Your Customers No Longer Search, They Brief the AI

The Google Search Console for this website holds a search query 46 words long. It opens with "chatgpt beschreibt unser unternehmen mit veralteten produktdaten" and closes with "suche online nach anbietern und nenne mir konkrete namen". It was served 84 times in August. Nobody clicked, because the page sat at position 62 on average.

Two rows below sits a query no person can have typed: the word "chatbot" in quotation marks, followed by eleven exclusions such as -site:reddit.com and -site:youtube.com. The website ranked at position 1 for it. Nobody clicked here either.

Key points: People looking for a provider today often stop typing keywords. They describe their situation to an AI assistant and ask for provider names. The assistant then researches on its own, and both land verbatim in Search Console: the written-out briefs and the machine-shaped queries the system searches with. In our data from 2 to 29 August 2026, briefs averaged position 61.2 and short keywords 9.8; visibility is not what is missing, fit is. Machine-shaped queries, by contrast, rank better than average (6.58 instead of 10.03), grew by a factor of 7.8 within five weeks and produced exactly one click across 218 impressions. Click-through rate therefore increasingly measures only what share of demand is human. What counts are the details on your own page: trigger, audience, territory and exclusions.

What is actually sitting in Search Console?

Classic search queries are short. "ki agentur münchen", "whatsapp business api kosten" - three words, sometimes two. Alongside them, for some months now, sit rows that read like they were pasted out of an email:

"uns ist aufgefallen, dass ki-assistenten unsere marke diffus und teilweise falsch beschreiben, vermutlich weil unsere eigene positionierung online nirgends klar formuliert ist. wir wollen die markenpositionierung so schärfen, dass sie auch von suchmaschinen und ki korrekt aufgegriffen wird. welche beratungen oder agenturen können markenstrategie und digitale sichtbarkeit zusammen denken? recherchiere online und nenne mir konkrete anbieter."

Fifty-six words. Served sixteen times, position 80.

Rows like these appear when an AI assistant looks something up on the web to build its answer. The assistant hands the brief, in whole or in part, to a search engine, and Search Console logs it like any other query. So you end up reading what the customer dictated to the system before the answer existed.

How long is a query today?

We evaluated the period from 2 to 29 August 2026: 372 query rows carrying 8,672 impressions. A row counts as a prompt brief only when three traits coincide: at least twelve words, an instruction or complete question, and a reference to the querant's own company. That filters out long technical terms and pasted error messages.

The spread is the actual finding. Same website, same content, same technical foundation - and a gap of more than fifty positions between the two query types.

Real customers or just tooling?

This question has to be settled before anyone builds a strategy on 211 impressions. The honest answer: some of these queries almost certainly come not from prospective buyers but from tools that measure AI visibility.

The tell is repetition with a swapped country. The same wording about outdated product data appears three times: once with no country reference at 84 impressions, once with "in deutschland" at 21, and once with "in österreich" at 19. A genuine prospect writes their situation out once. Anyone pushing the same sentence through country variants is measuring.

That changes the conclusion less than it first appears. Such tools deliberately mirror the phrasing buyers actually use, because otherwise their measurement would be worthless. The position that comes out of it is real. And the rest of the set carries queries that read like individual cases: an aesthetic clinic with outdated prices, a fintech that surfaces sometimes and sometimes does not, a board demanding figures before releasing budget.

What do the market figures say?

The internal observation lines up with the surveys. In March 2026 G2 surveyed 1,076 B2B decision-makers across North America, EMEA and Asia-Pacific. On the research phase:

Two of those rows matter most to mid-sized providers. The 33 percent who bought from a previously unknown vendor describe a door that classic search never opened: name recognition is no longer a precondition for making the shortlist. And the 41 percent using deep research explain the word counts in Search Console. Deep research runs break a brief into sub-questions and ask them in sequence, which produces exactly these long, written-out rows.

Why does the position collapse on long queries?

A keyword sets one condition. "ki agentur münchen" asks for a service and a place. A prompt brief sets four or five at once, and they are the kind that rarely appear together on one page.

Take the query about the software category: 39 words, twelve impressions, position 65. Inside it: the company sells software. AI-generated comparisons describe it incorrectly. Competitors are portrayed correctly. An agency is wanted. That agency should be able to improve the company's portrayal in ChatGPT specifically. And actual provider names should be returned.

A page headed "AI visibility" that explains what GEO means satisfies one of those conditions. It is topically correct and still the weaker answer compared with a page describing the case itself: false product details in competitive comparisons, who is affected, how the correction runs.

This is not a ranking problem in the classic sense. The page is indexed, it is fast, it is structured. What it lacks is overlap with a description of a situation.

What does a prompt brief demand from a page?

The sixteen queries in the set share a set of recurring components. Every brief contains at least four of them.

That last point is the one most often missed. When a system is told to name providers, it needs a paragraph from which a name, a region and a service can be lifted together. If the company name lives only in the legal notice and the service only in a heading, the page works as a source but not as an answer to "name specific providers".

Which queries could no person have typed?

Next to the customers' written-out briefs, Search Console holds a second class: the queries the AI system itself searches with. Anyone who cannot separate the human part of their visibility from the machine part is reading their Search Console incorrectly. Three features rule out a human origin in practice.

First, operator syntax on a scale nobody types by hand. The query mentioned at the start looks like this in full:

"chatbot" -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

That is roughly 200 characters of pure filter logic for a single search term. A person wanting to keep forums out of their results types two or three exclusions and then gives up.

Second, characters a search box will not pass through. The Bing data for the same website holds this query:

anthropic/claude-opus-4-8#N##N#heute um 15:04#N##N#klar, hier ein paar alternative überschriften mit anderem thematischen fokus:#N##N##N##N#klassiker mit netz

The #N# is an encoding for a line break. A search box does not accept line breaks; the return key triggers the search. A query like this arises when a program hands over a multi-line text as a search term; the embedded timestamp and the half-finished sentence show it comes from a chat transcript in progress.

Third, conversational leftovers with no topic of their own. The data includes also bis wann?, ja beides, wie nutzt man es?, when did it come out?, er det lovligt, daj link do artykulu and что ты путаешь. Not one of these is a search query. Each is an answer or a follow-up inside a dialogue that an assistant passed verbatim to a search engine.

Taken individually, each feature can be explained away. Taken together they form a pattern that can be measured cleanly.

How fast is the share of machine queries growing?

More than twice as fast as total volume. For this analysis we compared two non-overlapping windows and converted both to impressions per day. A row counts as machine-shaped if it begins with a plus or percent sign, contains quotation marks, or carries a -site: exclusion. That is a deliberately narrow definition: conversational fragments and fully worded briefs fall outside it, so the true figure is higher.

The factor of 3.5 in total volume has an identifiable cause: a new AI model was released on 1 September, and this website reported on it early. Part of the increase is pure news traffic. That is precisely why the comparison matters: the machine-shaped class is not merely riding the news surge; it is gaining share on top of it.

Why do machines rank better and still not click?

Because a search system can serve a precise query precisely, and because there is nobody at the other end to click.

Putting "chatbot" in quotation marks and excluding eleven platforms narrows the result set so far that specialist pages rise to the top. The system then reads the result list, fetches the page directly where needed, and composes an answer inside the chat. The retrieval happens; it simply does not show up as a click.

This shows most sharply in a paired comparison from the same period on the same destination page:

Near-identical impression counts, near-identical positions, the same destination page. The two human queries return 12 clicks across 57 impressions, the two technical identifiers zero across 54. The only difference is who was asking.

The written-out briefs follow the same pattern. Using the simpler word-count definition alone, they averaged position 59.4 at the end of August and 18.6 in the window up to 7 September. The position improved by 40 places, and still nobody clicked. A position can be improved; a machine sender cannot be made to click.

What do the Search Console AI reports show, and what not?

Impressions yes, clicks and queries no. On 3 June 2026 Google announced dedicated performance reports for generative AI features on the Search Central blog, and the rollout to every property was completed on 31 August 2026. There are two reports, one for AI Overviews and AI Mode in Search and one for generative features in Discover. Both show impressions by page, country and date, and the Search report also by device.

They show no clicks, no click-through rate, no position and, above all, no queries. They are also unavailable through the API and can only be exported by hand. The reports confirm that your page appears in AI answers, but withhold which question prompted it and with what result.

The analysis presented here therefore takes a different route. It looks for the machines not in the AI reports but in the ordinary query report, where they remain identifiable by the shape of their queries and where wording, position and click-through rate are available.

What has to be on your own website?

The details briefs ask for, complete on each page. There are two reasons.

The first lies in the briefs themselves: they set four or five conditions at once, and the page that wins is the one answering as many of them as possible in a single place. The second lies in how the systems search. The eleven exclusions in the opening query hit forums, social networks, review and booking portals, the very platforms many visibility guides treated as especially important. Published analyses of ChatGPT fan-out queries show the same movement from the other side: according to Peec AI data, the share of the site: operator, which restricts a search to exactly one domain, rose from around 0.3 percent to about 23 percent of all fan-out queries when GPT-5.6 became the default. Nectiv, testing around 4,000 prompts through OpenAI's API, even found 64 percent, and there the average number of fan-out queries per prompt rose from 2.17 in its previous year's study to 7.61.

The consequence is uncomfortable. In an open search, a system can fill gaps from third-party sources. In a search restricted to your domain, that fallback does not exist: what is not there does not exist for the purposes of that check. Four details cover most of the conditions:

• Name the trigger, not just the service. Instead of "we improve your visibility in AI systems", the sentence that describes the situation: "A customer quotes a product detail that stopped being true two years ago." A prospect's brief almost always opens with an event like that.
• Spell out who it is for. Not "companies of every size", but the actual range you are willing to be measured against, for instance law firms, medical practices, software vendors and specialist retailers with ten to two hundred staff.
• State the territory, even when it seems obvious. On site, remotely, within what radius. Briefs nearly always carry a geographic reference; writing only "nationwide" loses to a provider who names regions.
• Include the exclusions. What a service does not cover is worth more to a verifying system than another list of strengths, because it answers a question conclusively.

Not by building a separate page for every conceivable phrasing: that produces thin content that gets recognised as such, and the number of possible phrasings is unbounded anyway. These are four paragraphs on the pages you already have, not a new architecture. Whether any of this reaches AI answers at all depends on which sources the systems draw from; that is covered in the article on why your company does not appear in AI answers and which sources change that.

How do you find these rows in your own data?

With three filters, in a few minutes, on any property. Open the performance report in Search Console, add a new filter on "Query" and choose "Custom (regex)". Then enter each of the following in turn:

Note impressions, clicks and average position for each group, and convert your windows to impressions per day, or you will be comparing periods of different length. Read the rows themselves as well: the wording shows which question somebody asked before an assistant turned it into a search. How a full measurement of AI visibility is built beyond this is covered in the piece on measuring AI visibility.

What metric replaces click-through rate?

Three metrics together. If a growing share of demand structurally does not click, click-through rate increasingly measures only the human share of senders. That is exactly the case in this website's data: between the two windows, volume rose by a factor of 3.5, average position improved from 10.31 to 9.72, and click-through rate fell from 2.08 to 0.50 percent. Read the last figure alone and it looks like a collapse.

The number of distinct queries you are served for. It shows how many different situations your pages are considered a possible fit for. It rises before positions rise, and long before clicks arrive.

Position split by class. The overall position blends groups with entirely different behaviour. Split apart, you can see whether the work on your content is landing where decisions are made; for briefs, the position gap to keywords is the direct gauge.

The number of first contacts who name an AI assistant as their source. This number appears in no tool, because there is no referrer for it. It exists only if somebody asks, in the first conversation, how the contact came about. It is the most inconvenient and the most important, because it is the only direct evidence that machine visibility turns into business.

What should you do now?

Four steps, ordered by effort.

1. Measure once how large your machine-shaped class is. The three regex filters from the previous section, two time windows, half an hour of work. Without that figure, every further measure is guesswork.
2. Add the source question to the first conversation. One line in the conversation guide. Without it, the entire AI channel stays unprovable, however good the impression figures look.
3. Check your most important pages for five details. Scope of service, area covered, target audience, named contact, explicit limits. Not spread across the site but complete on each page, because a search restricted to your domain checks individual pages, not your website as a whole.
4. Stop reporting overall click-through rate as a success metric. As long as the machine-shaped class grows, that figure falls even when everything is going right. Anyone keeping it in the monthly report will regularly have to present good work as a setback.

If something is already said about your company in AI answers and it is wrong, the article on false and vague AI claims is the right starting point: this piece covers the shape of the query, that one the correction of the content. Both rest on the same foundation: a documented baseline of what the systems say about your company today and where they got it. How that baseline is built is described on the AI visibility page.

Frequently asked questions

What is a prompt brief? A search query that is no longer a keyword but a written-out instruction to an AI system. It describes a situation, names a problem and usually ends with a directive such as: search online for providers and give me actual names. In our data the length runs from 13 to 56 words, averaging 43.

How can I tell that a search query came from a machine? By three features that almost never appear in a search box. First, operator syntax: quotation marks around whole phrases, a leading plus sign, several minus-and-site exclusions in a row. Second, characters a search box will not even accept, such as encoded line breaks or a timestamp in the middle of the query. Third, conversational fragments with no topic of their own, which only make sense inside a dialogue. Any one feature is weak on its own; together they are conclusive.

Are these real prospects or automated tools? Both, and Search Console data cannot draw the line cleanly. When the same wording appears in variants for Germany, Austria and Switzerland, that points to a monitoring tool. Those tools deliberately mirror the phrasing buyers use, however, which is why the positions remain meaningful.

Why does our page rank worse on long queries? Because a forty-word brief sets several conditions at once: service, region, company size, trigger. A page optimised for a single keyword satisfies one of them. The system looks for the source that answers as many conditions as possible in one place.

Why do machine-shaped queries rank better yet never click? They rank better because a search system can serve a precisely worded query with quotation marks and exclusions more accurately than a vague keyword. In our data this class sits at position 6.58 against 10.03 across the whole property. The click never comes because there is nobody at the other end to click. The system reads the result list, fetches the page directly if it needs to, and delivers the answer inside the chat. The visit happens; it simply is not counted as a click.

Does Search Console not show these AI impressions by now? Partly. Google announced dedicated performance reports for generative AI features on 3 June 2026 and completed the rollout to every property on 31 August 2026. Those reports show impressions from AI Overviews and AI Mode, broken down by page, country, device and date. They show no clicks, no click-through rate, no position and no queries, and they are not available through the API.

Do I now need a page for every conceivable prompt? No. That does not scale and is recognised as thin content. The better move is to extend existing pages with the details that recur in briefs: who the service is for, which area it covers, what event triggers it and what it explicitly excludes.

What does this mean for my own website? When an AI system is asked to answer a question about your company, it increasingly checks your own domain specifically. What is not there, it cannot confirm, and what it cannot confirm it either omits from the answer or replaces with an older figure from a third-party source. Scope of service, area covered, target audience, named contact and explicit limits therefore belong in plain text on your own pages.

What metric replaces click-through rate? No single one, but three together still carry meaning. First, the number of distinct queries you are served for at all. Second, position within the machine-shaped class, measured separately from the rest. Third, the number of first contacts who say they found you through an AI assistant. That third number appears in no tool; it has to be asked for in the first conversation.

Do I now have to write for machines instead of people? No, and attempting it would backfire. What machine-shaped queries reward is exactly what rewards a reader in a hurry: a clear statement up front, verifiable figures with a date and a source, clean headings, and no claim without evidence. The difference lies not in style but in completeness. A person fills gaps from context; a system does not.

Is this worth it at only one or two percent of impressions? The share is small today, but the intent behind it is the strongest in the entire dataset, and it is growing: in our data the machine-shaped class grew almost eightfold within five weeks, while total volume grew three and a half times. Someone who writes out a situation and asks for provider names is close to a decision, unlike someone looking up a technical term.

Sources

• Google Search Console, arkeontech.de property: own data, retrieved on 30 August and 7 September 2026
• Bing Webmaster Tools, arkeontech.de property: the quoted query verbatim, including the #N# encoding for line breaks
• G2: The Answer Economy: G2's 2026 AI Search Insight Report - market figures, surveyed in March 2026 among 1,076 B2B decision-makers across North America, EMEA and Asia-Pacific, supplemented by interviews with 39 B2B software marketers
• Google Search Central blog (3 June 2026): Introducing Search Generative AI performance reports in Search Console - announcement of the performance reports for generative AI features, with the note on the completed rollout on 31 August 2026
• Search Console Help: Generative AI performance report (Search) - impressions as the only metric, breakdowns and export via the export button, retrieved on 14 October 2026
• Search Console API: Search Analytics: query - permitted search types, none of them for generative AI features, retrieved on 14 October 2026
• Search Engine Journal (31 August 2026): Google Search Console AI Reports Rolled Out Worldwide - report on the completed rollout, without click data in the report
• Search Engine Roundtable (1 September 2026): Global: Google Search Console AI Performance Report & AI Control - report on the completed rollout, without click or query data in the report
• Peec AI, David Konitzny (10 August 2026): ChatGPT's New Default Model GPT-5.6: More Retrieval, More Fan-Outs, More First-Party Content - share of the site: operator in fan-out queries before and after the switch to GPT-5.6
• Nectiv, Chris Long (13 August 2026): New Research: ChatGPT Tripled It's Fan-Out Queries + Looks For Authoritative Sources - share of site: searches and fan-out queries per prompt compared with the previous year's study
• Lily Ray (17 August 2026): What We Can Learn from Evolving ChatGPT Fan-Out Queries - overview of the analyses of ChatGPT fan-out queries, retrieved on 7 September 2026

Methodology: the time windows are given in the text. The prompt brief analysis rests on 372 query rows carrying 8,672 impressions from 2 to 29 August 2026, delimited by the three traits named in the text. The machine-shaped class was formed using the regular expressions given above and can therefore be reproduced on any other property.

This article was published on 30 August 2026. On 23 September 2026 the article on machine queries in Search Console from 7 September 2026 was merged into it. On 14 October 2026 we made three statements more precise: the G2 row on thinking and reasoning models, the fan-out measurements with their originators, and the description of the Search Console AI reports, which exist separately for Search and Discover. The sources are now listed, together with the primary studies and Google's pages on the reports.