Strange queries in Search Console? An AI wrote them
One line in our Search Console was an AI agent's own instruction, sent to Google word for word. It was not the only one: 35 of the 148 queries Google shows us were written by something other than a person. Here is how to spot yours.
What we found. Over 28 days, 35 of the 148 visible queries on cittago.com carry the fingerprints of a machine: search operators, half-sentences from a chat, and in one case a language-model instruction that leaked into Google's search box intact.
Why it is not for everyone. You need Search Console and a few hundred impressions a month. Below that, Google shows you too few queries for any pattern to surface. This is a reading exercise, not a tool purchase.
Why a small business should care. These are the narrowest questions that exist, so they are the ones you can win fastest. They put us nearly twice as high as the queries people type themselves.
- The full text of a robot's instruction that reached Google, reproduced exactly.
- Four tells that separate machine queries from human ones, with our own examples.
- What Google states in writing, and the number it has never published.
- The one conclusion almost everybody draws from this data that the data does not support.
The query that was really a robot's instruction
Start with the line that made us write this. It appeared in our Search Console as a query, and we ranked tenth for it. Reproduced in full, nothing removed:
search the web for the following query and provide a concise, factual answer. list all brands, products, and services mentioned. include source urls. be thorough but avoid filler text. query: top databricks companies in india
Read it again. It is not a search. It is an instruction written for a language model: search the web, be concise and factual, list every brand, include source URLs, be thorough but skip the filler. Only at the end, after “query:”, comes what the human actually wanted — Databricks companies in India.
Somebody built an agent that searches the web. The agent was supposed to take the part after “query:” and send that to Google. It sent everything.
If a client showed us this line in their account, we would expect them to ask whether they had been hacked. They had not. It is a programming slip at the other end, inside a tool somebody else built, and we only see it because we happened to land in the top ten for it.
We are not claiming it was ChatGPT. There is no way to know. What we know is that the text was written for a language model rather than for Google, and it arrived at Google anyway.
What Google says it is doing
No speculation needed here. Google published the mechanism on 20 May 2025, announcing AI Mode:
“AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf”
The Keyword, Google's blog, 20 May 2025
Definition: a fan-out query is a search the user never wrote — the AI assistant composes it on their behalf to gather the material for its answer. One question goes in; one or more searches come out.
The developer documentation says the same about AI Overviews and adds the part that matters to a site owner: sites appearing in AI features are included in overall Search Console traffic and reported in the Performance report, under the “Web” search type. Not separately.
So the circle closes. Google confirms its systems compose searches. Google confirms the results land in your ordinary report. Google never tells you which is which — which is why the sorting is done by eye, on the text.
Four tells
We read all 148 queries in the account. Four patterns repeated.
Whole sentences, first person
The hardest to spot, because it looks almost human. Almost.
what service enables a/b testing of new urls to see if chatgpt will cite them for a given query?
Nine impressions in 28 days — the most visible machine query in the whole account. Complete sentence, question mark in place, “a/b testing” spelled properly. People type into Google like a telegram: “ab test url chatgpt citation”. Full sentences show up when somebody talked to an assistant and the assistant passed the question along without shortening it.
One more, two impressions, position 65. It is the longest line in our account and it is effectively a paragraph:
my brand keeps getting mentioned in chatgpt responses but i have no idea which websites or sources are actually getting cited when that happens — is there any geo tool in india that can track exactly which urls or domains are being used as citations when ai recommends my brand?
Em dash included. Someone in India has exactly the problem this article is about, told a chat window the whole story, and the chat window forwarded the story to Google.
Conversation fragments
The opposite failure: far too short, and meaningless on their own. “anche in italia?” at position 1. “is it available in the eu?” at 7. “waar wel?” at 10 — Dutch, and it means “where then?”.
Nobody opens Google to type “where then?”. These are replies. They only make sense as the second line of an exchange whose first line happened somewhere else.
Specifications rather than questions
google ads official performance max lead generation best practices 2026
Seven words welded together, with “official” and “best practices” and the year bolted on the end. Same shape: “meta ads ecommerce creative best practices 2026 official”. A person writes “performance max best practices” and hits enter. This is a list of conditions.
Search operators
Six queries in our account carry the same tail, word for word:
"cloudflare" -site:reddit.com -site:twitter.com -site:x.com -site:wykop.pl -site:tripadvisor.com -site:youtube.com -site:yelp.com -site:booking.com -site:facebook.com -site:instagram.com -site:tiktok.com
The same eleven exclusions, six times, with one word swapped at the front: cloudflare, hosting, paradise, conductor. Nobody shopping for hosting types eleven exclusions, and nobody ever types -site:wykop.pl, which is a Polish forum.
You are not looking for a report. You are looking for a register. A machine writes too completely or too technically; a person writes too briefly. Ten minutes of reading the list and you see it unaided.
The numbers on our own account
This is the part that surprised us.
We rank twice as high on questions nobody wrote
All 148 queries Google shows for cittago.com over 28 days, split by the four tells. Press the second button to see where the operator queries find us.
Same 28 days, same site. On the queries people typed we sit at 48.5. On the ones carrying a machine signature, 25.7 — nearly twice as high.
cittago.com · Search Console API, classic performance report, 6 July – 2 August 2026
| Group | Queries | Impressions | Avg. position | In the top 10 |
|---|---|---|---|---|
| Written by people | 113 | 507 | 48.5 | 21 (19%) |
| Machine signature | 35 | 63 | 25.7 | 18 (51%) |
| All visible | 148 | 570 | 43.1 | 39 (26%) |
The line we kept staring at: 18 of the 35 machine queries find us inside the top ten. More than half. For the human ones it is 21 out of 113 — fewer than one in five.
The explanation is not flattering to us, which is exactly why it belongs here. We are not better at machine questions. On a twelve-word query with a year and a set of exclusions, five pages in the world are competing. On “marketing agency”, a few hundred thousand are. You win easily where nobody has bothered to write.
Nobody knows how many searches an AI actually runs
The obvious follow-up question, and the honest answer is that there is no published number.
Google writes “a multitude of queries” and gives no figure at all. The industry filled the gap by itself: search for “query fan-out” in English and the first page repeats “8–12 parallel sub-queries”. We went looking for the end of that chain. The most-cited source is a Surfer SEO study from December 2025 across 173,902 URLs. A well-known SEO platform simply writes “dozens”, citing nothing.
We also checked a third figure that circulates widely — “95% of fan-out phrases have zero search volume”. It does not appear in the article it is attributed to. We are not using it.
Meanwhile somebody did measure. On 29 July 2026, Suganthan Mohanadasan published an experiment in Search Engine Journal: he intercepted Perplexity's raw data stream rather than reading its answers. Eight captures, seven query types plus Deep Research, across two builds. For six of the seven, Perplexity ran a single search with the query near-verbatim. Only local queries escalated, to four searches across two rounds.
“8–12” and “1” can both be true, because they describe different systems. The practical conclusion is not a number — it is that there is no single number, and anyone selling you one without naming the system is selling an assumption.
What this data cannot tell you
The section where we ask ourselves the hard questions, because otherwise the article promises more than it delivers.
| What you see | What you cannot conclude |
|---|---|
| A query with a machine signature | That an AI assistant sent it. An SEO monitoring tool or a scraper produces the same shape. Google does not say. |
| An impression on that query | That you were cited or recommended in the answer. An impression only means you were in the list the system read. |
| A good position | That you have traffic. For us, those 35 queries produced 63 impressions and negligible clicks. |
| 148 queries in the table | That this is all of them. Of 1,986 impressions, only 570 carry a visible query. |
That last row is the rule everyone forgets. Google's own Search Console documentation states: “Some queries are omitted from the report to protect user privacy. These are called anonymized queries. They're included in chart totals”. The chart shows you the total; the table shows you a slice. Here the slice is 29%, and every percentage in this article is calculated on it.
One more limit, for honesty: 63 impressions across 35 queries is an observation on one account, not a study. We went looking for a published study analysing real Search Console queries attributed to AI assistants — rather than simulated ones — and found none. If you run a site with serious traffic and see something different, we would genuinely like to know.
What we changed
Not “what we recommend”. What we actually changed once we started reading this list.
| Step | Before | Now |
|---|---|---|
| Reading the list | Sort by impressions, look at the top 20. | Read every row and mark the four tells. |
| Long queries | Ignored — “nobody searches for that”. | First source of topics. That is where you win in days, not months. |
| Article format | A well-written essay. | Question, literal answer in the first sentence, table, FAQ. A system can lift a block; it cannot lift an essay. |
| Monthly reporting | Impressions, clicks, average position. | The same, plus how many machine queries appeared and where they placed us. |
The second row is the one that mattered. The old habit was to look at queries with volume and leave the rest. The new one is the reverse, and the logic is simple: a query a machine composed appears in no keyword volume tool, because nobody searches it monthly. It appears once, when somebody talks to a chat window. But it appears the same way for every competitor you have — which is to say for nobody, because nobody is looking.
It is the same conclusion we reached from a different direction when a four-day-old article ranked second in five countries. Narrow questions are won quickly. These are the narrowest that exist.
Where you are, in three thresholds
- You see no strange queries at all. Most likely you do not have the impression volume yet, or your subjects are not the ones people ask assistants about. The first step is not GEO — it is one page that answers a real question.
- You see a few, all in poor positions. You are where we were three weeks ago. The systems have found you, but late. That is fixed with pages that answer the question in the query directly, not with more text on existing pages.
- You see a few and you rank well on them. Then you already have the raw material: every machine query where you sit in the top ten is a validated topic with almost no competition. The next article is written from that list, not from an editorial calendar.
We are at the second, with a foot in the third. Worth saying where we start from: this site went live on 6 July 2026, and the SEO and GEO work on it is a few weeks old. These numbers do not measure how good an agency is — they measure what happens in one month. We like the experiment, and we publish the parts that do not flatter us, because otherwise there would be no point.
If you want to see this applied to your own site, this is where we usually begin.
Questions nobody has actually asked us
The article is a few days old, so nobody has asked anything yet. These are the questions a reasonable reader would have, plus the ones that come up in real client conversations.
If I appear for an AI-written query, does that mean ChatGPT recommended me?
No, and this is the most common confusion, so it is worth stating bluntly. Search Console tells you only that your page appeared in Google's results for that search.
It does not tell you whether the assistant read the page, used anything from it, or named you in the answer it gave the person. Those are three different things, and only the first is visible. We wrote separately about the gap between them in why AI knows your company but will not mention it.
Is this the same as Search Console's AI report?
No. The AI report tells you how much you appeared inside Google's AI features, as a total. It gives you neither the individual queries nor useful click data.
What this article describes is the ordinary Performance report, read differently: the same data you have had for years, which now happens to contain searches written by machines. More on the AI report itself here.
Should I block AI crawlers instead?
Different question, and mostly a separate decision. What lands in the Performance report is Google traffic, not a crawler visit — blocking a crawler in robots.txt does not remove these lines.
Blocking is a reasonable choice for a publisher whose business is licensed content. For a small business trying to be found, it removes you from the shelf the assistants read from.
Isn't this just GEO with extra steps?
It is one input to it, and a cheap one. Most GEO work is about being the kind of page an assistant can quote. This is about noticing which questions the assistants are already asking on your behalf.
The difference is that this costs fifteen minutes and no subscription, and it uses data you already own rather than a simulation of it.
How many impressions do I need before this works?
A few hundred a month is where patterns start to show. Below that, Google's privacy threshold hides most of your queries and you will be reading a list of five rows.
Our own numbers give a sense of the ratio: 1,986 impressions produced 148 visible queries.
Should I optimise for these queries?
Not for them individually — a fan-out query appears once and never repeats in the same form.
Optimise for what they reveal: the shape of the question. Ten machine queries on the same subject tell you what people are asking assistants about your field, and no keyword volume tool will show you that, because their monthly volume is zero.
Does it work for an online shop?
Yes, and usually better, because assistants field a lot of product comparison questions. One of the queries in our own account was “enterprise solutions for product citation in chatgpt shopping results”.
The tells are the same. The difference is that a shop will see more queries carrying a brand, a price and a “vs”.
How often is it worth checking?
Every couple of weeks, on the 28-day window. More often shows you nothing new: the window moves slowly and these queries arrive in drips, not waves.
We do it on Mondays, with the rest of the measurements.
What if I find none?
Most likely a volume problem, which is the common case for a small site. Google hides most queries: on cittago.com, 71% of impressions have no visible query attached.
Check again in a month, after publishing two or three pages that answer concrete questions. If still nothing, your subject has not reached assistant conversations yet — which is itself useful information.
Last updated: 2 August 2026. The figures come from the Search Console API for cittago.com, on the 28-day window ending 2 August 2026, and they will move. We update this page when Google changes what it reports, or when enough reader examples arrive to justify a second round.
