Checking and presenting the answer

Test an estimate before you say it as final in a PM interview. The slips that wreck an estimate are easy to make under pressure and easy to catch once you look. Cracking the PM Interview has you give your number and test it right away in front of the interviewer, fixing what you find, so treat your first figure as a working answer.

Here is the estimate we'll test. You are the PM at a language-learning app, and the team wants to sell live speaking practice with an AI partner as an upgrade, at $10 a month on top of the existing subscription. How much extra yearly revenue could it bring in from the US? These inputs are my assumptions:

• 5 million monthly users in the US, 8% of whom pay: 5,000,000 × 8% = 400,000 subscribers.

• 10% of subscribers upgrade and stay on it all year: 400,000 × 10% = 40,000.

• Each upgrade brings in $10 × 12 = $120 a year.

• 40,000 × $120 = $4.8 million a year.

Shrink the answer to a size you can picture

A total means little until you divide it by people. $4.8 million ÷ 400,000 subscribers = $12 per subscriber a year, or $1 a month. This division can catch a math slip or a wrong group, but it can't test whether 10% or $120 is right, since the $12 is only those two multiplied.

To test a guess, check the parts, because a total can look fine while a part is absurd. The US has about 340 million people (US Census Bureau, 2024), and 340,000,000 ÷ 5,000,000 = 68, so one person in 68 uses this app. A town of 68,000 would have 1,000 users, 1,000 × 8% = 80 subscribers, and 80 × 10% = 8 people paying for speaking practice. That seems believable. At 800, I would stop and look.

Check with a second route

A second route is a fresh estimate that shares as few guesses with the first as you can manage. If both rest on the same assumption, their agreement can't expose an error in it. So I'll start outside the app, with people who already pay for speaking practice. These are my assumptions:

• About three in four people are adults: 340 million × 3/4 = 255 million, call it 250 million.

• 1% pay for language classes or a tutor in a given year: 2.5 million.

• 10% of those would pay for an AI partner at this price for a full year: 250,000.

• 25% of those would choose our app, subscribing first if they aren't subscribers yet: 62,500.

• 62,500 × $120 = $7.5 million. Both routes share the $120 and the guess that buyers stay all year, so this check can't catch an error in either.

$7.5 million ÷ $4.8 million ≈ 1.6, inside the factor of two I'd accept for a rough model. I chose the second route's guesses already knowing the first answer, so this agreement is encouraging and falls short of proof.

Compare to something known

Hold your inputs up against a close cousin of your product. Duolingo reported about 133 million monthly users and about 12.2 million paid subscribers at the end of 2025. 12.2 ÷ 133 ≈ 9%. That ratio is worldwide and my 8% is a US assumption, so this is rough, but it gives no reason to doubt 8%.

Duolingo also said its top tier, Max, was about 5% of paid subscribers at the end of 2024, and its CFO put it at 9% on an earnings call in November 2025. That is a different product at a different price, so I wouldn't copy either figure. Going from 5% to 9% in about a year makes 10% plausible once a tier is established. I'll use 5% as my low case, knowing a new tier could start lower: 400,000 × 5% = 20,000 upgrades, and 20,000 × $120 = $2.4 million. For a high case I'll double my guess to 20%: 400,000 × 20% = 80,000 upgrades, and 80,000 × $120 = $9.6 million.

Compare with your own business too. If subscribers pay $80 a year on average (my assumption), the current business is 400,000 × $80 = $32 million. An add-on worth $4.8 million would be about 15% of that, which sounds possible for a feature only a minority buy.

Spot an answer ten times too big or too small

An error of that size is often one of two kinds: a slip while multiplying, or a correct multiplication of the wrong input, such as monthly treated as yearly, or all users where you meant paying users.

Count the digits. When you multiply two whole numbers, add the two digit counts: the product has that many digits, or one fewer. 40,000 has 5 digits and 120 has 3, so the product needs 7 or 8. 4,800,000 has 7. An answer of 480,000 has 6, so a zero went missing. Percentages and decimals don't follow the rule, and a slip can still pass: 48,000,000 has 8 digits. A wrong input multiplied correctly always passes.

Divide back to one person. For a wrong input, you need a test that looks at meaning. Suppose I had applied the 10% to all users by mistake: 5,000,000 × 10% × $120 = $60 million. Divide by the people who can buy it: $60 million ÷ 400,000 subscribers = $150 each per year. The upgrade costs $120, so the answer is impossible.

Compare to the whole. $60 million is nearly double the $32 million the whole subscription business brings in.

The error can run the other way. Forget the × 12 and you get 40,000 × $10 = $400,000, small enough to kill a decent project over a spreadsheet mistake. Divide back and it shows: $400,000 ÷ 40,000 upgrades = $10 a year for something priced at $10 a month.

When a test fails, work out where

Start with what's fastest to rule out, and rerun the arithmetic from your written lines. Then read back through the numbers you assumed for one that now looks wrong. Question the equation last, since changing it means rebuilding.

Then change the number out loud. Say I realize beginners may not be ready for speaking practice. Assume a quarter of subscribers are beginners, none of them upgrade, and the rest upgrade at the same 10%. The pool is 400,000 × 75% = 300,000, upgrades are 300,000 × 10% = 30,000, and the answer is 30,000 × $120 = $3.6 million. I'll treat that as a what-if and stay with $4.8 million below.

Lead with the number

When you finish, give the result, then the range, then the assumption that matters.

"About $5 million a year in the US, and I'd believe anything from about $2.5 million to $10 million. That range comes from one assumption, the share of subscribers who upgrade, which I set at 5% to 20% with 10% in the middle. I held users and paying share fixed, so the true range is somewhat wider. A second route, starting from people who already pay for tutors, gave about $7.5 million. That's encouraging, though I picked those guesses knowing my first answer."

Practice this on Size a New IKEA Store's Opening Month for a CEO Briefing and give your answer in a few sentences.

Connect the number to a decision

A PM estimates in order to decide something, so compare the number with the decision's cost. Assume a team of five at $250,000 each, which is 5 × $250,000 = $1.25 million a year, and AI usage costing $3 per upgrader per month, which is $3 × 12 = $36 a year. If upgrades are billed on the web with no app-store fee, each one keeps $120 minus $36 = $84.

To break even: $1,250,000 ÷ $84 ≈ 15,000 upgrades. 15,000 ÷ 400,000 ≈ 3.75%, or about 4% of subscribers.

App stores take 15% to 30% on mobile subscriptions. Take the top of that range, such as Apple's first-year rate, knowing not every upgrade pays it over its life. At 30%, the store takes $120 × 30% = $36, so each upgrade keeps $120 minus $36 minus $36 = $48. Break-even becomes $1,250,000 ÷ $48 ≈ 26,000 upgrades, and 26,000 ÷ 400,000 = 6.5% of subscribers, above my 5% low case. The rest of this lesson keeps the $84 and its web-billing assumption.

"With web billing, we break even at about 4% of subscribers. My low case is 5%, which is close, so I'd test before building: show the offer and price to a sample of subscribers for two weeks and count who tries to buy. If fewer than 4% even try, I wouldn't build it."

That test is an early read. It counts attempts over two weeks, so it misses people who would upgrade later and can't show who would really pay or still be paying after 12 months, which the $120 assumes. The 4% line is a judgment call, and clearing it earns only a small launch to measure those. I'd also watch a guardrail, a metric that warns of harm elsewhere, such as cancellations among subscribers who see the offer.

Handle pushback

"That seems high." Ask what they would use, then rerun it in front of them. "At 5%, it's $2.4 million in revenue, and after usage costs we'd clear the $1.25 million team cost by about $430,000." (20,000 upgrades × $84 = $1.68 million, minus $1.25 million.)

"Are you sure?" Restate your reasoning and how sure you are. Change the number only when you've heard a reason to.

A new fact. Recompute. If only 4% of users pay, you have 5,000,000 × 4% = 200,000 subscribers, and 15,000 upgrades now means 15,000 ÷ 200,000 = 7.5% of them. That is above my low case, so a loss is now possible, and the test matters more.

How people lose the room

• A long silence followed by a number. The interviewer can't credit reasoning they didn't hear.

• Too many segments. Twelve groups, each with a guess, leave no time to finish.

• Talking yourself down. "Sorry, my arithmetic is rusty" tells the listener nothing useful. A range tells them how far off you might be.