To practice estimation and market-sizing questions with AI, have a chatbot play the interviewer: it gives you one question, answers your clarifying questions, waits while you structure and calculate out loud, then challenges your assumptions and checks your arithmetic. It is good at generating varied questions and spotting gaps in a structure. It is not reliable at arithmetic or at real-world figures, so make it show every step, check the math with a calculator, and treat anything it presents as a fact as an assumption until you have verified it.
Below: what interviewers listen for, five prompts, a fully worked example with every assumption labelled, a table of the arithmetic slips chatbots make and a two-week plan. For the method itself, start with our guide to estimation interview questions; for full consulting cases, see case interview practice with AI.
What interviewers listen for in an estimation answer
| Step | What a strong answer does | How AI practice helps |
|---|---|---|
| Clarify | Pins down geography, time period, units, and revenue versus volume | The AI answers your clarifying questions the way an interviewer would |
| Structure | Breaks the number into a clear chain of factors, top-down or bottom-up | It points out gaps and double counting |
| Assumptions | States each one out loud, rounded and defensible | It pushes back on the ones that are most uncertain |
| Math | Written down, round numbers, no skipped steps | It checks your working, if it shows its own |
| Sanity check | A second method or a known benchmark, ending in a range | It asks whether the result passes a smell test |
| So what | Says what the number means for the decision | It asks the business follow-up |
These questions turn up in consulting cases, product manager loops (often inside a product sense interview), operations roles and the capacity-planning part of system design rounds. If the format is new to you, read what a case interview is first.
Set the chatbot up as an interviewer, not an answer key
Left to itself, a chatbot will answer its own question in the same message. Each prompt below closes that door.
1. Interviewer mode, one question at a time
Act as an interviewer for a [consulting / product manager / operations] role. Give me one market-sizing or estimation question suited to that role, and nothing else. When I ask clarifying questions, answer briefly as an interviewer would. Do not show a solution. When I give my answer, ask me two follow-up questions about my assumptions, then give feedback on structure, assumptions, math and sanity check, scoring each from 1 to 5.
Why it works: "do not show a solution" gives you the blank page you will face in the room, and the follow-ups imitate the probing that comes after your first number.
2. Challenge my assumptions
Here is my estimate with my assumptions numbered: [paste]. For each assumption, say in one line whether it looks reasonable, too high or too low, and why. Then tell me which assumption is the most uncertain, give a sensible low and high value for it, and show what range that produces for my final answer, step by step.
Why it works: interviewers probe the shakiest number, and seeing the range it produces tells you where to spend your words.
3. Check the arithmetic line by line
Check my arithmetic line by line. For each line, write the calculation, the result in full digits (for example 18,000,000, not 18M) and whether my result matches. Do not skip or combine steps, and say wherever you round. Then recompute the final number a second way, multiplying in a different order, and confirm both results match.
Why it works: full digits and a second route catch most dropped-zero and unit slips, including the chatbot's own.
4. Cross-check top-down with bottom-up
I estimated [quantity] top-down as follows: [my steps and answer]. Without giving me any numbers, describe one bottom-up approach (starting from a single store, vehicle, worker or server) I could use to cross-check it. I will do the math myself. Then tell me whether my two answers are within a factor of two and, if not, which assumption to revisit.
Why it works: reconciling two methods is one of the strongest sanity checks you can show, and here you still do the thinking.
5. A question bank across interview types
Write 12 estimation questions: 4 typical of consulting cases, 4 of product manager interviews and 4 of operations or system design interviews. For each, give the unit the answer should be in and whether a top-down or bottom-up start is more natural. No solutions.
Why it works: mixing types stops you from forcing every question into one memorized template.
A worked example: bike repairs in a city of 2 million
Question: Estimate the annual market for bicycle repairs in a city of 2 million people. Every figure below is an assumption chosen for this example, not a fact about any real city.
First attempt (spoken): "Hmm. Lots of people cycle, so maybe half the city has a bike, that's a million bikes, and each one gets fixed a few times a year at maybe a hundred dollars, so around 300 million dollars? It feels like a big market."
What the AI flagged (Prompt 2): no clarification of scope; "half the city" counts bikes owned, not bikes ridden; "a few times a year" is the most uncertain number and drives the answer; no cross-check and no range.
Improved answer (spoken): "I'll size it as one year of paid repairs and servicing at shops, not do-it-yourself. Assume one bike for every four people: 500,000 bikes. Assume 40 percent are ridden regularly: 200,000 active bikes. Assume 1.5 paid repairs or services a year each: 300,000 jobs. At an assumed 60 dollars a job, that's 18 million dollars a year. Cross-check bottom-up: say 60 shops doing 20 repair jobs a day for 300 days, which is 360,000 jobs, within 20 percent of my first figure. If repairs per active bike range from 1 to 2, the market is 12 to 24 million dollars, so I'd say about 18 million."
| Factor | Assumption | Result |
|---|---|---|
| People | 2,000,000 (given) | 2,000,000 |
| Bikes owned | 1 per 4 people | 500,000 |
| Bikes ridden regularly | 40% of bikes owned | 200,000 |
| Paid repairs a year | 1.5 per active bike | 300,000 jobs |
| Average price | 60 dollars a job | 18,000,000 dollars |
| Bottom-up check | 60 shops × 20 jobs a day × 300 days | 360,000 jobs |
| Range | 1 to 2 repairs per active bike | 12 to 24 million dollars |
Notice what the improved version does: it defines scope first, says "assume" before every number, rounds, cross-checks and ends with a range and a central figure. The first attempt multiplied correctly; its problem was unexamined assumptions.
Catch the chatbot's arithmetic slips
Chatbots generate text one piece at a time, and they can and do make arithmetic mistakes, especially across several steps, with large numbers and with percentages. Treat the AI's math the way you would treat a teammate's: check it before you learn from it.
| Slip | What it looks like | How to catch it |
|---|---|---|
| Dropped or extra zero | "2 million households × 15% = 30,000" (it is 300,000) | Ask for full digits on every line |
| Percent of a percent | 40% of 20% treated as 60% or 20% (it is 8%) | Write percentages as decimals: 0.4 × 0.2 = 0.08 |
| Unit drift | A daily figure multiplied by 365 and then again by 52 | Ask it to label the unit of every line |
| People versus households | A per-person rate applied to a household count | Name what each count is counting |
| Rounding drift | Rounding at every step until the answer moves a lot | Round inputs once and say where |
| Assumptions dressed as facts | "The city has 4,000 bike shops" | Ask "is that an assumption or a sourced figure?" and treat it as an assumption |
If you want a few real anchor figures (your country's population, number of households, a typical working year), look them up once on an official statistics site, round them and keep them on your one-page prep sheet. If the arithmetic itself is what slows you down, drill it separately with our guide to practicing mental math with AI.
A two-week practice plan
| Days | Focus | Daily reps |
|---|---|---|
| 1–3 | Top-down from population: how many of something are bought or used a year | 2 questions, 6 to 8 minutes each, out loud |
| 4–6 | Bottom-up from capacity: one store, one vehicle, one server | 2 questions, each with a cross-check (Prompt 4) |
| 7–9 | Revenue sizing with price and frequency | 2 questions, then Prompt 2 on each |
| 10–12 | Mixed and timed, with the AI interviewer's follow-ups | 3 questions in interviewer mode (Prompt 1) |
| 13–14 | A full practice session: behavioral questions plus one estimate | 1 recorded session, then review |
Say every answer out loud before you type it. Estimation is spoken reasoning, and a typed answer lets you hide the pauses and backtracking an interviewer would hear. For the spoken side of the interview, run an AI mock interview built from your résumé and the job description; it reads questions aloud and scores each answer you speak. For how estimation practice fits into the rest of your preparation, see our guide on how to use AI for interview prep.
Where CoPilot Interview fits
CoPilot Interview is an AI interview assistant with a practice-first free plan. Use a general chatbot for estimation drills as described above, and use CoPilot Interview's AI mock interview to rehearse the behavioral and motivation questions that come with consulting and product interviews, out loud and scored out of 10. The browser free practice demo adds 145 typed questions across 10 tracks, free and with no signup. The free plan gives 15 minutes a day of live help in sessions of up to 15 minutes, with AI mock interviews included. Paid plans are Standard at $24.99 a month and Pro at $39.99 a month. Estimates in a real interview, a case round or an online assessment are yours to reason through: follow your employer's rules, and prepare for assessments that forbid outside help rather than looking for help during them.
Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly.
Frequently asked questions
Can ChatGPT check my market-sizing answer?
It can check your structure and arithmetic if you make it show every step in full digits, and it can challenge your assumptions. It cannot tell you whether a base figure is true, so verify anything you plan to rely on.
How accurate does a market-sizing estimate need to be?
Usually within the right order of magnitude. Interviewers care more about a clear structure, stated assumptions, clean math, a sanity check and what the number means than about the exact figure.
Should I memorize population figures for estimation questions?
A few round anchors help, such as your country's population and number of households. Look them up from an official source, round them, and always say them as approximations.
Is top-down or bottom-up better for market sizing?
Neither in general. Pick the one the question suits, and if time allows, do a quick version of the other to cross-check your answer.
How long should an estimation answer take?
In practice drills, five to eight minutes including clarifying questions is a common target. In the interview, follow the interviewer's pace and keep talking through each step.
Try a practice question
The browser demo has 145 real interview questions across 10 tracks. Type your answer and get a score out of 10, strengths, areas to improve and a follow-up question. Free, no signup.
Try a practice question