A device graph is a stored map of which digital identifiers belong together. It records that a browser cookie, a phone advertising ID, a television identifier and a hashed email address are, in the judgement of whoever built it, attached to the same person or home. Nothing in the identifiers says so: a cookie on a laptop and an advertising ID on a handset are separate strings from separate systems with no shared key. The graph supplies the missing link, which lets a buyer count one viewer instead of four, cap frequency across screens, or credit a mobile purchase to a desktop impression.

The Federal Trade Commission's January 2017 staff report on cross-device tracking used the term for the set of devices a company associates with one consumer, that consumer's household, or each other. That definition has held; what has changed is the material the map is built from.

How the map is built

The structure is a graph in the mathematical sense: identifiers are nodes, and an assertion that two belong together is an edge. A vendor collects identifiers seen together, most often through IP address co-occurrence, then weights each candidate pair on behavioural features: how often the pair appears at the same address, at what times of day, and against which unusual combinations of sites or apps. A classifier scores the pairs, and a clustering step groups the nodes into components, each representing one person or home.

Patent filings on shared device graphs describe the deterministic layer as an undirected graph with potentially millions of connected components, traversed with tooling such as recursive SQL queries or Apache Spark GraphX. The probabilistic layer estimates edges nobody observed.

According to The Trade Desk, which operates the Identity Alliance graph coalition, identity resolution means associating pseudonymised identifiers at household and individual level. It places streaming and connected television identifiers at household level, since several people watch one screen, and cookies, mobile advertising IDs, hashed emails, Unified ID 2.0 and RampID at individual level. That split matters commercially: a household edge supports reach and frequency claims, while an individual edge is required before a buyer can claim a person was reached.

Deterministic and probabilistic construction are usually mixed rather than chosen between. The FTC noted that companies holding login data frequently work with firms doing probabilistic matching, and that a small deterministic core is used to train the probabilistic model. ID5 describes the same loop running in reverse, with probabilistic linkages validating deterministic matches.

Where it sits in the transaction

The graph is not passed in a bid request. It is applied on either side of one. On the sell side, a publisher or exchange adds an identifier the graph asserts is related to the device in front of it. On the buy side, a demand-side platform expands the identifiers in an ordinary bid request against a licensed graph before deciding what to bid.

OpenRTB 2.6 made the sell-side version visible. The extended identifiers array in the user object carries a source field for the party that defined an ID, an inserter field for the party that put it in the bidstream, and a matcher field for the party that made the link, which IAB Tech Lab implementation guidance says may be a device graph vendor. A match method value declares how the link was formed. Values of 1 and 2 cover an observed identifier and a cookie sync; values of 3, 4 and 5 declare that matching occurred, and the guidance requires the value to be inherited downstream, so an exchange receiving a bridged ID cannot relabel it as observed. The specification's worked example is a connected television request where a publisher adds the advertising ID of an Apple device in the same home, sourced from household IP and a device graph, tagged with match method 5.

Delivery has historically been by file. Experian described the Tapad Graph as a weekly file of advertising IDs clustered at individual or household level, built from customer-supplied input IDs. Adobe Audience Manager exposed graph selection as a configuration choice: a co-op graph combining deterministic and probabilistic data, a private graph using only the company's own deterministic data, or a third-party graph from LiveRamp or Tapad at person or household level.

Origin and evolution

The category dates to 2010, when Kamakshi Sivaramakrishnan founded Adsymptotic, later renamed Drawbridge, and Are Traasdahl founded Tapad. Both sold probabilistic cross-device matching as an alternative to the login data Google and Facebook already held.

Validation came through public competition. Drawbridge and the IEEE International Conference on Data Mining ran a cross-device connections challenge on Kaggle from June 1 to August 24, 2015, drawing 340 teams asked to link cookies to devices without names, emails or phone numbers. The metric was F0.5, weighting precision above recall: an early acknowledgement that a wrong link costs more than a missing one.

Regulators followed the same year. The FTC held a cross-device tracking workshop on November 16, 2015, and published its staff report on January 23, 2017. Preparatory research found cross-device tracking companies on at least 87 of 100 tested websites, 861 third parties connecting to both test devices, and only three of the 100 privacy policies explicitly disclosing third-party cross-device tracking. The Digital Advertising Alliance began enforcing cross-device guidance on February 1, 2017, requiring an opt-out on one device to stop that device's data informing behavioural ads elsewhere. The FTC noted the limit: a consumer still opts out device by device.

Consolidation then took the independents out. The Trade Desk acquired Adbrain's assets on October 25, 2017, having carried it in a cross-device marketplace since October 2015 alongside Drawbridge, Tapad and Oracle's Crosswise. Drawbridge sold its managed media business to Gimbal in May 2018 and was acquired by LinkedIn in May 2019. Roku bought dataxu and its device graph technology in 2019, and Experian completed its purchase of Tapad from Telenor for roughly 280 million dollars in early 2021. Adform brought its graph in-house in August 2019, and Adobe deprecated its Experience Cloud Device Co-op on August 31, 2022.

Why it matters to marketers

Three campaign functions depend on the graph. Frequency capping across screens is impossible without one, because a cap per identifier is not a cap per person. Deduplicated reach has the same dependency, as does attribution across a click on one device and a conversion on another.

Vendors have long quantified the gap. Adform said advertisers without a cross-device solution attribute on average 20 to 30 percent fewer conversions, and that its graph identified an average of 3.2 devices for around 70 percent of unique cookies. Adsquare, integrating Drawbridge, Tapad and Crosswise in December 2017, cited more than 90 percent accuracy verified by Nielsen. Both figures are self-reported.

Accuracy and criticism

The strongest external evidence is unflattering, and lands on the probabilistic layer carrying most connected television targeting. Research commissioned by the Coalition for Innovative Media Measurement and Go Addressable, conducted by Truthset across nearly one billion IP records from six providers and benchmarked against two internet service providers and a distributor, found IP-to-postal accuracy of 13 percent and IP-to-email accuracy of 16 percent. Providers agreed on the same IP-to-household link only 6.4 percent of the time, and the best was 4.5 times more accurate than the worst.

FreeWheel used the same data in February 2026 to argue that IP-based targeting can miss 87 percent of households, and disclosed that IP addresses account for 17 percent of audience targeting on its own platform. Adstra and InterMedia Advertising found in July 2026 that 23 percent of residential IP addresses reached the intended geographic target. Both sellers have a commercial interest in that conclusion.

Two structural objections remain. Probabilistic recognition sits close to fingerprinting, which the UK Information Commissioner's Office treats as a technique users cannot clear as they clear cookies. And the household unit that makes graphs scale is the unit that produces the disclosure the FTC warned about, where activity on a private device surfaces as an advert on a shared one.

Disambiguation

An identity graph is the broader category. Device graphs map devices; identity graphs may also hold offline records, postal addresses and purchase history, and resolve to a named person. LiveRamp's construction, joining an offline recognition layer to a device-to-person map, is an identity graph containing one.

A universal ID is an identifier, not a map. ID5, Unified ID 2.0 and RampID are values passed in the bidstream; graphs consume and connect them.

Cookie syncing matches two vendors' identifiers for one browser. That is one browser, not several devices, and OpenRTB treats it as a distinct match method.

A household ID is an output level rather than a rival technology. Viant markets Household ID against probabilistic device graphs, but the household is a resolution setting.

Recent developments

Graph economics shifted when Google confirmed on April 22, 2025 that Chrome would keep third-party cookies, then retired most Privacy Sandbox technologies that October 17. It had already dropped its prohibition on fingerprinting effective February 16, 2025, permitting IP address use for connected television, and extended IP-based ads to European Economic Area publishers from August 3, 2026. The ICO called the December 2024 announcement irresponsible.

The buy-side answer has been to move upstream of inference. PayPal launched PayPal Ads ID in April 2026, positioning deterministic transaction linkage against paid access to probabilistic graphs; Infillion bought Catalina in February 2026 for purchase-based identity. Roku's authenticated account-level graph across more than 100 million streaming households was named an asset in Fox's acquisition. Local television runs the other way, with Madhive supplying Gray Media a device graph for cross-environment recognition.

Timeline

  • 2010: Drawbridge and Tapad founded, commercialising probabilistic cross-device matching
  • June 1 to August 24, 2015: Drawbridge and ICDM run the cross-device connections challenge on Kaggle with 340 teams
  • October 2015: Adbrain integrated into The Trade Desk's cross-device marketplace
  • November 16, 2015: FTC holds its cross-device tracking workshop
  • January 23, 2017: FTC publishes its cross-device tracking staff report
  • February 1, 2017: Digital Advertising Alliance begins enforcing cross-device guidance
  • October 25, 2017: The Trade Desk acquires the assets of Adbrain
  • May 2018: Gimbal buys Drawbridge's managed media business
  • May 2019: LinkedIn agrees to acquire Drawbridge
  • August 2019: Adform launches an in-house cross-device graph
  • 2019: Roku acquires dataxu, including its device graph technology
  • Early 2021: Experian completes its acquisition of Tapad from Telenor
  • August 31, 2022: Adobe deprecates the Experience Cloud Device Co-op
  • December 18, 2024: Google announces it will stop prohibiting fingerprinting for advertising
  • February 16, 2025: Google's revised platform policy permitting fingerprinting takes effect
  • April 22, 2025: Google confirms Chrome will keep third-party cookies
  • October 17, 2025: Chrome retires most Privacy Sandbox technologies
  • November 5, 2025: Truthset study reports 13 percent IP-to-postal and 16 percent IP-to-email accuracy
  • August 3, 2026: Google begins serving IP-based ads to publishers in the European Economic Area

Summary

Who: Independent vendors including Drawbridge, Tapad, Adbrain and Crosswise built the first commercial device graphs; demand-side platforms, data companies and streaming platforms now operate them, with LiveRamp, Experian, The Trade Desk, Roku and Viant among the named operators.

What: A stored map of identifiers, cookies, mobile advertising IDs, connected television IDs and hashed emails, clustered into components representing one person or one household, built from deterministic links and probabilistic inference.

When: Commercialised from 2010, examined by the FTC in a November 2015 workshop and a January 2017 report, consolidated into larger platforms between 2017 and 2021, and re-priced from 2025 as browser and platform policy shifted.

Where: Applied on the sell side by publishers and exchanges enriching bid requests, and on the buy side by demand-side platforms expanding identifiers before bidding, with OpenRTB 2.6 match method and matcher fields disclosing when a graph was involved.

Why: Cross-screen frequency capping, deduplicated reach and cross-device attribution require a link between identifiers that the identifiers themselves do not provide, and accuracy studies published since November 2025 indicate the probabilistic version of that link is far weaker than its pricing implies.