- Public prompt datasets capture a tiny fraction of usage. WildChat, a well-known dataset of real ChatGPT conversations, corresponds to less than a thousandth of what ChatGPT processed in a single week in July 2025.
- Whoever ends up there has selected themselves. English, the USA, Russia and China are strongly represented, and German accounts for only 1.3%.
- Such data are useful for phrasings and new ideas. Search data show more reliably how big the demand for a topic is.
How big is the slice?
Datasets of “real AI prompts” are sold as an insight into customer behavior. A look at the orders of magnitude shows how small the slice is.
| Source | Size | Period |
|---|---|---|
| WildChat-1M (paper) | 1 million conversations with more than 2.5 million turns between user and ChatGPT | 9 Apr 2023 to 1 May 2024 |
| WildChat-4.8M (dataset) | 4,804,190 conversations, of which 3,199,860 without toxic content are public | 9 Apr 2023 to 31 Jul 2025 |
| ChatGPT according to OpenAI | 18 billion messages per week from 700 million users | July 2025 |
If you generously assume 2.5 messages per conversation, as in the WildChat paper, the whole dataset comes to around 12 million user messages from 28 months. That is less than 0.1% of the messages ChatGPT received in one week in July 2025. Claude, Gemini, Perplexity and the AI features in Google and Bing are not even counted.
Who is in such datasets?
For WildChat, researchers provided a chatbot on Hugging Face. Users got free access to ChatGPT and in return agreed that their conversations would be collected and published for research. So the dataset contains people who know this chatbot, do not use their own subscription and are willing to share their conversations. This group is probably more tech-savvy than average. That is our assessment, we have not measured it.
Languages and countries
The distribution in the WildChat paper shows how strongly individual languages and countries dominate.
| Language | Share of turns | Country (by IP) | Share |
|---|---|---|---|
| English | 52.9% | USA | 21.6% |
| Chinese | 13.4% | Russia | 15.6% |
| Russian | 11.6% | China | 10.0% |
| French | 3.4% | Hong Kong | 4.6% |
| Spanish | 2.7% | United Kingdom | 3.8% |
| German | 1.3% | Germany | 3.6% |
If you sell in German-speaking markets, for example, you will find only a few conversations in your language in this dataset. Whether they match your customers remains open.
Behavior under observation
People often behave differently when they know they are being observed. It would be natural to assume that WildChat users therefore avoid sensitive topics or experiment more. That has not been measured. The study “Trust No Bot” examined WildChat conversations and rather shows the opposite. Many users still share personal data, for example in translations and when editing code.
Everyday questions such as product comparisons, travel planning or looking for a tradesperson are probably rarer in such datasets than in real usage. That too is an assessment, not a count.
Data from browser extensions
Some providers advertise larger volumes of conversations collected via browser extensions and apps. A report by The Register from 29 September 2025 describes doubts as to whether users really understand that their AI conversations are passed on. The providers named reject this. Regardless, this sample is selected too. People who install extensions often use free tools and less often work in tightly secured corporate environments.
What do the AI providers themselves show?
Only the providers themselves have representative data. In September 2025, OpenAI published a study that automatically classifies around 1.1 million randomly sampled conversations, without humans reading the content. The figures show broad categories and no individual prompts.
| Topic | July 2024 | July 2025 |
|---|---|---|
| Practical guidance | approx. 29% | approx. 29% |
| Seeking information | 14% | 24% |
| Writing | 36% | 24% |
| Technical help | 12% | approx. 5% |
The study describes seeking information as a “very close substitute” for web search. That supports a simple idea. The need behind a question stays the same, only the place where it is asked changes.
What are prompt data good for?
- Ideas for questions and angles you have not thought of yet
- Phrasings showing how people talk to AI systems
- Collecting content topics and then checking them
- One source among others
- Strategic decisions on this basis alone
- Statements such as “this is how people use AI”
- Conclusions about competitors' customers
- A substitute for surveying your own target group
Search data as a cross-check
Someone who used to google “fix dripping tap” may now ask ChatGPT. The problem has stayed the same. Search data show demand across a wide range, with seasonality, trend and region. That is why a combination pays off.
- Search volume for the size of demand and how it develops.
- Prompt datasets for phrasings and longer conversation flows.
- Your own customer research such as interviews, support requests and surveys as a check.
- Behavioral data from analytics and Search Console to measure what works in the end.
For monitoring your own AI visibility, this means deriving prompts from the topics for which there is proven demand. A list from someone else's dataset hardly reflects your customers.
Date of the figures
The calculations use the figures from the dataset description (as of 11 August 2025) and OpenAI's figures for July 2025. The 4.8 million count conversations, not individual prompts. The language shares come from the WildChat paper.
Common mistakes
- Presenting a prompt dataset as “user behavior”.
- Transferring English-language prompts one-to-one to another language market.
- Taking prompts for monitoring from someone else's list without checking them against your own keywords.
- Inferring demand from how often a topic appears in the dataset.
Frequently asked questions
Is there search volume for AI prompts?
The AI providers do not publish a reliable volume per prompt. Public datasets show examples, but no representative quantities.
What is WildChat?
A research dataset of conversations that users had via free ChatGPT access on Hugging Face and released for publication. The public version contains around 3.2 million conversations.
How do I choose prompts for AI monitoring?
Derive them from keywords with proven demand and phrase them the way your target group asks. Prompt datasets help with the phrasing.
Are data from browser extensions more representative?
They are often larger, but also selected. There are public doubts about consent, which the providers reject.
Are AI answers replacing Google Search?
Part of information seeking is moving to chatbots. According to the OpenAI study, its share in ChatGPT rose from 14% to 24%. The questions behind it stay similar. That is why search data remain a good yardstick.
Sources
Retrieved on October 15, 2026. If something is out of date or missing, let us know.
- Zhao et al., WildChat: 1M ChatGPT Interaction Logs in the Wild (ICLR 2024): https://arxiv.org/abs/2405.01470 · May 2, 2024
- WildChat paper, full text with language and country distribution: https://arxiv.org/html/2405.01470v1 · May 2, 2024
- Allen AI, WildChat-4.8M dataset: https://huggingface.co/datasets/allenai/WildChat-4.8M · Updated Aug 11, 2025
- Chatterji et al., How People Use ChatGPT (NBER Working Paper 34255): https://www.nber.org/papers/w34255 · Sep 15, 2025
- Mireshghallah et al., Trust No Bot (COLM 2024): https://arxiv.org/abs/2407.11438 · 2024
- The Register, Your AI conversations are a secret new treasure trove for marketers: https://www.theregister.com/2025/09/29/profound_browser_extension_privacy_concern/ · Sep 29, 2025