A simple budget-split that adds an explicit 'unknown' gender allocation meaningfully reduces gender skew in Google Ads delivery without shutting out users platforms can't infer; it offers resource-constrained public advertisers a lower-cost compromise between ignoring demographics and expensive granular targeting.

Main Finding

A field case study of a government Google Search advertising campaign found persistent gender-based ad delivery skew (over-delivery to users labeled male). Implementing a novel "unknown-aware" budget split intervention — splitting the original campaign into separate campaigns that target Google-inferred male, female, and unknown users with tailored budget allocations — reduced delivery skew while mitigating the cost increases typically associated with demographic budget-splitting. Including the platform’s "unknown" demographic category avoided excluding users with missing labels (who are non-random and likely to include socioeconomically disadvantaged and non-binary individuals) and provided a practical middle ground between ignoring demographics and expensive granular targeting.

Key Points

• Problem: Algorithmic ad delivery can introduce demographic skew (predicted relevance, competition, and baseline engagement drive this), undermining public-interest outreach.
• Data limitations: Advertisers often must rely on platform-inferred demographics (male, female, unknown) rather than ground-truth sources (voter files) or purchased broker data; "unknown" is common and non-random.
• Intervention: Designed and deployed a budget split that explicitly included an "unknown" campaign in addition to male/female campaigns, allocating budget to enforce population-proportional delivery goals while preserving reach to unknown users.
• Outcome (qualitative): The unknown-aware budget split reduced gender delivery skew and moderated cost increases relative to more granular demographic targeting approaches that exclude unknowns or rely on external data.
• Trade-offs: Using platform-inferred labels raises fairness-privacy and fairness-utility tensions — improving parity may raise costs and involves operating on imperfect/invasive demographic inferences.
• Broader significance: The approach is pragmatic for resource-constrained public-sector advertisers and illustrates how handling missing demographic data changes mitigation design and outcomes.

Data & Methods

• Setting and partnership
  • Collaborated with a state-level government agency promoting a website with small-business resources.
  • Focus platform: Google Search (dominant paid-search market share).
• Timeline
  • Pre-intervention evaluation: April–November 2024 (historical campaigns).
  • Intervention deployed: January–February 2025.
• Metrics used
  • Impressions (primary measure of ad delivery), clicks, conversions (site actions indicating meaningful engagement).
  • CTR (clicks/impressions), CVR (conversions/impressions), CPM (cost per 1,000 impressions), spend attributed to clicks.
  • Demographic attribution via Google-inferred gender labels: male, female, unknown.
• Audit and intervention design
  • Audit: measured delivery shares by Google-inferred gender and compared to target distribution (e.g., population parity).
  • Baseline finding: consistent over-delivery to users labeled male.
  • Intervention: "Unknown-aware budget split" — replaced the single campaign with multiple targeted campaigns (male, female, unknown), allocating budgets to achieve desired delivery shares while allowing the platform’s auction and relevance scoring to operate within each targeted sub-campaign.
• Rationale for unknown inclusion
  • Platform "unknown" users are common on search platforms (higher logged-out share), and missingness is correlated with low online footprint, lower socioeconomic status, and non-binary gender identities. Excluding them risks systematic exclusion of disadvantaged groups.
• Limitations
  • Demographic labels are platform inferences (measurement error, unknown accuracy).
  • Single case study on one platform and one state campaign; quantitative effect sizes are context-dependent.

Implications for AI Economics

• Market structure and platform power
  • Platforms’ inference systems and delivery optimization create externalities: advertisers’ objectives interact with platform auction dynamics and relevance models, yielding unpriced distributional effects (who sees ads).
  • The presence of a sizable "unknown" population (and opaque inference quality) affects bidders’ effective demand and competition across audience segments.
• Pricing and competition
  • Targeting splits campaigns into separate auctions for subgroups; this changes competition and prices (CPM/CPV) across segments. Unknown-aware splitting can mitigate price spikes caused by attempting to force parity across costly subpopulations.
  • Platforms that offer or withhold controls (e.g., ability to target unknown) influence advertisers’ feasible strategies and thus surplus distribution between advertisers and platforms.
• Valuation of demographic data
  • Demographic inferences are economically valuable but noisy. The uncertainty and missingness of such signals reduce the precision of targeting, affecting advertiser returns and the social value of outreach (especially for public-good advertising).
  • Data-broker alternatives are costly and ethically fraught; reliance on platform labels creates a market for demographic inference that has distributional and privacy externalities.
• Welfare and regulatory implications
  • Skew in ad delivery creates allocative harms (inequitable information distribution) that are not internalized by platforms’ objective functions. Economic analysis should account for these social externalities when evaluating platform policies or interventions like VRS (variance reduction).
  • Interventions (platform-side algorithmic mediation vs. advertiser-side budget splits) have different cost, transparency, and control properties. Budget splits give advertisers control but may increase costs; platform mediation may be opaque and costly.
• Design and mechanism implications
  • Auction and relevance mechanisms should be studied under missing attribute information (how auctions allocate and price impressions when demographic labels are missing or noisy).
  • Mechanism design solutions could incorporate parity constraints, priced fairness, or subsidization for public-interest campaigns to internalize distributional goals without imposing undue cost on advertisers.
• Research directions for AI economics
  • Formal models of ad auctions with missing/noisy demographic signals: characterize equilibrium outcomes and cost-fairness trade-offs.
  • Welfare analysis comparing platform-side fairness enforcement (e.g., VRS) versus advertiser-side strategies (budget splits), including cost transfers and impacts on user privacy.
  • Empirical work on how missingness correlates with socioeconomic attributes and how that changes the marginal value of impressions.
  • Policy analysis on mandates for platform transparency (accuracy of inferences) and controls (explicit unknown targeting), and the economic effects of such regulation on small and public-sector advertisers.

Practical takeaways for economists and policymakers: missing demographic information materially affects mitigation strategies and economic outcomes in ad markets; designing auctions and regulation that account for missingness and distributional externalities can better align platform incentives with public-interest outreach goals.