From demographics on who uses ChatGPT most and total weekly active user numbers to details such as splits between topics, work vs non-work and user intents there’s so much to take in but I’ve tried to immediately pull out the most important 10 insights for the SEO industry (and those of you calling it GEO or AEO too).

OpenAI reference in the abstract of their new paper that ‘Despite the rapid adoption of LLM chatbots, little is known about how they are used.’ So they’ve set about finding out, being uniquely positioned with the data they require and helped out by no less than Harvard university & Duke University academics.

Do read the full, 63-page, paper when you have time, https://cdn.openai.com/pdf/a253471f-8260-40c6-a2cc-aa93fe9f142e/economic-research-chatgpt-usage-paper.pdf

For now though I have pulled out 10 of the most interesting stats and findings, especially for those of us working in SEO, including AI SEO (and GEO, AEO, etc.).

Notes on the Report’s Methodology

The paper focuses on a random sample of messages across different account levels (free, plus, pro), belonging to over-18s, with most data comparing June 2024 and June 2025.

The paper dives into the methodology more but, for speed and privacy, they have used an LLM to classify the messages in a number of ways, mentioned in the main insights below, such as work, vs non-work, topic, type of interaction etc.

1. ChatGPT Weekly Users – 10% of adults Worldwide.

As of July 2025, each week ChatGPT sees 18 billion messages sent by 700 million unique users, which as OpenAI point out is equal to around 10% of the global adult population (there are presumably significantly more unique users when looked at monthly).

So it seems that within each signup period-cohort some users do drop off completely or almost completely over time, but, as you might expect, the earliest adopters, who actually stuck with ChatGPT, are the true power users now.

2. Work vs Non-Work Use – Non-Work outpacing Work

Non-Work use is now growing faster than work use and while at 53% of messages in June 2024 by June 2025 non-work accounted for 70% of usage.

Work related messages still grew by +236% but Non-Work messages grew by a massive +703% to 1.9 Billion a day in the year to June 2025.

From what I have seen B2B companies have been slightly faster to move to optimise for LLM Chat Bots, with many B2C brands waiting to see how the use of ChatGPT et al progressed. Well it seems if you haven’t heard the starter’s gun yet, then this could be it for you.

(See the paper, section 5.1 for the prompts used to identify work vs non-work)

3. Topics of conversation – 3 conversation topics = >3/4s of messages.

OpenAI’s classification of conversation topics includes 7 top level topics, though these are broken down into 24 sub-categories which give even more interesting insights*.

The above chart taken for reference purposes from Chatterji et al (2025) How People Use ChatGPT.

The 7 are shown below, with the top 3 dominating with a combined percentage of 77% of messages

Of the above Practical Guidance has maintained a similar share year on year (YoY) but Writing has fallen 12 p.p (percentage points) or 33% YoY to 28.1%.

I have dived into Writing in more detail in point 5 below, but, and this is speculation, is this a sign that some people are realising that the quality of AI written content and correspondence just isn’t good enough and they are getting caught out? Or just that this was a core use for early adopters but the latest wave of new users are using it differently, after all writing is more likely to be for work purposes.

Seeking Information meanwhile grew 10 p.p or c.70% to 21.3% again this type of conversation, people using ChatGPT to get information and inform themselves, is hugely significant to us as SEOs and is covered in more depth below.

Key Insights given on a Sub-Category Level

Other interesting, arguably surprising insights, include that only a small percentage of messages are for Computer Coding, 4.2% . This is a ‘skill’ that ChatGPT is known for, and I’ve heard people claim things along the lines of ‘people aren’t using ChatGPT for search, they’re using it to make cartoon versions of their cats and to try to build their ‘unique’ idea for an app’.

Speaking of which 4.2% of messages are for Creating Images, by far the biggest share within the total of 6% for Multimedia conversations, but small compared to the 21.3% Seeking Information.

Practical Guidance conversations still a threat to Traditional Search

Practical Guidance is an interesting area for us SEOs as well, this includes subcategories like Tutoring & Teaching at 10% of all messages and How to Advice 8.5% and Health, Fitness Beauty or Self Care 5.7%.

These are different to simply seeking information, but do replace search in some cases, often the kind of searches where Featured Snippets used to show and if we had written great E-E-A-T content we could expect to get a lot of traffic and good engagement as people looked not for a simple answer but to read around a topic.

What ChatGPT does better here though is make this kind of guidance personalised and that is something long form web content can’t compete with, if AI does it well.

Lastly, while not especially relevant to SEO, Relationships & Personal Reflection account for 1.9% of messages and it is worth remembering that under 18s aren’t included.

With the concerns many of us have of people, especially young people, reaching out to ChatGPT for this kind of advice, and potentially getting bad advice, it may feel relieving to see this percentage is so low, but this could still account for around 34.2 Million messages per week. This highlights how important it is that ChatGPT, and other AI Chatbots put measures in place to handle these kind of queries in the right way.

*(See section 5.2 for the full classification on conversation topics)

4. ‘Seeking Information’ conversations (in other words ‘search’) grew 70%+ in 12 months.  

The most interesting top-level topic for SEOs is probably Seeking Information, OpenAI themselves compare this to Web Search and break it down in to:

1. Specific Info 18.3% (of all conversations)
2. Purchasable Products 2.1%
3. Cooking & Recipes 0.9%

It’s tempting to jump straight to Purchasable Products with that number seeming unbelievably low. However, let’s start with Specific Info.

Specific Info then is information with one definitive answer but also broader research around topics and questions open to some interpretation. OpenAI gives example prompts for each sub-category of conversations in their appendix and for Specific Info they are:

– “What is regenerative agriculture?”

– “Whats the name of the song that has the lyrics I was born to run?”

– “Tell me about Marie Curie and her main contributions to science.”

– “What conflicts are happening in the Middle East right now?”

– “Quelles ´equipes sont en finale de la ligue des champions ce mois-ci?”

– “Tell me about recent breakthroughs in cancer research.”

While some of these wouldn’t be phrased this way when using traditional search engines like Google, less conversational terms like “marie curie contributions to science” would have been used with a similar research aim in mind. It is easy to see where these kind of searches are replacing, not just complementing, Google Search.

What the research sadly doesn’t tell us is for these kind of terms is how often does ChatGPT decide it needs to add to its knowledge base and search? How often it fetches pages? How often are sources cited? And How often do users click on them?

Purchasing Products still only 2.1%

The number that will have those who said that ‘ChatGPT doesn’t drive sales and that users still go to Google when they want to buy’ rubbing their hands at the thought of being proven right: only 2.1% for Purchasing Products, albeit that’s 37.8 million conversations per week.

I am surprised it is that low, but not surprised it is low, ChatGPT is arguably better than some AI ChatBots at finding and comparing products, but still has massive blind spots. ChatGPT relies on access to feeds mostly, such as those it gets direct from Shopify, but it’s still missing most retailers and products for now.

There’s little doubt in my mind the potential is there for ChatGPT and its competitors to be the ultimate shopping comparison engines, but they aren’t their yet.

Where people may be using ChatGPT more though is when researching higher up the funnel, for solutions to problems, the types of product and different brands that are available not the kind of terms categorised here as Purchasing Products, which focuses mainly on a specific product or brand or on a type of product and may include questions about price, the examples from OpenAI being:

– “iPhone 15.”

– “What’s the best streaming service?”

– “How much are Nikes?”

– “Cuánto cuesta un Google Pixel?”

– “Recommend a good laptop under $1000.”  

5. Writing – 40% of work-related messages

While Writing is just 28.1% of all messages, within those messages that are work related, 40% are related to Writing. That seems to make a lot of sense and there’s nothing too surprising there though it has fallen YoY from around 43%.

That’s still growth in total messages Year on Year of course but it again shows that maybe using ChatGPT for Writing is less of a focus for newer users, and my opinion is that as there is more AI generated text out there it is less valuable and easier for people to spot and users may be starting to realise this.

2/3rds of Writing conversations are for improvements to users’ text.

Within writing though we can see it is split across 5 sub-categories of conversation:

1. Edit or Critique Provided Text – 10.6% (of all conversations)
2. Personal Writing or Communication 8.0%
3. Translation 4.5%
4. Argument or Summary Generation 3.6%
5. Write Fiction 1.4%

Three of these are focused on taking user’s text though and improving (Edit or Critique Provided Text), augmenting (Argument or Summary Generation) or translating (Translation) it.

These account for 66.5% of all Writing messages. Personal Writing or Communication, in other words generating new content or communications, is a relatively small share, again perhaps users are realising that ChatGPT and other AI Assistants are not great at just being left to generate text, and that their real value is as a tool to help writers be more efficient and improve the quality of their work.  

6. User intent: Asking vs Doing vs Expressing*

While the classification for topics is based on one OpenAI used internally already, they also came up with a new taxonomy based on the type of output a user is seeking, or their user intent.

This is different to the standard search intent classifications we may be used to, though maybe Asking isn’t that far from informational and Doing from Transactional.

49% of messages were Asking, i.e. when a user “is seeking information or clarification to inform a decision, corresponding to problem-solving models of knowledge work”.

40% of messages were Doing, i.e. the conversation is task based and the user wants to “produce some output or perform a particular task”.

Only 11% of messages were Expressing i.e. when the user wants to express “views or feelings” to the model but isn’t expecting any specific information or action. Should this be ‘only’ though or is this another one where actually that number is surprisingly high?

When comparing Intent classifications to Topic classification, OpenAI found “Asking queries are more likely to be Practical Guidance and Seeking Information” so Askingis arguably a better direct comparison to Web Search than ‘Seeking Information’ alone.

It is true that many Asking intentconversations will result in information that isn’t directly comparable to traditional search results, the answers may well replace information that users would get from visiting a page and reading an article to get answers and information, as well as wider topic research.

For work messages Doingjumps up to 56% of messages (ahead of 35% for Asking), and 35% of the total are Doing messages related to Writing.

While not directly relatable to Writing being done for web-content or copy, the sheer volume, 630 Million requests for writing to be updated or created each week, is difficult to ignore.

*(See OpenAI’s definitions, and examples, of Asking, Doing & Expressing in section 5.3. of their report)

7. User Gender Gap has narrowed to almost 50/50

When ChatGPT first launched they report that 80% of active users’ first names were masculine (excluding those that couldn’t be identified as masculine or feminine). As of June 2025 however that was only 48%, meaning ChatGPT users are slightly more likely to have feminine first names.

Many assume that early-adopters of technology tend to be male, and sometimes decisions for marketing make this a self-fulfilling prophecy. However, some research (source) points to this not being the case, but only where women have access.  The low cost of ChatGPT and the fact it can be used on a mobile app  means that anyone with a smartphone can access it now, which may not have been the case when it first launched.

Those assuming that optimising for LLMs and specifically Chatbots is only useful for targeting tech savvy, slightly nerdy, men working as developers or running tech start-up businesses, again, need to see these figures and have a rethink.

What is worth mentioning though is that the report did find that those with feminine names are more likely to send messages around ‘Writing and Practical Guidance’. Whereas for masculine first names they over indexed for ‘Technical Help, Seeking Out Information, and Multimedia’.

8. Ages of users – Nearly 50% of adult users under 26.

As mentioned earlier OpenAI excluded under 18s so data is only available for adults. But of adult users 46% were under the age of 26 years old (18-25).

Whether that seems high or low will depend on your perception of ChatGPT users, you may not be surprised that older generations are less likely to be using ChatGPT, you may even be surprised that the percentage for 18-25s isn’t higher.

There’s still significant numbers for over 25s but OpenAI don’t break that down for all age cohorts annoyingly.

While there is data missing from the report to understand, for over 25s, just what age groups are using ChatGPT most, maybe some demographics you previously assumed were all still safely on Google and even ignoring AI Mode and AIOs, you need to think about as being at least occasional ChatGPT users now.

However, with around 17.4% of adults being aged under 26 (source: UN), this all means that an estimated 27% of 18-25s year olds are using ChatGPT each week worldwide (nearly x3 the average adult rate of 10%).

That doesn’t mean they don’t use other search platforms but if that’s your target market you can no longer ignore AI Agents like ChatGPT, just because it’s harder to track and report on. Yes most of what you do is not that different to optimising for Google and Bing, but those few vital things that are, they are really important.

9. Highest Growth Rates in Low-Medium Income Countries

OpenAI found that in the year to May 2025 usage grew disproportionately in low- and middle-income countries ($10,000–40,000 GDP-per-capita).

Usage grew for countries with all income levels (based on splitting countries with populations of 1M+, where ChatGPT is not blocked, into deciles).

The higher growth in low-medium income countries may indicate users, including businesses, using ChatGPT, as a cheap, flexible tool, for a range of business, professional and educational purposes.  

This may allow them to more effectively compete with workers and businesses in other countries who have access to more expensive and unaffordable tools and education. This points to what many have already written about that LLM chatbots can be a great leveller within society.

10. How is ChatGPT used for Work – 58% of Work messages related to just 2 work activities.

For the study, OpenAI mapped messages to work activities using the Occupational Information Network (O*NET) job characteristics and Generalised Work Activities (GWAs); this allowed them to map messages to specific types of work activity.

58% of messages were related to either ‘obtaining, documenting, and interpreting information’  or ‘making decisions, giving advice, solving problems, and thinking creatively’. This pattern is similar across different professions, including STEM, management and business, administration and sales roles.

Interestingly, though not unexpectedly, those working in Computer-Related roles were likely to send a significantly higher proportion of work related messages, 57% compared to 50% for management and business; 48% for engineering and science; 44% for other professional occupations and 40% for non-professional occupations. This gives a basic but interesting insight for those selling B2B to know which professions are best targeted through ChatGPT currently.

For those marketing B2B products and services it is perhaps even more useful to understand if and where ChatGPT and other AI Assistants fit into different types of work activity (GWAs). Making Decisions and Solving Problems, while not limited to purchase decisions does include these, and, compared to 7 other GWAs, ranks first for the most common requests for 10 out of 16 occupations (& 2nd for 4 more).

Making Decisions and Solving Problems ranks first for the occupations:

• Management
• Business
• Engineering
• Science
• Social Service
• Arts / Design / Media
• Sales
• Administrative
• Transportation
• Military

Meaning for B2B products and services selling to these occupations, ensuring that you have optimised so that ChatGPT is likely to cite you or recommend your offerings may be especially valuable.

Above the full Occupation and GWA table from page 35 of Chatterji et al (2025) ‘How People Use ChatGPT‘.

Thank you to the academics and researchers of OpenAI, Harvard University and Duke University for creating such an in-depth and useful report. Full Reference:

Chatterji, A., Cunningham, T., Deming, D., Hitzig, Z., Ong, C., Shan, C. and Wadman, K. (2025) How People Use ChatGPT. OpenAI. Available at: https://cdn.openai.com/pdf/a253471f-8260-40c6-a2cc-aa93fe9f142e/economic-research-chatgpt-usage-paper.pdf (Accessed: September 16th 2025).