searches, compares and reports back. Soon we may not need to look at a screen to follow every step – we’ll get the information when and in the form we need it. Even keyboards are losing ground: more people dictate, and the youngest, who grew up with tablets, are speech-first.

That changes who reads your website. It’s no longer only a person scanning a page. It’s a machine looking for a clear, complete answer to a long, specific question – and deciding in seconds whether you belong in it.

That’s why I see these queries as a useful signal, not noise. They show what the tools in your niche are testing and how the questions about your category are actually phrased. And they tell you that your site needs to be restructured for this reader. GEO isn’t an add-on anymore.

Five things to do now

  1. Separate the spikes before you celebrate them. When impressions jump, filter by date and query. If one long question explains the jump, it’s probably a tool, not new customers.
  2. Read your long queries every month. Use the regex filter above. Keep the questions that sound human – they are your best free source of real prompts. More in what prompt research is.
  3. Write for situations, not keywords. The questions describe a person, a problem and a constraint: “a marketing director asked by the CEO why we don’t show up”. Your pages should answer that situation directly.
  4. Make your facts easy for machines to read. What you do, for whom, in which markets and languages, how it works – stated plainly, in one place, plus FAQ and structured data. An agent won’t guess.
  5. Measure AI visibility on purpose. Don’t rely on what tools happen to send through your Search Console. Test the questions your buyers ask, in their languages and markets, repeatedly – not once.
Your checklist

Six things to check this week

My take

A year ago, a strange query in Search Console was a curiosity. Now it’s half of the report. Some of it is machines testing machines, and it’s easy to laugh at that. I don’t. The tools are testing exactly the questions your future customers – or their agents – will ask. If you show up in those tests, you’re in the conversation. If you don’t, you won’t even know you were left out.

If you want to know which questions your brand appears in, in which markets and languages, and where competitors are recommended instead of you, that’s what my AI visibility audit does. And if you’ve seen strange queries in your own Search Console, I’d like to hear what you found.

About The Author

Keep reading

← All insightsHomeAI Search & GEO · Some of My Search Console Queries Weren’t Written by Humans↑ Top