In September 2024, Adlook — a cookieless brand growth platform operating across the RTB House Group ecosystem — published research that should have ended a conversation. The study examined a random sample of 151,032 programmatic impressions matched against demographic segments from popular data providers, and then ran a validation survey with 1,335 U.S. online respondents who self-reported their actual demographics against their assigned cookie-based segments.
The results were not close. Of users targeted with the "Women 18–24" segment, 18 percent were actually women between 18 and 24. Forty-three percent were men. Sixty-one percent were older than 24. Thirty-five percent were over 55. The "Parents" segment contained 67 percent of people with no children. The "Moms" segment was 52 percent male.
More structurally: 55.57 percent of all users in the sample were eligible for two or more age groups simultaneously — segments designed to be mutually exclusive. Thirty-five percent were eligible for both the Male and Female segments at the same time. Twenty-eight percent of users under 34 were also eligible for the over-55 segment.
This is not noise. This is the targeting system firing at a target that isn't there, at a rate that makes random selection look precise.
The data powering these segments does not come from declared identity. It comes from inference — probabilistic modeling built on cookies, device graphs, browsing behavior, and data broker records. The industry refers to this as "socio-demographic targeting," and it is the substrate beneath billions in programmatic bid requests.
The foundational claim is that behavioral signals correlate with demographic facts. The evidence for this claim is weak.
A landmark study by Neumann, Tucker, Kaplan, Mislove, and Sapiezynski, published in Management Science (Vol. 70, Issue 11), tested profiling accuracy across fifteen data brokers. Offline profiling — using records like voter registration, property ownership, and credit headers — achieved 84.2 percent accuracy. Online data, the kind powering real-time bidding, achieved between 19.0 and 23.3 percent accuracy. The gap is not a technical problem awaiting a better algorithm. It is a structural property of the data.
The researchers found something further: profiling accuracy is systematically unequal. Higher-income, more affluent consumers have more robust digital footprints — more accounts, more transactions, more behavioral signal — and are profiled more accurately. Blue-collar workers, single households, and non-white individuals have narrower digital footprints and are profiled less accurately. Vendor choice explains virtually none of the variation in accuracy. Who you are determines whether you are profiled correctly, not which data broker is used.
Acxiom, which maintains profiles on approximately 2.5 billion people with over 3,000 data attributes per person, has acknowledged this directly. The company's own language: "our inferences, all they are, are informed guesses."
When Facebook's own researchers studied the platform's ad delivery — Ali, Sapiezynski, Korolova, Mislove, and Rieke (arXiv:1904.02095, 2019) — they found that identical ads with identical neutral targeting parameters delivered to dramatically different demographic audiences depending on budget level. A low-budget version reached over 55 percent male users. A high-budget version reached under 45 percent male. The algorithm was optimizing for engagement and relevance score, which correlates with demographic skew in ways that have nothing to do with advertiser intent.
The business case for precise demographic targeting has always been intuitive: reach the right people, waste less money, convert more efficiently. The empirical case is harder to make.
In 2016, Procter & Gamble — then the world's largest advertising spender at approximately $7.2 billion annually — began a deliberate experiment. The Febreze brand was running targeted Facebook campaigns aimed at "pet owners and households with large families," assuming these segments represented the core odor-problem demographic. Sales stagnated.
P&G's chief marketing officer, Marc Pritchard, described what happened next with unusual candor for a corporate setting: the brand opened its targeting to everyone 18 and older. Sales went up. The implication — that the precision targeting was not finding additional buyers but was somehow preventing them, or that the assumed correlation between pet ownership and Febreze purchase intent was simply wrong — led P&G to publicly reconsider the value of narrow demographic targeting for brand-building objectives.
P&G subsequently reduced its narrow Facebook targeting across multiple brands. When this became public, Facebook's stock fell.
The ad tech industry did not rewrite its targeting philosophy as a result. The intuition was too useful, the alternative — buying contextually against relevant moments, or broadening audiences and letting performance data sort the signal from noise — was operationally less convenient, and the measurement systems rewarded narrow targeting through lower apparent CPMs even when the delivered audiences were incorrect.
Meta developed a response to documented demographic skew in its ad delivery system. After the Department of Justice sued the company over algorithmic discrimination in housing ads — the platform's own delivery algorithm was systematically delivering housing and employment credit ads to racially skewed audiences despite neutral targeting parameters — Meta built what it called the Variance Reduction System.
VRS is a reinforcement learning system that adjusts ad pacing to align delivered demographics with advertiser-targeted demographics. It launched for U.S. housing ads in January 2023 and expanded to employment and credit categories. Meta's stated target: reduce variance to within 10 percent for 91.7 percent of housing ads for gender, and within 10 percent for 81 percent for estimated race and ethnicity.
Academic research studying VRS after deployment found that it reduced demographic variance — it did not eliminate the underlying optimization pressure that caused the variance in the first place. The system trades some delivery efficiency for demographic compliance. It does not resolve the fundamental tension between an algorithm optimized for engagement and a regulatory regime that expects equal delivery across demographic lines.
The DOJ settlement required Meta to deploy VRS. It did not require anyone to explain why the skew occurred in the first place, or to disclose how frequently it occurs in categories beyond housing, employment, and credit.
The Adlook study's 18-percent accuracy figure for "Women 18–24" is not an outlier. It is a point on a distribution.
Truthset's 2026 State of Data Accuracy report estimated that $7.36 billion in CTV advertising is wasted annually due to inaccurate identity and demographic data — not from fraud or invalid traffic, but from the foundational mismatch between who advertisers are trying to reach and who they are actually reaching. IP-to-household matching is accurate 13 percent of the time. Email-to-postal matching is accurate 51 percent of the time. Age-based demographic signals are accurate 45 percent of the time. Households-with-children signals are accurate 41 percent of the time.
The Association of National Advertisers' Programmatic Transparency Benchmark (Q2 2025) estimated $26.8 billion in global programmatic advertising is wasted per quarter — supply-chain waste encompassing made-for-advertising sites, fraud, viewability, and the audience accuracy problem embedded in the targeting infrastructure itself. This is not a niche finding. It is the industry's own measurement of its own failure.
When demographic targeting errors are layered on top of attribution measurement inflation — where platform-reported ROAS exceeds true ROAS by 30 to 50 percent depending on the platform — the misallocation compounds. Advertisers shift budget toward the channels and segments that appear to perform best in distorted measurement, away from the channels that actually created demand, toward retargeting that captures organic conversions rather than induced ones.
A brand spending $500,000 per month in programmatic display, applying the Adlook-era accuracy figures, is likely allocating between $150,000 and $250,000 per month to mis-targeted impressions — impressions that reach the wrong age group, the wrong gender, the wrong household composition, the wrong income bracket — before accounting for any other category of waste.
The targeting system fires. The bullet goes somewhere. The target is rarely where the crosshairs suggested.
The data broker accuracy study by Neumann and colleagues found that behavioral data — what you click, what you buy, what you search for, what you leave unfinished — achieves higher accuracy for some demographic categories than others. Facebook likes alone were 95 percent accurate at distinguishing Black users from white users, and 93 percent accurate at distinguishing men from women, according to the company's own research cited in the academic literature. Personality prediction accuracy, the original promise of psychographic targeting, maxed out around 30 percent.
This is not a contradiction. It is a description of what the models actually learned to predict. The algorithms did not find personality. They found demography wearing personality's clothing.
The advertising industry has invested the better part of two decades in building targeting infrastructure on the premise that behavioral data reveals who people are and what they will buy. The empirical record suggests that it reveals something more particular and less useful: what kinds of ads a person has been likely to respond to in the past, which is a function of what kinds of ads they have seen, which is a function of what kinds of ads have been targeted at demographically similar people, which was determined by the last targeting model's best guess about who those people were.
The system is confident. It fires with great conviction. The 55.57 percent of users eligible for multiple age segments suggests it is also not particularly precise about where the target is.