I wrote the first version of this piece in May, when OpenAI opened self-serve ads inside ChatGPT. My thesis was that targeting would move from the keyword to the conversation. This autumn I put it to the test with motive, the product I'm building to observe what AI assistants answer when someone asks about a market. The first real measurement took apart half my reasoning. This piece is about that half.
The corrected version: a keyword is no longer enough as a brief, you have to describe the buyer's situation. Then you have to measure separately three things the “AI acquisition” story lumps together: the ads you pay for, the recommendations the assistant makes on its own, and business outcomes.
Describe the whole buyer, and remember the platform decides the match.
In the May draft I compared two requests. “best CRM for startups” is compressed intent. “I run a seven-person agency, our leads live in spreadsheets, Salesforce is too heavy, I need something cheap that works with Gmail and can be live this week” is the whole need. OpenAI's creative guidelines ask for exactly this: what you offer, for whom, when it helps.
That part holds. What was missing was a boundary. The hints you give the platform steer who sees the ad and guarantee nothing, so users' questions aren't inventory you can buy. OpenAI keeps ads separate from the answer: advertisers don't buy the assistant's recommendation, they get aggregate reports, and conversations stay private. In my first take, “mapping conversations” was sliding towards “buying conversations”. Only the first promise is honest.
A complete demo with made-up numbers can't prove you wrong.
motive started at a hackathon in Paris as an AI-native acquisition demo: from the brand's link to customer intent, then human review, ads, creative, a fake launch and a dashboard whose numbers told a coherent story. It was convincing precisely because it was complete: beautiful, and the numbers were fake.
This autumn I removed the simulated launch and the synthetic numbers. An incomplete experiment with real observations is worth more than an end-to-end demo with invented data, because only the first can prove you wrong. If you build a demo like that, label the simulations and keep three questions apart: does the process work, does the measurement hold up in the real world, will anyone pay.
Asking a model who it would recommend isn't observing a recommendation.
motive's first check asked a model which brands it would mention. It answered from memory, once, with no language or country, on questions it had invented itself. Out came a plausible list of names, and that's the trouble: the wrong method also produces credible results.
Now motive puts the real question to each assistant, with web search on, repeats it several times per language and per assistant, and keeps every answer. Visibility becomes a frequency with its uncertainty stated. And an observation made from outside stays one: the app people actually use has memory and personalisation we don't reproduce, and that needs saying.
Your questions have to ask for someone to choose, and they have to come from buyers.
The first real measurement on our own company was partial and we discarded it. It still exposed the biggest flaw. The questions were almost all “how do I” questions, many lifted from the titles of our blog. No answer named Yempik or a competitor, by construction: an assistant doesn't name vendors if nobody asks for them. The same day we tried a single buyer question, “Which software houses in Italy build custom AI agents for SMEs?”, and Claude named Yempik first. It was a single check, so no champagne. It was enough to show where the mistake was.
Now every question states what it asks the assistant: who to choose, how to choose, how something works. The panel is mostly the first kind and results are read separately. Informational questions still show one useful thing: which sites the assistant trusts.
The second mistake was where the questions came from. Deriving them from the company blog looked like a serious method. Your site tells you about your positioning and says little about real demand: it kept the experiment inside our own assumptions, and one assistant cited our site precisely because the questions echoed our pages. Now questions come from what people actually search for, each with its origin, and if search returns nothing the system stops instead of making them up. One detail confirmed the May thesis from an unexpected side: search suggestions only come back when you start from a few words. When people search, they still use keywords.
Read, cited, named and recommended are different outcomes.
In the first measurement one assistant listed pages from our site among the sources of many answers that never named us. A single “visibility” score would have hidden the most interesting signal: the assistant used us as a source without putting us forward as a vendor.
Now being named and being cited are two separate measures, never merged into one score, and pages cited in the answer are kept apart from pages that were only read. Where an assistant shows no sources, the report says “no sources”. A zero there would describe how the assistant searches, and the client would read it as a verdict on the brand.
Web search being on also needs checking. In the first measurement one assistant had the tool and used it in very few answers, and we were keeping only text and sources. Now we save every answer in full, along with whether the assistant searched, with which words and which pages it opened. Those searches are the reason for the title. The keyword has disappeared from the interface and stayed inside the machine: the assistant rewrites the conversation into short searches and then picks pages. You can read those rewrites and work on the pages that get picked.
A better method can fake growth for you.
Moving from “how do I” questions to buyer questions changes the results even when the market stands still. Our history chart put measurements made with different questions and methods on a single line, so a movement could come purely from how we had changed the questions. The fix is still in progress: new questions are added without replacing old ones, every trend compares the same questions with the same method, and a change of method visibly breaks the line.
The strongest objection: your number isn't what a real user sees.
It's true. An observation from outside doesn't reproduce the app with its memory and personalisation, and motive is still being built. My answer is to state exactly what you measure: how many times out of how many, with which method and how much uncertainty, next to the answers that produced the number. I'd rather have a small number the client can rebuild on their own.
From Monday
- Write the brief as a situation: goal, constraints, alternatives, objections, fitting offer. Separately, list the controls the platform really gives you.
- Label every question in your panel with what it asks the assistant, and read results by type of question.
- Take your questions from what buyers search for and keep the origin of each one. Your blog measures your positioning.
- Record separately whether the assistant searched, what it read, what it cited and who it named.
- Keep the original answers and track every change of method and questions. Only compare measurements taken the same way.
- Keep separate reports for ads, unprompted answers, visits and outcomes. Write down what links them and what you aren't measuring.
- Label as simulated everything that is simulated.