# Design an ad conversion prediction system: system design interview question

> Purchase probabilities with rare outcomes, delayed reports, incomplete attribution, and strict score latency.

Source: https://sysround.com/questions/design-ad-conversion-prediction-system
Updated: 2026-11-05


Difficulty: hard. Levels: Senior, Staff+.

## How the interviewer opens it

"Design a system that predicts the probability of a purchase or another conversion for an advertisement. The system must return predictions for live ad requests."

## What to clarify first

- Prediction target
- Traffic
- Outcome rate and read:write mix
- Purchase delay
- Attribution and missing links
- Features and privacy
- Freshness and retention
- Latency and availability
- Event delivery and finality
- Scope limits

## Scope of a strong answer

Core requirements:

- Define the conversion target and create labels from ad and purchase events
- Train and evaluate a model that returns conversion probabilities
- Return probabilities for live ad requests
- Handle delayed events and incomplete purchase links

Usually out of scope:

- Auction design and bid selection
- Payments and conversion fraud detection
- Multi-touch attribution
- Estimates of purchases caused by an ad

## What interviewers probe

### What the probability means

A purchase can follow many ads. Missing links vary by user group, so reported purchases do not reveal every purchase.

### When labels become final

Future purchases and late exposures can change credit. A model can mistake incomplete labels for low purchase probability.

### Probability accuracy

Most impressions have no credited purchase. Ad bids need accurate probabilities, not only a useful order of ads.

### Scores for live ads

The service has 20 ms to return a probability. Feature delays and ad selection can change the data that the model sees.

## Common mistakes

- Candidates define a purchase probability without a time window or an attribution rule.
- Candidates finalize labels without complete impression, click, and purchase data.
- Candidates use accuracy to judge data with only 0.1% positive labels.
- Candidates sample negatives without a correction for probability estimates.
- Candidates treat the absence of an attributed purchase as proof that no purchase occurred.
- Candidates use feature values that the score service did not have during inference.

## The bar by level

- **Senior:** Calculates scale and label maturity across all event sources. Uses exact score inputs for training. Designs probability checks, bounded latency, compatible releases, and label repair. Evaluates incomplete attribution by cohort.
- **Staff+:** States which probabilities the available data can identify. Separates coverage changes, purchase delays, and ad selection feedback. Defines release evidence and revision policy. Justifies privacy limits, sparse-cohort conclusions, and cost choices.

Practice this question with an AI interviewer: https://sysround.com/practice/new?problem=ad-conversion-prediction


---
SysRound (https://sysround.com) is an AI practice tool for system design interviews. It has a voice or typed AI interviewer and a shared whiteboard. At the end, you get feedback on each skill and a model answer. The AI feedback can be wrong.