OpenAI's sample of ChatGPT work messages shows the ten most common activities falling from 56.0% to 51.1% of the total. For CPG brands the lesson is practical: pick specific jobs, measure them against a baseline, and run that alongside AI shopping visibility work.
AI at Work Is Spreading Out. What That Means for CPG Brands
By Stephen Honight, Founder of Lmo7
OpenAI has published a large-scale look at what people ask ChatGPT to do at work, and the finding that matters for a CPG team is about mix. Over the period the release covers, the ten most common work activities fell from 56.0% to 51.1% of work messages (OpenAI, “ChatGPT and the expanding world of work”). More of the work reaching ChatGPT is landing on a wider range of jobs.
I would not read that as a case for an AI transformation programme. My sense is that the useful reading is smaller and more practical. People are bringing AI into specific, everyday jobs, and a specific job can be measured. You can time it and count the errors. A programme with a grand name is much harder to judge.
My recommendation would be to pick one job, measure it against a baseline, and run that alongside the work that decides whether AI shopping assistants recommend your products. Those are two tracks with different jobs to do, and I would run both, in sequence.
A note on sources
Every figure about ChatGPT use in this post is from OpenAI. The activity count (165), the country coverage (nearly 100), the Q1 2025 to Q3 2026 window, the 74 countries used for the average and the 56.0% to 51.1% shift come from the release, “ChatGPT and the expanding world of work”. The sample size, the account types included and excluded, the way OpenAI classifies messages and the definition of its “seeking information” category come from the Signals methodology document, which I read on 28 October 2026.
The pilot workflows, the view on shopping assistants and every recommendation are Lmo7’s own interpretation, and I mark them as such. The one result from Lmo7’s client work carries its own source where it appears.
What OpenAI measured
OpenAI classified 1,000,000 consumer ChatGPT messages each month, from July 2024 to September 2026, which comes to about 27 million messages (methodology document). That is a count of messages, so I will not call it users.
The work-related messages were classified into 165 work activities across nearly 100 countries, covering Q1 2025 to Q3 2026 (release). The categories are inferred from the message itself (methodology document).
The sample has edges worth knowing. It includes Free, Go, Plus and Pro accounts held by adults aged 18 and over. It excludes Enterprise, Codex, deleted accounts and anyone who opted out of training (methodology document). My reading is that this makes it a view of individuals on personal accounts, some of whom use them for work, and that it tells us little about the managed AI tools a large company gives its own teams.
The shift from 56.0% to 51.1%
OpenAI reports that the ten most common activities fell from 56.0% to 51.1% of work messages. That figure is averaged across the 74 countries tracked every quarter (release).
If the top ten take a smaller share, the remaining activities take a larger one. The number describes mix, so it cannot tell you whether people use ChatGPT more overall. What it shows is work reaching ChatGPT becoming less concentrated over the period.
Why does mix matter more than a bigger total? My interpretation is that when use spreads, the useful question for a brand changes. It stops being “what is the one thing AI is for” and becomes “which of our own jobs sits in the wider mix”. In the CPG teams I work with, plenty of the week goes on narrow jobs such as checking a spec sheet against a live listing, or rewriting a title for a new pack size.
I am not claiming those jobs appear in OpenAI’s data. My point is that a broader mix leaves room for narrow jobs, and narrow jobs are the ones a brand can test properly.
Three jobs worth piloting (Lmo7 interpretation)
These three are my picks, not findings from OpenAI’s study, which did not test any CPG workflow. Each is repeatable, has a clear before and after, has a human who can check the output, and connects to how products show up in AI shopping.
Product data quality checks
The job is making sure product information is correct and consistent before it reaches a listing or a feed. Title against attributes. Pack size and count against the spec sheet. Claims that read one way on Amazon and another on your own site. Today a person usually does this by eye, and often only after something has gone wrong.
The AI-assisted version gives a model the spec sheet as the source of truth and a record to check, then asks it to list every mismatch. A human reviews the exceptions and decides what changes.
This suits a pilot because the answer key already exists. The spec sheet says what is right, so you can count the errors found and the errors missed against a manual audit of the same SKUs. It also feeds the visibility work. My working assumption is that an assistant describes a product from the data it can find, so a clean data set is where the pilot meets the shopping side.
Listing preparation
The job is turning a spec sheet and a brief into listing copy for a new SKU, a new pack or a new marketplace: title, bullets, A+ Content outline and image alt text.
OpenAI’s classification of “doing” covers messages that ask ChatGPT to perform a task, which includes drafting text (methodology document). It does not tell us how well AI-drafted copy performs on a listing, which is the question your pilot answers for your own range.
A human owns the claims. In food and supplements a wrong claim is a compliance problem, so the reviewer checks every claim against approved copy. The pilot measures the time from spec sheet to approved listing, and the number of edits and errors at review.
Reporting and analysis
The job is the weekly or monthly pack. Pull the sales and advertising exports, work out what moved, write the commentary.
AI can draft the commentary and flag unusual movements. The risk is that a model misreads a table or offers a confident reason for a movement that has none. So the control is mechanical. Every figure in the draft must tie back to a cell in the source export, and the analyst signs off the explanation. You measure the hours to produce the pack, the figures that fail the tie-back check, the recipient’s rating of the commentary and how many flagged movements turned out to be real.
How to measure a pilot
Four things, and the order matters: baseline, time saved, output quality and error rate. The table shows what I would record for each job.
| Job | Baseline to record before the pilot | Quality check | Error check |
|---|---|---|---|
| Product data quality checks | Hours to audit a fixed sample of SKUs by hand, and the errors found | Reviewer agreement with the flagged exceptions | Mismatches missed and false alarms, against the manual audit |
| Listing preparation | Hours from spec sheet to approved listing | Reviewer score against a listing checklist | Wrong specs or unapproved claims reaching review |
| Reporting and analysis | Hours to produce the pack | Recipient rating of the commentary | Figures that do not tie back to the source export |
Start with the baseline. Run the manual job once and time it with the person who normally does it. Use the same sample for the pilot, whether that is the same SKUs or the same reporting period, so the comparison is fair.
Time saved has to include review. A draft that appears instantly and takes an hour to correct has saved nothing. Count the full loop from request to approved output.
Output quality needs a checklist written before the pilot starts. Error rate is the count of errors per item reviewed. Where an error is costly, as with product claims, weight this measure above speed.
Agree the decision rule in advance too. What result earns the pilot a place in the weekly routine, and what result ends it? My recommendation would be four weeks, which is long enough to see several cycles of the job. Use the result to decide for your own team, and repeat the pilot before you roll it out to another category.
Where AI shopping assistants fit (Lmo7 interpretation)
This is the Lmo7 view, and it is interpretation. If AI use at work is broadening across jobs, my sense is that shopper behaviour is broadening in a similar way. A shopper who once typed a product keyword into a search box can now ask Rufus, ChatGPT or Gemini to compare options, or to check whether a product suits a dietary need. Each of those is a different job, and each draws on a different part of your product content.
To be plain about the evidence, the OpenAI release is supporting context for that view and not proof of it. It measures work messages in consumer ChatGPT. It says nothing about shopping behaviour, and nothing about Rufus or Gemini. I am drawing a line from one to the other, and it is a judgement.
It still shapes how Lmo7 positions its work. If the questions shoppers ask are getting more varied, a brand needs to know which of those questions it turns up in and how the assistant describes it when it does. That is what Share of Model measures: mention rate and position in AI answers for a defined set of prompts. The content then has to answer the wider mix of questions in a form the assistants can read.
One example from our own work, with its limits. Trip Drinks, a functional drinks client of Lmo7, ran a 60-day engagement covering LLM visibility tracking, site signals, content upgrades on its direct-to-consumer product pages and citation-source optimisation. Average position across major LLMs moved from 7th to 3rd, and AI referral traffic rose 33%. Both figures come from Lmo7’s own LLM visibility tracking across that 60-day window, as published in our Trip Drinks case study. It shows visibility can be measured and moved, which is why I would start the baseline now.
Why I would run both tracks, in sequence
Internal pilots change how your team works. Visibility work changes what a shopper sees when they ask an assistant about your category. A faster listing process will not get you recommended on its own, so I would fund the pilots alongside the visibility work.
The sequence is where the two meet. The product data pilot can start inside your team straight away, and its output, a corrected data set, feeds the visibility work. The visibility baseline starts alongside it, because tracking needs time to show a trend. The listing preparation pilot then uses what tracking shows about the questions shoppers ask and the sources assistants cite. The reporting pilot comes in once there is tracking data worth reporting on.
Visibility work also runs at two speeds. Content and structure changes can move quickly. Authority, meaning the citations and links that make an assistant willing to use a source, takes longer and depends on third parties as well as your own work.
What the data does not show
The study is useful context with clear limits.
- It counts messages, not people. About 27 million sampled messages is not 27 million users.
- It does not measure productivity gains or job impact. It records what people asked for, not what the result was.
- “Seeking information” in OpenAI’s taxonomy covers purchasable products, but also specific information requests and recipes (methodology document). It is not a shopping share.
- “Doing” means messages that ask ChatGPT to perform a task, including drafting text and writing code. It is not proof of autonomous agents.
- The categories are inferred behavioural classifications, not self-reported.
- Enterprise and Codex accounts are excluded, so business AI use is only partly covered. The data is ChatGPT only, with nothing on Rufus or Gemini.
What I would do next
Choose one pilot and baseline it this month, starting with product data quality checks. Owner: your ecommerce lead, with one analyst to run the manual baseline. Reason: it is the most checkable of the three jobs, and its output feeds the visibility work.
Baseline your AI shopping visibility in your category, in parallel. Owner: your ecommerce lead sets the brief, and Lmo7 runs the tracking. Reason: you need a baseline before anyone can show change, and tracking takes weeks to show a trend. Route: Lmo7’s Challenger tier, where we run the work for you. The Agentic Tracking & Content Optimisation module is £1,000 per month and covers Rufus and Brand Radar tracking, long-form LLM-readable content, FAQ pages and answer-snippet optimisation, with no minimum contract.
Act on what the baseline shows, on both speeds. Owner: your ecommerce lead approves, and Lmo7 delivers under Challenger. Reason: the baseline names the gaps. Content and structure fixes can move quickly, and the baseline also shows which sources the assistants cite, which is where the slower authority work starts. Route: Amazon Content Optimisation is £750 per month for PDP copy, A+ Content, Brand Story and SPN-mapped bullets, plus image alt text and weekly ASIN keyword rank tracking. If technical setup for AI discovery is also needed, such as JSON-LD schema and product feed optimisation, Agentic Ops is £500 per month. For a multi-brand portfolio where stakeholders need alignment first, the Enterprise route is a workshop from £5,000, or a tracking pilot at £5,000 to £10,000, with your own teams executing.
If you want step 2 before committing to anything, I can run a category-level AI visibility check: which assistants mention your brand, and where you sit against competitors. Get in touch through the Lmo7 website and we will set it up.
Sources
- OpenAI, “ChatGPT and the expanding world of work”: https://openai.com/index/chatgpt-expanding-world-of-work
- OpenAI Signals data dictionary and methodology: https://cdn.openai.com/signals/data-dictionary.pdf
- OpenAI Signals data: https://openai.com/signals/data/
- Trip Drinks result: Lmo7 LLM visibility tracking, 60-day engagement window, as published in the Lmo7 Trip Drinks case study (not an OpenAI source)
Stephen Honight is the founder of Lmo7, the AI-native agency helping consumer brands win in AI-powered discovery and agentic commerce across Amazon and D2C.