Ghost Ads

A digital advertising incrementality testing methodology where both the test and control groups participate in the same ad auction — the control group's slot shows organic content rather than the advertiser's ad, and the platform logs a "ghost impression" recording that the user would have received the ad. This contrasts with Incrementality holdout designs (ITT) that remove the control group from the auction entirely, and with PSA designs that serve a paid charity/placeholder ad to the control group.

Origin and naming

Ghost Ads were named and formalized by Garrett A. Johnson (Boston University), Randall A. Lewis (Google), and Elmar I. Nubbemeyer (Google) in a working paper first posted to SSRN in June 2015, published in the Journal of Marketing Research 54(6): 867–884 in December 2017. The paper won the 2017 Paul E. Green Award from the American Marketing Association for most potential to contribute to marketing research practice. (Johnson, Lewis & Nubbemeyer, JMR 2017)

DoorDash's 2025 implementation documentation states that "in 2017, Google and Amazon pioneered the ghost ads technique," attributing the origination to both companies based on the authors' institutional affiliations (Lewis was at Amazon by the time of DoorDash's reference). (DoorDash Ads, 2025-10-04)

How it works — the three designs

1. Intent-to-Treat (ITT) / holdout

The control group is ineligible to receive the ad; they are removed from the auction when their slot is reached. Problem: a large portion of the treatment group never wins the auction and therefore never sees the ad, diluting the measurement signal. (Johnson, Lewis & Nubbemeyer, JMR 2017; Remerge, 2019)

2. PSA (Public Service Announcement) test

The control group sees a charity or PSA ad in the same slot. The advertiser pays for those PSA impressions — the cost of running the control group is identical to the treatment group. Additionally, the platform's targeting algorithm optimizes differently toward users likely to engage with PSA content vs. commercial ads, introducing selection bias. (Johnson, Lewis & Nubbemeyer, JMR 2017; Tinuiti, 2025-12-15)

3. Ghost Ads

Both groups participate in the same auction normally. When a control-group user's slot would have served the advertiser's ad, the ad is replaced with organic content — the platform logs a "ghost impression" to record the counterfactual. No cost to the advertiser for the control-group exposure. (Johnson, Lewis & Nubbemeyer, JMR 2017)

The founding paper's result: "compared to Intent-to-Treat or PSA experiments, advertisers can measure ad lift just as precisely while spending at least an order of magnitude less." (Johnson, Lewis & Nubbemeyer, JMR 2017)

DoorDash's auction implementation: "ghost ads participate in auctions but are skipped when establishing second prices, preventing real ads from ever bidding against ghosts." The control slot is filled with "organic content, blended the same way real ads are, ensuring relevance and avoiding duplication." (DoorDash Ads, 2025-10-04)

Variants

Predicted Ghost Ads

A recall-optimized machine-learning model is trained on the bid and impression data from the treatment group. The model is then applied to the control group to identify "would-be-impressions" — users who would have received the ad based on their auction behavior — without requiring the platform to actually participate in the auction on their behalf. This variant is "compatible with online display advertising platforms" where direct auction-level ghost ad infrastructure does not exist. The founding paper's implementation recorded "more than 100 million predicted ghost ads per day" (as-of 2017). (Johnson, Lewis & Nubbemeyer, JMR 2017; AMA/dataxu, 2019)

Ghost Bids

Developed by Remerge for retargeting contexts where Ghost Ads are structurally problematic. A bid is placed into the auction on behalf of the control-group user, but is deliberately designed to lose. The losing bid is logged as the counterfactual. Users not visible on RTB exchanges at all are excluded from both groups entirely. Unlike Ghost Ads (which require a competing advertiser to win the slot for the control user), Ghost Bids work in narrow retargeting audiences where only one advertiser is typically interested. (Remerge, 2019; Viant, updated 2026-01-12)

Platform implementations

Meta Conversion Lift

Meta randomizes users at the Facebook/Instagram account level into test and control groups. When a control user would have won the auction, the second-place ad is served instead. Conversions are measured via Pixel, Conversions API, or offline event uploads. Minimum 7-day test period; at least 10% of audience required in each test cell. (Haus.io, undated/2026-sourced)

Measured.com describes Meta's Conversion Lift as a holdout design rather than a pure ghost ad: "the control group is withheld from receiving that campaign" — but confirms the auction mechanism is intact. (Measured.com, updated 2025-11-12)

Google Conversion Lift

Google "may use a ghost ads approach or a geo approach in the background to run the study and report back results to the brand." The ghost ads framework identifies audiences who matched campaign criteria but were not served an ad "because of other constraints, like budgets and competitive auctions." (Measured.com FAQ, 2025-08-15)

DoorDash Sponsored Listings

DoorDash implemented Ghost Ads as its standard incrementality measurement for Sponsored Listings, documented publicly in October 2025. Results from DoorDash's implementation (as-of 2025-10-04):

• Reduced experimentation dilution by 92% vs ITT design
• Improved iROAS confidence intervals by 35%

DoorDash case-study results (as-of 2025-10-04):

• Heritage Kellanova: incremental ROAS 2.4×–2.9× across three studies
• MALK: 2.7× iROAS and 48% sales lift

(DoorDash Ads, 2025-10-04)

RTB House

RTB House's proprietary ghost ads implementation uses an 80/20 split (80% treatment, 20% control) rather than 50/50, minimising opportunity loss while preserving statistical power. The test runs for at least 21–35 days to compensate. Control users see ads at "limited frequency, typically around one impression every eight days." (RTB House, updated 2026-01-08)

IAB formalization (November 2025)

The IAB and IAB Europe jointly published Guidelines for Incremental Measurement in Commerce Media on November 3, 2025 (updated January 22, 2026), developed by the IAB Commerce Board, IAB's Task Force on Incrementality, and IAB Europe's Retail Media Committee. This is the first cross-industry standards document to formally name and classify Ghost Ads. (IAB/IAB Europe, 2025-11-03)

Ghost Ads are categorized within the "Experiment-based" tier and assigned Strong causal strength — the highest rating in the framework.

IAB's four-method causal hierarchy (as-of November 2025):

Method	Causal strength	Holistic scope
Experiment-based (RCTs, holdouts/Ghost Ads, matched markets)	Strong	Low
Model-based counterfactual (synthetic control, ML propensity)	Strong to Moderate	Medium
Econometric (Media Mix Modeling (MMM), time-series)	Moderate to Weak	High
Hybrid proxies (new-to-brand %, platform-reported lift, Multi-Touch Attribution (MTA))	Weak	Low

IAB defines incrementality as "the additional business outcomes directly driven by a campaign or tactic, compared to what would have occurred in the absence of marketing activity." (IAB/IAB Europe, 2025-11-03)

IAB's named weakness of Ghost Ads/experiments: "costly, prone to data contamination without proper controls, time-intensive" and "usually confined to one platform unless usage of multi-platform holdouts." (IAB/IAB Europe, 2025-11-03)

Benchmarks (as-of 2017 study data)

• Retargeting ads in founding study: +17.2% website visits, +10.5% purchases (Johnson, Lewis & Nubbemeyer, JMR 2017)
• DoorDash: 92% dilution reduction, 35% tighter iROAS confidence intervals (as-of 2025-10-04)
• Yahoo Research double-blind study (insurance sector): last-touch attribution undervalued programmatic by 87% vs ghost-ads-style measurement (as-of 2021 WWW Conference paper)
• Industry adoption (as-of 2025): ~52% of US brand/agency marketers use incrementality testing; 71% of retail media advertisers rank it as top KPI (via Dataslayer, ANA finding)

What practitioners report

Ghost Ads vs geo holdout: Practitioners (Haus, Measured, Northbeam) primarily use geo/market holdout testing as their primary methodology, treating ghost ads as one input. Measured.com recommends independent geo tests as the "gold standard" over on-platform lift studies. RTB House positions ghost ads as superior to geo testing for user-level and segment-level granularity, framing geo testing as "slow, expensive, only works on large groups, prone to interference from market noise." (RTB House, updated 2026-01-08; Measured.com, 2025-11-12)

Retail media adoption: DoorDash, Walmart, Ahold Delhaize, Instacart, and Albertsons Media Collective are all participants in the IAB incrementality task force, suggesting ghost ads are being adopted across retail media networks beyond the social-ad context where the methodology originated. (IAB/IAB Europe, 2025-11-03)

The Haus Meta Report (2025): Analysis of 640 Meta incrementality experiments shows that platform attribution significantly overstates effectiveness. Example: if Meta reports 500 conversions but the experiment shows 300 incremental, the Incrementality Factor (IF) = 0.6 — meaning only 60% of Meta's reported conversions were truly incremental. "For omnichannel brands, 32% of the channel's impact went to non-DTC sales." (Haus.io, 2025)

Limitations

Platform dependency: Ghost Ads require the ad platform to implement auction-log infrastructure. Advertisers cannot run them independently. Cross-platform ghost ads (same user across Meta and Google simultaneously) are technically very difficult. (IAB/IAB Europe, 2025-11-03; Tinuiti, 2025-12-15)

Retargeting incompatibility (original design): The original methodology requires a competing advertiser to win the control-group user's impression slot. In retargeting — which targets narrow, specific audiences only one advertiser typically wants — this condition is rarely met. Ghost Bids (Remerge) and RTB House's proprietary design were developed to solve this. (Remerge, 2019)

Multi-format auction complexity: When multiple ad formats compete in the same auction, isolating the specific ghost ad counterfactual becomes harder. (Johnson, Lewis & Nubbemeyer, JMR 2017; Tinuiti, 2025-12-15)

Opportunity cost: Holding 10–50% of the audience in a control group means forgoing potential conversions from that group during the test period. (Haus.io, 2026)

Volume requirements: Small audiences, niche products, or low-conversion-rate businesses may require 30+ day test windows or larger holdouts to achieve statistical significance. (Haus.io, 2026)

Privacy degradation: On-platform lift studies (Meta Conversion Lift) face accuracy problems in a privacy-first environment: low event match quality from iOS tracking limits and cookie deprecation reduces match rates between platform and advertiser data. (Measured.com, 2025-11-12)

Platform incentive conflict: Platforms with ad revenue incentives manage their own lift measurement infrastructure. Independent measurement vendors (Measured, Haus) position on-platform lift studies as insufficient as a "single source of truth." (Measured.com, 2025-11-12)

Key terms

Term	Meaning
Ghost impression	The logged record that a control-group user would have received an ad
ITT (Intent-to-Treat)	Holdout design that removes control users from the auction entirely; dilutes signal
PSA test	Control group sees a public service announcement ad; costly and introduces targeting bias
Ghost Bids	Bid placed on behalf of control user but designed to lose; logged as counterfactual (Remerge)
Predicted Ghost Ads	ML-based variant that simulates would-be impressions without auction participation
Incrementality Factor (IF)	Platform-reported conversions ÷ true incremental conversions from a lift study
Test cell	One arm of the experiment (treatment or control group)

Next frontier (dangling links from this page)

Predicted Ghost Ads · Ghost Bids · Conversion Lift Studies · Incrementality Factor · Double-Blind Ad Design · Incrementality Bidding · IAB Commerce Media Measurement Guidelines · iROAS