By Ajitesh

Estimation Interview Questions: How to Prep for Market Sizing, Guesstimates and Fermi Rounds

Estimation Interview Questions: How to Prep for Market Sizing, Guesstimates and Fermi Rounds

Estimation interview questions are in an odd place right now. The classic brainteaser, golf balls in a school bus, shows up as a standalone question less often in some product loops than it did a decade ago. The math itself has not gone anywhere. It runs inside consulting cases as market sizing, inside product rounds as “how big is this, and is it worth a team?”, and inside technical rounds as storage, requests per second and cost. People also search for the same thing under other names: guesstimate questions, Fermi questions, back-of-the-envelope estimation.

This post is for anyone doing estimation interview prep for a consulting, product or technical loop. It covers what the round scores, the question types with real examples, one estimate worked start to finish, the mistakes that cost the most, and a two-week plan. I teach it through a course we built for exactly this, Estimation Interview Prep: AI Coach Edition, because the course is how I would practise it. You can use the method without ever opening the course.

A one-minute walk through the course
Watch the tree: the learner taps a guess first, then the branches grow and the values multiply up into the answer.

What estimation interview questions actually test

Every estimation question is a small model of real work. A decision needs a number, nobody has it, and you have ten minutes. Product and strategy work runs into this problem all the time, usually right before a meeting. The interviewer is watching how you get from a vague prompt to a number you can defend. They are not checking whether you guessed the true count.

Interviewers rarely write “correct answer” on a scorecard. They write down signals, and the same five come up in almost every room:

  • Clarify. You pinned down what counts, where, when and the unit before any math.
  • Structure. You built a tree whose parts are easier to guess than the whole.
  • Assume. Each number came with a reason or an anchor.
  • Math. Friendly numbers, units kept on every line, said out loud.
  • Check. A second route, a range, and what the number means.

One correction to common advice. People say “the number doesn’t matter.” The digits don’t. The order of magnitude does. An answer ten times off usually means an assumption broke, and a strong candidate notices before the interviewer does.

The mock interviews in the course are scored on a hiring rubric built from those signals. The weights are worth knowing, because they show where the points are.

How a mock estimation round is weighted
Clarify and scope
15%
Structure
25%
Assumptions
20%
Math and units
15%
Sense-check and so-what
25%
Weights from the course’s mock report rubric. They sum to 100.
Half the score sits in the structure and the check. The arithmetic is 15 percent.

Most candidates prepare in the opposite proportion. They drill multiplication and skip the check and the closing sentence, which is where a quarter of the score sits.

Types of estimation questions, with real examples

The questions look endless, but they fall into a handful of families. Each family has its own tree, so recognising the family in the first ten seconds is most of the battle.

Market sizing interview questions. Customers times frequency times price, or TAM narrowed honestly to a share you can defend. Examples: size the US market for driverless ride-hailing in 2030, or estimate the market for in-flight ads. This is the heart of consulting market sizing, and it shows up in product strategy rounds too.

Population and funnel questions. Start from people, homes, devices or accounts and filter down. Examples: how many credit card users are there in India, or how many Gmail users are there in the US. The trap is the unit: users are not accounts, and people are not households.

Capacity and throughput. Units times rate times how busy each unit can be, sized for the peak. Examples: how many drivers does Uber need for the San Francisco Bay Area, or how many planes are in the air over the US right now.

Physical guesstimate questions. Volume, weight and stacking. Examples: how many ping pong balls fit in a Boeing 747, or how many quarters stacked flat reach the top of the Empire State Building. These are the classic Fermi questions, and they reward subtracting what is already inside the container.

Revenue and pricing. How the company gets paid, then payers versus everyone, then what it keeps. Example: how much money does the Play Store make in a year, and what does a five-point fee cut do?

Rates, growth and technical estimates. Stocks versus flows, per day to per second, bytes and bandwidth. Examples: how many YouTube videos are watched each day, then per second at the peak, or, as the PM for Gmail, how much storage will you buy for next year? I worked a full answer to a sibling of that question in the Google Photos storage estimation post.

The same method in four different rooms

What changes between interviews is the purpose of the number. Before you build anything, ask yourself what decision the number is for. The answer tells you which room you are in, even when nobody announces it.

One method, four rooms
ConsultingShould the client act?Ask for the data. Within about two times is fine. Close with a recommendation.
ProductWhat should we build or change?Assume and say why. Close with a product call and a metric.
TechnicalWill it hold, and at what cost?Exact units, sized for the peak. Close with capacity and headroom.
Ops and financeDoes each unit make money?Per unit, to the cent. Close with break-even and payback.
The five steps stay the same. The data you ask for, the precision and the closing line change.

The practical difference is biggest between consulting and product. A case interviewer usually holds a fact sheet, so guessing a number they would have handed you costs credibility. A PM interviewer expects you to make the assumptions and then do something with the result. If you are preparing for case interviews more broadly, the MBB case interview post covers the rest of the case around the sizing.

A worked example: how many cafés are in a city of 600,000?

This is the example Coach Tara teaches in the first module of the course. Port Alder is a made-up port city of 600,000 people, and every number below is a practice assumption. The question: how many cafés does it have?

Clarify. Four choices in under thirty seconds: sit-down coffee shops with chains included, inside the city limits, open today, counted as places. The unit is not pedantry. If you are counting places, a café’s capacity has to be measured in cups a day, not chairs.

Break it down. The rule that decides the split: break a number down only when its parts are easier to guess than the whole. “City area times cafés per square kilometre” fails that test, because cafés per square kilometre is the answer in disguise. “Cups drunk a day divided by cups one café sells” passes, because people are countable and coffee habits are familiar. Say the tree before any number. It gives the interviewer a map and lets them redirect you early.

Estimate and compute. One anchor and four assumptions, each with a reason, then the math out loud with units on every line.

The Port Alder estimate, line by line
600,000 peopleThe anchor: the city’s census.
↓
× 20% buy a café coffee today = 120,000 buyersThe shakiest number. It could be 10% or 30%.
↓
× 1.5 cups each = 180,000 cups a dayMost buy one, some buy two.
↓
÷ 300 cups one café sells a day12 hours open × 25 cups an hour.
↓
About 600 cafésLikely 200 to 1,400.
One anchor, four assumptions, and the weak one named before anyone asks.

Notice the 20 percent. It has no anchor and could plausibly be 10 or 30, a threefold swing. Strong candidates point at their weakest number themselves. In the coach session this is a question Tara asks you directly: which number will the interviewer push on first?

Sanity-check. Check from a side that does not share your shakiest assumption. Per person, 600 cafés is one for every 1,000 people, plausible for a busy city. Per café, 300 cups at $4 for 365 days is about $440,000 a year, enough for rent and a few staff. The strongest check is a supply route: 200 km of busy streets times 3 cafés per km is also 600, and it never touches the 20 percent.

Say the range. Push each shaky input to a plausible edge. The low case, 10 percent buying and 400 cups per café, gives about 225. The high case, 30 percent and 200 cups, gives about 1,350. So the closing line is: about 600 cafés, likely 200 to 1,400, the share who buy coffee on a given day moves it most, and I would check it by counting one high street. The range is wide, but it sits on one step of the powers-of-ten ladder. Six thousand would be a different city.

Common mistakes in guesstimate and market sizing answers

A weak answer and a strong answer often land on similar numbers. The difference is what the interviewer could follow. Compare two openings to the café question. The weak one: “Let’s say 30 percent of people drink coffee, so…” The strong one: “I’ll count sit-down cafés inside the city, open today, and I’ll get there from cups drunk a day over cups one café sells.” The second has earned two signals before a single number.

The traps I see most:

  • Multiplying before scoping. The tree ends up answering a different question.
  • A split that hides the answer. “Restaurants per block” is the same guess in a new coat.
  • Personal habit as the main anchor. “I buy two coffees a day” is the easiest number in the room to knock over. Use it only as a check.
  • Silent math. The interviewer cannot credit what they cannot follow.
  • False precision. “3,742 cafés” sounds made up. A round number with a range sounds confident.
  • Averaging two routes that disagree. If demand says 1,400 and a pilot district scaled up says 3,000, the gap is a diagnosis to make, not a number to split. Look for different definitions, a biased sample, a unit slip, or a stock mixed with a flow.
  • Caving or digging in under pushback. When the interviewer says “that’s too high”, ask which part looks high, walk that one branch, then change one input for a reason or hold it with one.

A rough clock helps with all of these. For a fifteen-minute question: two minutes to clarify, three on structure, seven on numbers, two on the check and range, one to say it. Most people run out of time in the numbers and never check or close.

How the estimation interview prep course is built

The course has 15 modules. The first 14 each follow the same loop, and the last is one full round.

What one module looks like
Each moduleRead the notesTrain with Coach Tara in the labMock with OwenWritten report
14 readings, 14 coach sessions and 15 mock interviews: 29 voice scenarios in all.

The reading is a short written lesson with a worked example and practice questions. The coach session is about 25 minutes with Tara, a voice AI coach who teaches the idea live and then runs two answer drills: she asks a real interview question, you work it out loud for two or three minutes, and she gives one strength, one gap and the one move that would lift it, then lets you retry the weak part. The mock is about 30 minutes with Owen, the interviewer. He asks one main question, presses with follow-ups like “why that number?” and “check it another way”, throws one curveball, and gives no hints. A written report with a score per dimension and a hire band arrives after the call.

#ModuleThe mock question
1The MethodRestaurants in San Francisco
2Top-Down and Bottom-UpShampoo bottles sold in the US each year, built from both sides
3Market SizingThe US market for driverless ride-hailing in 2030
4Population FunnelsCredit card users in India
5Capacity and ThroughputDrivers Uber needs for the San Francisco Bay Area
6Volume and WeightPing pong balls in a Boeing 747
7Revenue and PricingWhat the Play Store makes in a year
8Rates and GrowthYouTube videos watched per day, then per second at peak
9Technical EstimatesStorage to buy for Gmail next year
10Mental MathA rapid round, then quarters to the top of the Empire State Building
11Sanity Checks and TriangulationZomato delivery riders in Bengaluru, checked from the other side
12Sensitivity and RangesAR glasses shipped in 2031
13Tailoring by InterviewA grocery chain’s first store in Montreal, as a consulting case
14Presenting Your AnswerIn-flight ads, presented to a VP in sixty seconds
15Full RoundA 40-minute round on a bike-share launch

The Estimation Lab: a Motion App that draws the estimate while the coach talks

The coach sessions do not run on voice alone. Beside the call sits the Estimation Lab, which is what we call a Motion App: a live app next to the conversation that the voice agent controls while it talks, and that tells the agent what you tap. The idea behind the name is that the motion is the lesson. You watch the numbers move while the coach explains them.

Every important idea in the lab starts with a prediction. Tara asks a question, you tap a guess or say it, and the guess locks in with nothing revealed. Only then does the scene move to the truth.

Predict first, then the reveal
Coach asks You tapa guess Scene reveals Coachexplains
Committing to a guess before the answer shows is what makes the reveal stick.

A few real scenes from this course’s lab:

  • The factor tree. The café estimate as a 3D tree. Known facts are solid blocks with a pin, assumptions are lighter with a dashed rim, and the shaky one wears a red outline. When the coach runs the numbers, light climbs the rods and 600 lands in the top block.
  • Packing a bus. Golf balls pour into a see-through city bus cabin with the seats already taking their share. You guess 80 thousand, 800 thousand or 8 million before the count settles near 775 thousand.
  • The airport checkpoint. Travellers flow through ID podiums, X-ray lanes and body scanners at 2,500 an hour. You pick which stage chokes, then watch the queue pile up in front of it.
  • Two routes on one summit. A demand route and a supply route climb staircases toward one number. When a third route lands far away, the lesson is to find the broken assumption, not to average.
  • The sensitivity tornado. One bar per assumption around the base answer. The widest bar grows out last, and that is the assumption to defend and test.

The lab matters because estimation is hard to learn from a transcript. Seeing a guess land a full power of ten away from the plausible band teaches order of magnitude faster than being told. We have invested in these labs across our interview courses: Arch Lab for system design, Stats Lab and Query Lab for data science, Estimation Lab here, Deploy Lab, Defense Lab, and Merger, DCF and LBO labs for investment banking.

What it costs

The course content is free. There is no separate course fee, and nothing is locked behind a higher tier: every reading, coach session and mock in the course is included. Voice sessions use the minutes on your plan, and you can start on the free plan, which comes with 25 one-time minutes. That is about one coach session, enough to judge whether this way of practising works for you. Tough Tongue AI is a voice AI company, and the same minutes cover any scenario on the platform. All the interview courses are on the courses page.

A two-week estimation interview prep plan

Two weeks of steady work is enough for most people, as long as the practice is out loud. Reading solved examples feels productive and transfers poorly, because the round is spoken.

Two weeks, one module loop a day
  1. 1
    Days 1 to 2: the method and both routesModules 1 and 2. Five steps, then demand side against supply side.
  2. 2
    Days 3 to 7: the question types your loop usesModules 3 to 9. Consulting: market sizing and funnels first. Technical: capacity, rates and technical estimates first.
  3. 3
    Days 8 to 9: mental mathModule 10. Powers of ten, friendly rounding, units on every line.
  4. 4
    Days 10 to 11: checks and rangesModules 11 and 12. A second route, then low, base and high.
  5. 5
    Days 12 to 13: the room and the closeModules 13 and 14. Tailor to the interview, then answer first and handle pushback.
  6. →
    Day 14: the 40-minute full round
Read the notes the night before, take the coach session, then the mock, and fix the top improvement from each report before the next one.

Three habits make the plan work. Take the coach session before the mock, because the mock gives no hints. After each mock, read the report and redo only the weakest dimension, out loud, the next morning. And keep a one-page anchor sheet of numbers you can cite without thinking: populations, household sizes, seconds in a day, file sizes. If you are short on time, do modules 1, 2, 11 and 14 and the full round. That covers the method, both routes, the check and the close, which is most of the rubric.

If estimation is one round of a larger loop, the product manager interview preparation guide covers the rest of a PM loop. We wrote companion posts for the other coach courses too: data science interview prep, the system design interview course, the forward deployed engineer interview, and security engineer interview prep.

Closing thought

Estimation is one of the few interview rounds that maps cleanly onto the job. You will be asked for a number you do not have, with a decision waiting on it, and the people in the room will trust you in proportion to how clearly they could follow you. That is a skill you build by saying estimates out loud and having someone push on them. The technology for practising this with an AI coach is not perfect yet. A voice model still mishears a “million” for a “billion” now and then, which is why Tara is told to ask. But it is good enough that the hard part, getting many spoken repetitions with honest feedback, is no longer hard to arrange. If you want to try it, start with the first module of the estimation course and see whether your café count survives the second route.

A
Ajitesh
Tough Tongue AI
Share