Fun with Data – AI Usage
Unless you are very adept at using smartphone OS and browser applications, every time you use your phone or open a browser, lots of information gets collected. Most folks don’t know or care that information is being harvested about where they go, what they do, and what they do it on, and that gives those who know how (or pay to) collect that data, the ability to analyze all sorts of trends. In most cases, the data is relatively benign and (hopefully) anonymous, so it can be used to come up with thousands of metrics that are primarily used for marketing.
An example might be browser statistics, essentially the number of times a particular browser type is opened in a given period, or how often a website is visited and how long users remain on the site. These are routine statistics that are readily available, so in the fierce battle for Ai supremacy, such data is as important or more important than the benchmarks that are used to compare Ai model performance. Depending on which metric is used, AI models can be ranked according to the number of times they are accessed and can even be segmented by country, phone OS (Android, iOS, etc.) and a variety of other categories. However, when it comes down to who is using what AI and how often, such user data is a good start toward trying to calculate a model’s potential for profitability.
The big problem is the fact that there are typically a number of AI model pricing tiers , with the far most popular being ‘free’, and the data that is collected by your phone’s OS or by your browser does not know if you are paying for the model’s service or if it is free. This means the value of such user data is more general than might be hoped for, but we take what we can if it helps to put together the puzzle.
An example might be browser statistics, essentially the number of times a particular browser type is opened in a given period, or how often a website is visited and how long users remain on the site. These are routine statistics that are readily available, so in the fierce battle for Ai supremacy, such data is as important or more important than the benchmarks that are used to compare Ai model performance. Depending on which metric is used, AI models can be ranked according to the number of times they are accessed and can even be segmented by country, phone OS (Android, iOS, etc.) and a variety of other categories. However, when it comes down to who is using what AI and how often, such user data is a good start toward trying to calculate a model’s potential for profitability.
The big problem is the fact that there are typically a number of AI model pricing tiers , with the far most popular being ‘free’, and the data that is collected by your phone’s OS or by your browser does not know if you are paying for the model’s service or if it is free. This means the value of such user data is more general than might be hoped for, but we take what we can if it helps to put together the puzzle.
In the Table above we show US daily, weekly, and monthly engagements with each of the 5 most popular AI models, with the obvious leader being ChatGPT (pvt) at almost 6 times the closest competitor on a daily basis. That ratio drops considerably when calculated on a weekly basis, and further on a monthly basis but is still far and away the leader. GROK (pvt) engagements are quite close to Gemini’s (GOOG) on a daily basis, however weekly and monthly ratios are lower. DeepSeek (pvt), the newest model, is certainly a contender in the US even though it has been bad-mouthed by the US government as a tool of the Communist Party.
We note that ‘stickiness’, as it is used here is a ratio between daily users and either weekly (Ud/Uw) or monthly users (Ud/Um), with the higher number indicating that users are more likely to return to the model on a regular basis. Again, ChatGPT is the obvious ‘stickiness’ leadeer while surprisingly Gemini the weakest. We would have expected Gemini, with its ability to access some of the Google search index, to have been stickier.
We note that ‘stickiness’, as it is used here is a ratio between daily users and either weekly (Ud/Uw) or monthly users (Ud/Um), with the higher number indicating that users are more likely to return to the model on a regular basis. Again, ChatGPT is the obvious ‘stickiness’ leadeer while surprisingly Gemini the weakest. We would have expected Gemini, with its ability to access some of the Google search index, to have been stickier.
At the global level, while ChatGPT is the leader by far, DeepSeek takes the number 2 spot at more than 4x GROK’s position, and makes a stab at ChatGPT’s weekly stickiness ratio, likely due to its origins in China. We take things a bit further in the table below where we track the global user base for each model against the percentage of population represented by the US. As the population of the US currently represents 4.22% of the global population, it can be seen that while the absolute user base for ChatGPT is considerably larger than any of the others, it is a global phenomenon, as is DeepSeek, while Gemini, GROK, and Claude (pvt) are much more US centric, especially Gemini, with US users representing an average of 44% of its total user base.
One unusual datapoint we came across was the average session time spent on each model. While there was a relatively small difference between the leaders, ChatGPT (3:39) and DeepSeek (3:18) and GROK (2:47) and Claude (2:28), again the surprise came from Gemini with a meager 0:17 average user time. We expect this short stay is a function of Gemini’s use as a ‘summarizer’ where search results are boiled down to a paragraph or two, although given the model’s large US user base, it could also be due to the very short attention span of US users (joke!). We will dig further.
As noted above, the data that shows what percentage of each model’s users are ‘paid’ users, is not publicly available, as it would open each model to a level of transparency that would likely make management uncomfortable, especially as growth slows, so right now we have to rely on more typical ratios for subscription models, especially as AI models are a new category. Worst case would be a 2% to 5% ‘paid’ ratio, and best case between 10% and 25%, the average for ‘freemium’ subscriptions that offer a free trial period that converts to paid if the user finds value. The calculations get very complex even if we know the exact paid ratio for each model as the models are priced based on the number of tokens processed rather than a monthly fee. How each user breaks down as to token use is another ‘internal’ metric that will take time to discover. Until then estimates are just supposition, making future growth and profitability unusually difficult to pinpoint, at least for now, although we expect AI investors might disagree..
As noted above, the data that shows what percentage of each model’s users are ‘paid’ users, is not publicly available, as it would open each model to a level of transparency that would likely make management uncomfortable, especially as growth slows, so right now we have to rely on more typical ratios for subscription models, especially as AI models are a new category. Worst case would be a 2% to 5% ‘paid’ ratio, and best case between 10% and 25%, the average for ‘freemium’ subscriptions that offer a free trial period that converts to paid if the user finds value. The calculations get very complex even if we know the exact paid ratio for each model as the models are priced based on the number of tokens processed rather than a monthly fee. How each user breaks down as to token use is another ‘internal’ metric that will take time to discover. Until then estimates are just supposition, making future growth and profitability unusually difficult to pinpoint, at least for now, although we expect AI investors might disagree..
RSS Feed