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.

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. Those briefs land verbatim in Search Console. In our data from 2 to 29 August 2026 there are 16 of them, drawing 211 impressions. The positions reveal the real problem: on queries of up to four words the site averages position 9.8, on written-out briefs 61.2. Visibility is not what is missing, fit is. A forty-word brief sets 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.

What is actually sitting in Search Console?

Classic search queries are short. "ki agentur münchen", "whatsapp business api kosten", "ox alpha" - 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 a prompt brief demands 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".

How do you write for prompt briefs?

Not by building a separate page for every conceivable phrasing. That produces thin content, gets recognised as such, and the number of possible phrasings is unbounded anyway.

The more effective route runs through the pages you already have. Four additions cover most of the conditions that recur in briefs:

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. "For mid-sized companies" is too vague to satisfy a condition. "Law firms, medical practices, software vendors and specialist retailers with ten to two hundred staff" can be checked.

State the territory, even when it seems obvious. 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 as much to a selecting system as what it does. It keeps the page out of unsuitable briefs and makes it more credible in fitting ones.

Whether any of this reaches AI answers at all depends on which sources the systems draw from. Which ones those are, and how to get into them, is covered in the piece on the sources behind AI answers.

How do you tell whether it is working?

In the same place the finding came from. The evaluation repeats in Search Console without extra tooling: filter queries by word count, for instance with a regular expression matching twelve words or more, then read the average position of that group separately from the short queries.

Three figures carry meaning when tracked across several months:

• The position gap. In our case 51.4 points. It measures directly whether the content covers situations or only keywords.
• The count of distinct prompt briefs. If it rises, the site is being considered for more situations, even while positions remain poor.
• The first click above zero. At position 61 nobody clicks. The first click out of a prompt brief signals that the page has moved into visible territory.

These three numbers do not replace a full measurement of AI visibility; they are one cut through a data source you already hold. How a defensible measurement is built and where its limits lie is handled in the piece on measuring AI visibility.

What this means for the months ahead

The 2.4 percent share will shift, and the direction is foreseeable. When 51 percent of decision-makers begin research with an AI chatbot more often than with Google, and 41 percent regularly run deep research, the share of queries that start life as a written-out brief grows.

What stands out is less the growth than how intent is distributed. The 8,154 impressions from short queries come mostly from people looking up a technical term. The 211 from prompt briefs come from companies describing a concrete problem and asking for provider names. That is the difference between reach and demand.

For your own website this does not mean rebuilding the structure. It means extending every page that describes a service with the details briefs ask for: trigger, audience, territory, exclusions. Four paragraphs, not a new architecture.

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.

Why do these queries show up in Search Console at all? Because AI assistants research on the web to build their answers. When the system passes the brief on to a search engine, in whole or in part, Search Console logs it like any other query. That is why sentences of over forty words sit alongside classic keywords.

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.

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.

How do I spot these queries in my own data? Filter Search Console queries by word count, for instance with a regular expression matching twelve words or more. What remains is almost exclusively prompt briefs. Viewing the positions of those rows separately from short queries exposes the gap immediately.

Is this worth it at only two percent of impressions? The share is small today, but the intent behind it is the strongest in the entire dataset. Someone who writes out a situation and asks for provider names is close to a decision. By contrast, most short queries come from people looking up a technical term with nothing to commission.

Sources and status

The website figures come from Google Search Console for the period 2 to 29 August 2026, retrieved on 30 August 2026. The basis is 372 query rows carrying 8,672 impressions. Prompt briefs were separated out using the three traits named in the text.

The market figures come from the AI Search Insight Report by G2, published in 2026, surveying 1,076 B2B decision-makers across North America, EMEA and Asia-Pacific in March 2026, supplemented by 39 interviews with B2B software marketers.

If you want to check how your own company is currently portrayed in AI answers, the entry point is the piece on why AI does not mention a company at all.