What a million ChatGPT messages a month say about marketing

Summary

On personal ChatGPT accounts, marketing work is mostly production. About one work message in twenty asks for marketing or promotional material. The core strategy tasks get about a third as many messages as the core production tasks.

• Marketing material is a top-three work task. It took 5.1% of work messages in Q3 2026. It has ranked second or third of 164 work activities every quarter since early 2025.
• Production outweighs strategy about three to one. Marketing materials and commercial writing take 7.4% of work messages. Market analysis, consumer research, planning, pricing and similar strategy tasks take 2.5%. Count eight broader business-strategy tasks too, and production still leads by 1.7 to one.
• The growth is visual. Visual design rose from 3.1% to 7.9% of work messages between Q1 2025 and Q3 2026. Writing held between about 37% and 42% of work messages until mid-2026, then fell to 33% by September.
• Strategy is growing slowly. It is up about a fifth since early 2025, mostly from pricing work. Market analysis fell, and the latest quarter is noisy.
• The Netherlands sits in the upper middle. Averaged over the last four quarters, marketing material takes 5.5% of Dutch work messages. That is close to the US (5.2%) and the UK (5.1%), and about half the level in Nigeria, Brazil and Argentina.

The data covers personal accounts only. ChatGPT Enterprise, Edu and Codex are excluded, so strategy work on company seats does not show up here.

The data

OpenAI has published version 3.0 of its Signals data, with its own write-up of how ChatGPT use at work is spreading. It is one of the largest public samples of what people ask an AI assistant to do. Each month OpenAI samples a million messages from adult users of its personal plans, leaving out deleted accounts and anyone who opted out of training. Models label the messages and researchers never see the text. The published shares carry added privacy noise.

The most useful table for marketers sorts work messages into O*NET work activities, the US Department of Labor's catalogue of what jobs involve. It covers 164 activities in 94 countries, quarterly from Q1 2025 to Q3 2026.

We picked out the activities that describe marketing work and grouped them by stage: understanding the market, deciding the strategy, making the material and selling. Then we compared them across 74 countries and seven quarters. The charts below switch between the 74-country average and the US, and the country table sorts.

A caution before the numbers. The data describes tasks and has no job-title field, so the marketing work here could come from marketers, founders, freelancers or students. It also leaves out company accounts entirely. We come back to both points below.

The numbers

Better, not just faster

Making marketing material cheaper is a real gain. A client of ours spent about eight hours learning to make eBooks, flyers and print-ready PDFs in-house with Claude. That saved roughly €10,000 set aside for a freelancer. It also helped redirect six figures of budget earmarked for outside help toward building the capability in-house.

The question this data raises is what all that production is aimed at. Every flyer, deck and landing page rests on decisions made upstream. Which accounts do we go after, what do we tell them, what do we charge and what do we stop? On personal ChatGPT accounts, those decisions get about one message for every three spent producing the material.

Much of this usage comes from small businesses and freelancers, so it covers a wider market than B2B SaaS. Our own research points the same way for B2B teams, though. In October we analysed 466 B2B marketing job ads. 163 of them asked for at least one capability in goals, segmentation or targeting. Even so, those three areas made up only 5% of all the capabilities mentioned. ICP definition came up in 11 ads and market segmentation in 3.

Hiring and AI use both lean away from goals, segmentation and targeting. That matters because AI accelerates whatever system it lands in. In a team chasing this quarter's pipeline, cheaper output brings more of the same campaigns, sooner.

The same tools make it almost free to try work a team never had the time or skills for. A proper segmentation and a pricing test are good places to start. That is the better route. The rise in pricing work hints that some people are taking it, though a few quarters of noisy data is thin evidence.

If you lead a marketing team, three questions are worth asking of your own AI use:

1. What share of it produces more of what we already make, and what share helps us decide what to make?
2. Which capability have we always bought in, rushed or skipped, that we could now build ourselves?
3. If we try that capability, how long does it get, what would make us stop, and who owns it?

Company seats may hold more strategy work than this data can see. Treat it as a prompt to look at your own team's AI use, and at which side of that three-to-one ratio it falls.

What this data can and can't tell you

• Personal accounts only. The data covers ChatGPT Free, Go, Plus and Pro. ChatGPT Enterprise, Edu and Codex are excluded, and OpenAI says this likely understates business use. In-house teams on company seats, where more strategy work may happen, are invisible here.
• Tasks, whoever does them. There is no job-title field, so a message counts as marketing work because of what it asks for. The countries with the most marketing material, such as Nigeria, Brazil, Argentina and Pakistan, suggest a lot of small-business and freelance work.
• A generous work filter. The activity table only covers messages flagged as work. That flag marks about 32% of messages, against 21% on OpenAI's stricter work, school or other classifier. Some of it will be coursework or personal projects.
• One label per message. A model gives each message a single activity and needs direct evidence that the person is doing it. Explicit requests such as “write a flyer” are easier to label than strategic thinking, which may land under editing or business advice.
• Message shares. Every figure is a share of work messages. Coding's share fell from 7.8% to 3.1% over the period, and that fall lifts every other share. Codex and company plans sit outside the data. A flat share in a growing product can still mean more messages.
• Our grouping. The stages are our judgement, so the ratio is shown under four definitions, from 1.7 to 6.8 to one.
• Small numbers are noisy. The country figures weight every country equally. Most strategy activities sit below 0.5% of work messages. The latest quarter has been published once and may be revised.

Source and licence

Data. Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Drew Johnston, Alex Martin Richmond, Christopher Ong, Carl Yan Shan and Kevin Wadman, OpenAI Signals v3.0, OpenAI, 2026, downloaded 26 October 2026. Licensed under Creative Commons Attribution 4.0.

Method. Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman, “How People Use ChatGPT”, NBER Working Paper 34255, 2025, doi:10.3386/w34255.

Changes made. Gossamer selected 25 of the O*NET work activities and grouped them into stages. We averaged them across the 74 countries reported in every quarter, and over the last four quarters for the country comparison. We calculated the ratios and changes shown. OpenAI has not reviewed or endorsed this analysis.

Further reading

• The Activation Trap: Why Marketing Leaders Systematically Under-Invest in Long-Term Growth
• What B2B startups get wrong (and right) when hiring marketing leadership

Curious where these patterns show up in your own marketing? The Gossamer Audit scores 17 categories, traces every score to evidence and hands back one prioritised roadmap.