Microsoft — Ads Identity — Product Sense & Metrics

Role context: Senior Applied Scientist, Ads Data Management Platform · Est. study time: 35 min · 5 practice questions

How to prepare for this role

The day-to-day is building the identity graph — the system that resolves the scatter of identifiers a person leaves across devices (cookies, device IDs, hashed emails, Microsoft Account logins, IPs) into a single person and household — and proving a better graph measurably improves ads personalization, audience targeting, and measurement, including on Copilot / MAI surfaces. The instinct that separates strong candidates: refusing to celebrate match rate, reasoning about the distinct costs of merging two people vs splitting one, and treating privacy/consent as a hard constraint, not an afterthought.

Where to spend your prep time:

• Identity/measurement product sense (this article) — what the graph is, the precision-vs-coverage trade-off, and the metrics that mean quality (not match rate).
• Ads measurement & causal inference — you usually can't randomize identity, so proving the graph's impact needs incrementality (ghost ads / PSA), geo tests, and holdout audiences; measuring precision/recall without ground truth. See the A/B Test & Causal Inference section.
• Entity resolution & the statistics of matching — probabilistic record linkage, calibration of match scores, precision/recall estimation, missing-data (consent) modeling.

The through-line: resolve people accurately, at the coverage each use case needs, prove it lifts advertiser outcomes, and never overstep privacy.

What Microsoft Ads Identity actually is

Microsoft Advertising is Microsoft's digital advertising business (Search/Bing, the Edge/Windows ecosystem, native/display, the ad exchange, and increasingly Copilot). The identity graph is the infrastructure beneath it — officially, "a database of customer information that allows Microsoft Advertising to identify users on multiple devices," covering the US and key European countries. The identity to carry into every answer: the graph turns a scatter of anonymous identifiers into people — and every ad personalization, every frequency cap, every reach and conversion number depends on it being accurate.

Two things make this different from a normal product-analytics job:

• It's a precision-vs-coverage machine with asymmetric costs. Linking identifiers is a matching problem: deterministic links (same login / hashed email / authenticated) are high-precision but sparse; probabilistic links (same IP, device, behavior) add coverage but risk false merges. A false merge (two people collapsed into one) mis-personalizes ads, breaks frequency caps, mis-attributes conversions, and is a privacy risk; a false split (one person seen as several) under-counts reach and over-serves. The job is choosing the operating point per use case.
• Signal is shrinking and privacy is the frame. Third-party cookies are going away and consent (Consent Mode, GDPR) removes signal, so measurement gaps are increasingly modeled (Microsoft's Advanced Consent Mode estimates lost conversions from aggregate trends). Much of the science is measuring well with missing data while respecting privacy.