"How many gas stations are in the US?" "How much storage does a year of photos on a phone take?" Product, consulting and engineering interviews still use these Fermi questions. Nobody checks your final number against a database. They check whether you can turn a vague question into a structured chain of reasonable guesses.
Spend fifteen seconds on scope. Gas stations: the whole US, or including truck stops? Photos: one user or everyone? Saying "I'll assume X; tell me if you meant something else" costs nothing and avoids solving the wrong problem.
There are usually two routes: from the demand side (people → cars → fill-ups per week → fill-ups one station handles) or the supply side (towns × stations per town). Pick one, say why, and write the formula before plugging numbers in. The formula is the part being graded.
US population about 330 million, roughly 250 million cars, a fill-up about once a week. Round aggressively: 1, 2, 5 and powers of ten. Precision you don't have only slows the arithmetic and invites mistakes.
250M fill-ups a week ≈ 35M a day. If a station serves about 10 cars an hour over 18 hours, that's ~200 a day, so 35M / 200 ≈ 175,000 stations. Say each step so the interviewer can follow, and catch you, before the error compounds.
Try the supply side quickly: ~20,000 towns and cities, a few stations each in small towns and many more in cities. Low six figures again. When two independent routes land in the same order of magnitude, say so: it's the strongest signal you can give. (Published counts are around 150,000, so the method holds up.)
Finish with the assumption that moves the answer most ("throughput per station; if it's 20 cars an hour, the answer halves") and how you'd check it. That sentence shows judgment, not just arithmetic.
Disclosure: I'm part of the team behind Linkjob.ai, an AI interview assistant. The method above works with nothing more than a pen.