The Geometric Distribution: Trials Until the First Success

The binomial distribution answers "how many successes in a fixed number of trials?" The geometric distribution asks a related but genuinely different question: "how many trials until the first success?" Free throws until the first make, sales calls until the first sale, product tests until the first defect found — any repeated, independent trial with constant success probability, counted up to and including whichever trial first succeeds, follows a geometric distribution.

The Formula

For independent trials each with probability p of success, the probability that the first success occurs exactly on the k-th trial is:

The logic is straightforward: the first k−1 trials must all fail (each with probability 1−p, and they're independent so their probabilities multiply), and then the k-th trial must succeed (probability p). Unlike the binomial distribution, which asks about a fixed number of trials, X here is itself the random, uncertain quantity — the trial count is what's being predicted, not something fixed in advance.

Worked Example: Free Throw Shooting

Problem: A basketball player makes free throws at a 70% success rate. Assuming each shot is independent with the same 70% probability, what's the probability that her first make of a practice session happens on exactly her 3rd attempt? What's the probability she needs at most 3 attempts?

Here p = 0.7 and k = 3, so the first 2 shots must miss (probability 0.3 each) before the 3rd shot goes in:

P(X=3) = (0.3)² × 0.7 = 0.09 × 0.7 = 0.063

There's a 6.3% chance her very first make happens specifically on the 3rd attempt. For "at most 3 attempts," sum the individual probabilities for k=1, 2, and 3 — or use the cumulative formula directly:

P(X≤3) = 1 − (0.3)³ = 1 − 0.027 = 0.973

There's a 97.3% chance she makes at least one shot within her first 3 attempts — a high probability that makes sense given her strong 70% per-shot success rate; a 30% miss rate compounding across only 3 shots leaves relatively little room for a run of bad luck.

Mean and Variance

For the free-throw example: μ = 1/0.7 ≈ 1.43 — on average, the player needs only about 1.43 attempts to get her first make, reflecting how likely each individual shot already is to succeed. Variance = (1−0.7)/0.7² = 0.3/0.49 ≈ 0.612, giving σ ≈ 0.78.

A Second Worked Example: Sales Calls

Problem: A salesperson closes 15% of cold calls. What's the probability the first sale happens exactly on the 5th call? What's the expected number of calls until the first sale? What's the probability more than 10 calls are needed?

p = 0.15, k = 5: P(X=5) = (0.85)⁴ × 0.15 ≈ 0.0783 — roughly a 7.8% chance the very first sale lands precisely on call 5.

μ = 1/0.15 ≈ 6.67 calls expected on average before the first sale — noticeably higher than the free-throw example, reflecting the much lower 15% per-call success rate.

P(X>10) = (1−p)¹⁰ = (0.85)¹⁰ ≈ 0.197 — there's still roughly a 1-in-5 chance the salesperson needs more than 10 calls to land the first sale, a useful number for setting realistic daily call-volume expectations.

The Memoryless Property

The geometric distribution has an unusual and genuinely distinctive feature: it's memoryless. Given that the first k trials have already failed, the probability distribution of how many more trials are needed is exactly the same as it was at the very start — the process has no "memory" of the failed attempts already behind it. Concretely, if the salesperson has already made 5 unsuccessful calls, the probability distribution for how many more calls until the first sale is identical to the distribution for a fresh salesperson who hasn't made any calls yet. This can feel counterintuitive — a common but mistaken instinct is to think a sale becomes "due" after a run of failures (the same fallacy behind the gambler's fallacy in coin flipping) — but as long as each trial is genuinely independent with a truly constant success probability, the past failures carry no predictive information about what comes next.

Real-World Uses

Manufacturing quality control: the number of items inspected until the first defect is found follows a geometric distribution, useful for setting expectations around inspection workload and understanding how often a given defect rate produces early versus late detections.

Reliability and failure analysis: the discrete geometric distribution is the direct discrete-time analogue of the continuous exponential distribution used in reliability engineering — both share the memoryless property, and both model "time until an event" in their respective settings (trial count for geometric, continuous time for exponential).

A/B testing and conversion funnels: the number of website visitors until the first conversion, at a known conversion rate, follows the same logic as the sales-call example — directly relevant to the conversion-rate thinking covered in A/B testing.

Game design and random-drop mechanics: the number of attempts needed before a rare in-game item drops, at a fixed per-attempt drop rate, is a textbook geometric distribution — the same reasoning game designers use to calibrate how often, on average, a "grind" should reward a player, and the same reasoning behind the classic feeling that a hoped-for drop is "overdue" (an illustration of the memoryless-property fallacy discussed above, since the drop rate genuinely doesn't change based on how many unsuccessful attempts came before).

Common Mistakes

Confusing the geometric distribution with the binomial distribution. The binomial fixes the number of trials n in advance and counts successes; the geometric fixes the target of exactly one success and lets the number of trials vary. Mixing up which quantity is fixed and which is random is the most common conceptual error with this distribution.

Believing in a "hot streak" or a sale becoming "due." The memoryless property directly contradicts this intuition — for a genuinely constant-probability process, no number of past failures changes the odds of the very next trial.

Using P(X=k) instead of P(X≤k), or vice versa, for the wrong question. "Probability the first success is exactly on trial k" and "probability the first success happens by trial k" are different questions with different formulas — always check which one a problem is actually asking.