# Design a lookalike audience targeting system: system design interview question

> Behavior-based audience selection at scale, with daily refreshes, consent changes, relevance checks, and limits on advertiser access.

Source: https://sysround.com/questions/design-lookalike-audience-targeting-system
Updated: 2026-11-05


Difficulty: hard. Levels: Senior, Staff+.

## How the interviewer opens it

"Design a service that finds users whose behavior resembles an advertiser's seed audience. The service must respect user consent and exclude sensitive targeting attributes."

## What to clarify first

- Seed audience
- Audience scale and refresh
- Events, retention, and durability
- Read and write traffic
- Consent, audience status, and availability
- Allowed data and revisions
- Model and relevance
- Advertiser reports
- Request controls
- Access and scope

## Scope of a strong answer

Core requirements:

- Create an audience from an advertiser's approved seed audience.
- Refresh audience membership daily.
- Check audience eligibility during an internal ad request.
- Enforce consent, audience status, and sensitive data rules.

Usually out of scope:

- Match advertiser identities to platform identities.
- Design the similarity model.
- Design auctions, bids, or billing.
- Export user lists to advertisers.

## What interviewers probe

### Audience refresh deadlines

Thousands of audiences share the same user population. Each job must finish on time without publishing partial results or results from obsolete configurations.

### Audience relevance

The approved model supplies similarity scores, not proof of useful future behavior. The service must measure relevance without future data leakage.

### Consent and audience status

Stored membership can outlive consent or audience approval. Every ad decision needs current permission despite cache delays, regional failures, and restored data.

### Sensitive data and advertiser access

Approved behavior can correlate with sensitive traits. Repeated seed changes can reveal information, while policy changes can invalidate existing audiences and unfinished jobs.

## Common mistakes

- Designs the similarity algorithm but omits the audience data flow and eligibility API.
- Runs a separate raw event scan for each audience without estimating the work.
- Treats stored membership as permission despite consent withdrawal, an audience pause, or a policy disable.
- Rejects old job attempts but permits publication from obsolete seed, configuration, or policy versions.
- Assumes that removal of explicit sensitive attributes also removes sensitive proxies.
- Exposes exact counts or user-level membership through advertiser reports.

## The bar by level

- **Senior:** Estimates shared computation and membership traffic. Explains candidate selection, relevance tests, atomic publication, and recovery. Rejects obsolete seed and policy jobs. Bounds consent and status cache delays. Enforces seed minimums and report controls.
- **Staff+:** Defines control ordering across policy changes, publication, caches, and recovery. Proves permission deadlines under regional failures. Traces deletion through retained data. Explains privacy limits under repeated requests. Balances recall, refresh cost, and advertiser fairness with measurable limits.

Practice this question with an AI interviewer: https://sysround.com/practice/new?problem=advertising-audience-targeting


---
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.