The buyers who hold the budget turned ChatGPT into a search engine

On 26 October OpenAI published the third release of Signals, its own account of how people use ChatGPT: one million classified messages a month, from July 2024 to September 2026. The headline went round within a day. Writing fell from 34% of messages to 21%. Seeking information rose to 20% and passed writing for one month in August.

That’s the tide. If you sell something, the more useful questions are underneath it. Who is doing the searching? What do people research at work? Does it look the same in your market? And the one nobody has asked: what does this data say about your brand?

I spent a day inside the files. Here is what I found.

First, what the data is

Four things to hold in mind, so the numbers below land the right way.

• It counts messages, not people. Every figure is a share of messages. When a share rises, more of the conversation is about that thing. It says nothing about how many people, searches, clicks or purchases sit behind it.
• It’s consumer ChatGPT only. Free, Go, Plus and Pro accounts, adults only. Business accounts aren’t in it, so anything about work is understated.
• Product questions aren’t separated out. OpenAI groups “which laptop should I buy” with “what’s the capital of Peru” and “how do I cook lentils” under one label, seeking information. So the figures below are about search in general. The shopping part is inside them, and nobody outside OpenAI can isolate it.
• The latest months move. OpenAI revises earlier months with each release. July to September 2026 is out for the first time, and it will shift a little next quarter.

Everything else in this article comes straight from the files, with the arithmetic done by me. The data is published under a Creative Commons licence, so you can download it and check.

1. The young didn’t change. Everyone else did.

The common picture is that Gen Z searches in ChatGPT and everyone older still uses Google. The data says close to the opposite.

Take each age group on its own, and ask what share of that group’s messages are searches.

Users aged 18 to 24 are where they were two years ago. Every group from 25 upward moved, and the further up the ladder you go, the bigger the move.

Someone aged 55 to 64 now uses ChatGPT as a search engine more than one message in four.

Two things could produce a rising search share overall: older people joining, or people changing how they use it. I separated the two. If the age mix had stayed as it was in July 2024 and only behaviour had changed, search would be 19.6% of messages today instead of 20.2%. If behaviour had stayed and only the mix had changed, it would be 16.7%, which is where it started. Almost the whole rise is behaviour. People over 35 changed what they use ChatGPT for.

This isn’t a quirk of one market.

Across the 104 countries where both groups have enough data, 55 to 64 year olds search more than 18 to 24 year olds in 81% of them.

The searcher has changed in another way too. In July 2024, 61% of search messages came from accounts with a typically masculine first name. By September 2026, 53%.

Put plainly: the age groups that usually hold the household budget, sign the purchase order and approve the vendor are the ones who turned the assistant into a search engine.

If your buyer is over 35, the assistant is already in the room when they decide.

2. At work, people now research vendors in the chat

Signals also classifies work messages by the activity they serve, using the same taxonomy the US Department of Labor uses to describe jobs. Most work use is what you’d expect: editing documents, making visuals, marketing copy. Look at the activities that describe a buyer, though, and a pattern shows.

Obtaining information about goods or services. In the United States this was 1.7% of work messages in the first quarter of 2025. By the third quarter of 2026 it was 2.9%, and it moved from the 11th most common work activity to the 5th. The same rise happened in 70 of the 74 countries with data for every quarter, and most of Europe is ahead of the US.

Determining the value or price of goods or services. More than doubled over the same period: 0.3% to 0.7% of US work messages, up in 69 of 74 countries.

Explaining the technical details of a product or service. Fell by half, from 1.2% to 0.6% of US work messages, and by the same proportion across the 74 countries.

Fewer people ask the assistant how a product works. More ask it which one to get and what it should cost. That is the shape of a buyer moving from learning to choosing.

The choosing now happens in a chat window, on a personal account, two or three questions deep, where no vendor can see it.

Remember this is consumer accounts only. The company accounts, where most purchasing conversations happen, aren’t counted. Whatever these numbers show, the real picture is larger.

3. Markets don’t adopt AI the same way, and search follows adoption

Here is where it stops being one story.

In the third quarter of 2026, search was 31% of messages in Slovenia, 30% in Croatia, 29% in Slovakia, 28% in Czechia and 27% in Hungary. In the same quarter it was 14% in Indonesia, 15% in Nigeria, 16% in Brazil and 16% in India, where writing still leads. Spain sits at 26%, France at 26%, the UK at 23%, Germany at 23%, the United States at 21%. Japan is the outlier in a different direction: 17% of its messages are self expression, about double the United States.

There’s a pattern under this spread. OpenAI ranks countries by messages per person. Early in 2025 a country’s rank told you little about how it used the tool: the correlation between per person use and search share was 0.30. By the third quarter of 2026 it was 0.74. The countries with the heaviest use per person put 24% of their messages into search; the lightest quarter of countries, 16%. Markets that adopt deeper become search markets. Estonia, for example, ranks 4th of 173 countries for use per person, and nearly a quarter of its messages are searches.

For a brand this means “how are people using AI” has no single answer. The answer is per market, and in most of Europe the answer is already “to search”, at a share that is still rising.

4. The conversation changed shape

One more figure, because it changes how you should think about measuring any of this. OpenAI sorts messages by what the person is doing: asking for information, asking the assistant to do a task, or simply expressing something with no task attached.

People aren’t typing keywords into the assistant. They’re talking to it, thinking out loud, telling it what they want and why, and the question about your product arrives inside that.

Any method that tests an assistant with the same short query every week is measuring a shape of conversation that's shrinking.

What this data can’t tell you

Signals is the best public view of the tide that exists today, and I’d rather have it than not. But read the four caveats again. It counts messages. It can’t separate a product question from a recipe. It covers no business accounts. And it carries no brand names at all.

So it can tell you that your buyer, over 35, in France, at work, is now asking an assistant which vendor to pick and at what price. It can’t tell you what the assistant answers. Whether it names you. What it says when it does. Which competitor it reaches for instead, and why. Where in the conversation the buyer stops considering you: at the first list, at the comparison, at the moment they ask “should I go with them?”

That's the part that decides the sale, and it's different for every brand, every category and every market. There's no release to download for it.

It has to be researched, for your brand, in your market, question by question.

We do that research for brands. If you want to know what the assistants are telling your buyers at each step, and where you lose them, start with the free assessment.

Data: OpenAI Signals v3.0, OpenAI, 2026, released 26 October 2026, licensed CC BY 4.0. All figures are shares of classified consumer ChatGPT messages; age group figures are derived from OpenAI's published age by topic, topic and age tables. Country figures for a quarter are the mean of its three months. Full download: openai.com/signals/data-download.