Identity Graph

An identity graph is a database that links multiple identifiers, device IDs, hashed emails, cookies, mobile advertising IDs, back to a single representation of a person or household, so an advertiser or platform can recognize the same individual across devices and sessions rather than treating each identifier as a separate, disconnected user.

Deterministic versus probabilistic matching#

Identity graphs are built two ways, often combined. Deterministic matching links identifiers based on a direct, confirmed connection: someone logs into the same publisher account on their phone and their laptop, so the graph can confidently tie both device IDs to one person. Probabilistic matching, by contrast, infers a likely connection from patterns, like a phone and a laptop that consistently connect from the same IP address and location overnight, without a confirmed login tying them together. Deterministic matches are more reliable but require some form of authenticated identity, like Unified ID 2.0 built from a hashed email; probabilistic matches scale further but carry more risk of an incorrect link.

Why platforms build them#

An identity graph is the layer that lets a demand-side platform or a data management platform do things that no single identifier can do alone: frequency-cap an ad across a person's phone and desktop instead of showing it to what looks like two separate strangers, stitch together an attribution path that starts on mobile and converts on desktop, or extend a lookalike audience built from known advertiser customers to new prospects who share enough matched traits, often refined further by geo targeting within the target markets that matter most.

Without some form of identity resolution, every device and browser looks like an isolated, anonymous visitor, which is exactly the state that signal loss pushes more of the industry toward as third-party cookies get less reliable across the open web.

Where identity graphs run into limits#

The reliability of an identity graph depends entirely on the quality of the underlying matches, and probabilistic links degrade in shared-device households, on public networks, or wherever the behavioral patterns a graph relies on get noisy. Privacy regulation also constrains what can legally feed an identity graph in the first place, particularly around combining data from separate sources without clear consent, which is part of why data clean rooms exist: to let two parties get some of the benefit of matched identity without either side building a full combined graph outright.

If your priority is understanding what a specific advertiser is running right now, rather than modeling who their audience is, that's a simpler problem than identity resolution and one OpenAdLibrary's ad intelligence index solves directly by tracking the live creative itself.

A practical caveat#

Coverage varies a lot by vertical and geo, and a graph built mostly from one region's logged-in traffic won't necessarily extend cleanly to a market with different login habits or device patterns. Treat any vendor's stated match rate as a starting point for testing, not a guarantee that applies evenly across every audience you plan to target.