Estimate the number of ChatGPT users worldwide

[INTERVIEWER] Estimate the number of ChatGPT users worldwide. Estimate the number of ChatGPT users worldwide. Before you say a single number, understand this: nobody in that room is checking your arithmetic to the decimal place. They genuinely don't care if you land on 280 million or 310. They're checking whether you build a structure, state your assumptions out loud, and sanity-check the answer instead of just blurting a figure.

Top-down, one clear assumption per line, then a reality check at the end. That's the whole game. The person who blurts half a billion with no structure loses to the person who builds a defensible 200 million. Estimation questions aren't about the answer. They show how you think without data, which is most of a PM's real life. Can you decompose a scary number into pieces you can reason about?

Do you flag your softest assumption instead of hiding it? Do you catch yourself when you're an order of magnitude off? Get this right and you've shown the interviewer you can size a market, a workload, or an opportunity on the spot. Let me give you the structure. Step one, which many candidates skip: "users" is undefined. Registered accounts, monthly active, weekly active, daily active?

Those numbers differ by multiples, so pin it before you compute. I'll estimate monthly active users, meaning humans who use ChatGPT at least once in a month, and I'll say where daily active would land at the very end. I'll scope it to consumer web and the app, meaning real people. If they want API developers, I'll treat those separately because they're a different population entirely.

Forty-five seconds on this saves you from computing the wrong thing beautifully. Step two, pick your approach and justify it. Bottom-up, summing users country by country, is fragile here because of too many unknowns stacked on each other. Top-down, starting from the global online population and funnelling down through adoption, is cleaner, and every step is a single defensible assumption.

So I state the approach out loud before I touch a number. That sentence alone tells the interviewer you've done this before. Now the funnel. I narrate each line as an assumption, not a fact. Start with world population, about eight billion. Internet users, roughly two thirds of that are online, so call it five and a half billion. Now the reachable slice: literate in a supported language, not behind a hard block.

China is largely excluded, and you've got infrastructure and language gaps elsewhere, so I'll take it down to about three and a half billion addressable. Awareness of ChatGPT: it's a genuine household name in the online world now, so say half of those addressable people are aware, about 1.75 billion. Have tried it at least once: say forty percent of the aware, roughly 700 million ever-tried.

And still monthly active, the retention from ever-tried: say another forty percent, which lands at about 280 million monthly active users. Notice I kept every number round, and I flagged each percentage as a judgement, not a measurement. False precision is a tell that you don't understand what you're doing. Step four, and this is the step that actually scores.

Does 280 million pass a smell test? Publicly, OpenAI reported 300 million weekly active users in late 2024, climbing to 400 million by early 2025. Monthly active users must logically exceed weekly active users. So my 280 million monthly figure is actually a massive undercount. That's a red flag. I need to revise my assumptions upward. If awareness is closer to seventy percent, that gives 2.45 billion.

If trial is fifty percent, that's 1.2 billion. And if monthly retention is fifty percent, I land around 600 million monthly active users. That sits comfortably above the 400 million weekly active figure, which makes logical sense. Catching your own error out loud and correcting it is one of the strongest things you can do in an estimation. Now flex the structure using that corrected 600 million figure to get the whole family of numbers.

Apply a stickiness ratio. A habit product might see daily-over-monthly around 0.2 to 0.4, so say 0.25, which gives roughly 150 million daily active. And paid conversion, maybe five percent of monthly actives, lands you around 30 million paying subscribers. Being able to slide between monthly, daily, and paid off a single structure, without rebuilding anything, is a strong close and it shows the structure was real.

Next, name what would swing the answer most. The biggest levers here are whether a major blocked market opens up, mobile-app growth across the developing world, and the retention assumption, which is the softest number in the whole chain. Flagging that sensitivity honestly, saying here's the number I'd least trust, is exactly the maturity they're grading. Let me run the full corrected chain clean, top to bottom, so you can hear the rhythm.

Eight billion people, to five and a half billion online, to three and a half billion addressable once you filter for language and access, to 2.45 billion aware at seventy percent, to 1.2 billion ever-tried at fifty percent, to 600 million monthly active at fifty percent retention of triers. Sanity check: 600 million monthly sits comfortably above the publicly reported 400 million weekly active figure, so the logic holds.

Neighbours: at a daily-over-monthly of 0.25, about 150 million daily active; at five percent paid, about 30 million subscribers. Softest assumption: the fifty percent retention of ever-tried, which on its own can swing the answer by over a hundred million either way, so that's the number I'd flag to firm up with real data before I trusted the estimate for anything that mattered.

A sharp interviewer will now poke at your weakest number to see if you flinch. They might say your seventy percent awareness feels high and ask you to defend it. This is a gift, because a good estimator has a range, not a point. So I'd say: fair, awareness could plausibly be sixty percent rather than seventy. Watch what that does.

It drops aware to 2.1 billion. If trial stays at fifty percent, that's 1.05 billion. If retention drops to forty percent, monthly actives land around 420 million. So my honest range is roughly 420 to 600 million monthly, and the true answer probably sits in there. Giving a range on demand, instead of clutching your single number, is exactly the composure they're testing.

The second push you should expect is to estimate it bottom-up and see if the two agree. That's the strongest possible check. Quick bottom-up: the US has maybe 250 million adults online, say half have tried ChatGPT and fifty percent of those stay monthly, that's around 62 million US monthly actives. If the US is roughly twelve percent of global usage, you scale up to about 500 million.

It lands in the same zone as the top-down. When two methods built on totally different assumptions agree, that agreement is worth more than either number alone. Here's what makes them lean in. First, you pinned the metric, monthly versus daily versus registered, before touching a number, because the entire answer hinges on it. Second, you stated one clear assumption per line, kept the numbers round, and used no false precision.

Third, you ran a real sanity check against a public figure you actually know, caught an error, corrected it, then flexed the same structure into daily and paid. That combination of structure, sanity check, and flex is a clean pass. Avoid three traps. First, blurting a single number with no structure and no stated assumptions, which gives the interviewer nothing to grade and no way to follow your thinking.

Second, skipping the sanity check, so an answer that is ten times off just sails past uncaught. Third, getting lost in precision, like 8.1 billion people or 67.3 percent online, which signals you've missed the entire point. So the shape: pin the metric, choose top-down and say why, funnel down with one stated assumption per line and round numbers, sanity-check against something you know, then flex into the neighbouring metrics and flag your softest assumption.

Carry this into the room: pin the metric, funnel top-down with one stated assumption per line, then sanity-check against a number you know. The structure is the answer, not the digits.