There is an equation at the center of psychological targeting. It has been there since Kosinski, Stillwell, and Graepel published their 2013 PNAS paper showing that Facebook Likes could predict personality traits — sexual orientation at 88% accuracy, political views at 85%, the full OCEAN model of Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism. The equation reads:
β_realized = Δβ_main + β × corr(P̂, T) × E[T]
It describes the realized effect of a psychologically targeted message. The key variable is corr(P̂, T) — the correlation between your inferred personality and the actual personality trait. This is where the whole enterprise lives or dies.
In 2013, Kosinski reported correlations of r = 0.43 for Openness. Square that: 18.5% of variance explained. For a psychological trait measured by self-report, this seemed remarkable. Cambridge Analytica took this number and multiplied it by 230 million American adult profiles, 5,000 data points per person, and built a political persuasion machine that was supposed to have changed the 2016 election.
The machine did not work. The ICO found no evidence it had actually been used to influence elections. Internal emails showed the models had not performed well. Aleksandr Kogan, the academic who harvested the Facebook data, testified that the methodology "had not worked very well." The company collapsed in 2018. But the structure remained.
What remained was not a proof of concept. It was a cost center.
Raphael Perla and colleagues published the first end-to-end meta-analysis of psychological targeting in October 2025, in Psychology & Marketing. They examined 41 studies spanning the full pipeline: inferring personality from digital traces, then deploying personality-matched advertising, then measuring the lift.
The findings were not ambiguous.
For personality inference from digital footprints — the first step — the realized correlation was r ≈ 0.23. That is approximately 5% of variance explained. The optimistic numbers from earlier studies (r = 0.40 to r = 0.56, explaining 16–31% of variance) reflected a specific methodological problem the authors call data leakage: using variables that differ between treatment groups to train models that then predict those same groups on the same dataset. The circularity inflates apparent accuracy without adding real predictive power.
When the same studies were evaluated with proper held-out validation and cross-study controls, the signal collapsed to 5%.
The second step — personality-tailored persuasion — fared no better. In rigorous studies using randomized message assignment and independently measured personality traits, the effect size was r = 0.009. The 95% confidence interval crossed zero. In plain terms: there was no detectable effect.
Combine both steps in the full pipeline, controlling for leakage at every stage, and the end-to-end realized effect is approximately r = 0.002. The authors describe this as "essentially zero."
The paper's core finding is not that psychological targeting is ineffective. It is that the entire academic literature base, which generated the industry confidence that psychological targeting works, was produced by methodological choices that systematically overestimated the effect. When those choices are corrected, the effect disappears.
The ICO investigation in 2018 produced one finding that receives less attention than it deserves: Cambridge Analytica's pitch was internally consistent. The logic was correct. The inference pipeline worked in principle. The failure was not conceptual — it was quantitative. The accuracy of personality inference from available digital traces was not sufficient to drive the claimed outcomes at the scale promised.
This is an important distinction. The failure was not fraud, exactly. The failure was a category error: treating a research result produced under artificial conditions as an operational specification for a commercial system operating at population scale.
Kosinski's 2013 result was obtained from a dataset of 58,000 Facebook users with an average of 170 Likes per user — rich signal, well-behaved modeling. Cambridge Analytica was working with a different data substrate: fewer likes per person, noisier behavioral proxies, a population that had not consented to have their psychological traits inferred. The correlation that worked at r = 0.43 in the research context degraded to something closer to the meta-analytic mean — r ≈ 0.23 — and then degraded further when paired with messaging calibrated to those noisy inferences.
What CA got right was the cost structure. The infrastructure they built — the data pipelines, the psychographic segmentation, the message-to-personality matching engine, the media buying integration — was real and expensive. The per-person data acquisition costs, the modeling labor, the platform fees for activation. That cost was not conditional on the targeting working. It was fixed.
The industry inherited the cost structure without the conditions that made the research result plausible.
The costs of psychological targeting are not tracked as a line item anywhere. There is no "psychographic targeting surcharge" on a media plan. The costs are distributed across the layers of the programmatic supply chain: third-party data licensing fees, DSP platform fees, agency management fees, data provider margins.
The ANA Programmatic Transparency Benchmark study found that in 2023, only 36 cents of every dollar entering a DSP reached consumers as viewable impressions on quality inventory. In 2024, this improved to 43.9 cents. The median effective cost rate — the share of gross spend consumed by intermediary fees before working media — ranges from 20% in well-structured accounts to 40% or more in poorly negotiated ones. For the specific data layer that psychological targeting requires — third-party psychographic and behavioral segments — the documented licensing fees add 2–8% of media spend at the vendor level alone.
This is the tax. It is not theoretical. It is not a margin squeeze. It is a structural charge levied on every impression that passes through the targeting stack, whether the targeting works or not.
And the targeting, by the best available meta-analytic evidence, does not work.
The global data broker market was valued at $296–312 billion in 2024. A significant portion of this market services the advertising ecosystem's appetite for psychological and behavioral targeting data. The WARC estimated the "tech tax" — the sum of intermediary charges — at up to 60% of programmatic spend before fraud. The Guardian documented instances of 50–70% of ad revenue lost to middlemen in programmatic transactions. These figures are not about psychological targeting specifically, but psychological targeting lives inside this cost structure. It is one of the things the tax buys.
The companies still selling psychographic targeting are not fraudulent. Resonate, for example, offers personality-based audience segments built from multiple data sources. Oracle Data Marketplace lists over 30,000 audience attributes per consumer. Acxiom claims 11,000 data attributes per person across 2.5 billion consumers worldwide. Alliant, Tapestry, and dozens of smaller data providers operate in the same space.
What these companies are selling is not deception. They are selling infrastructure. The promise is operationalized: use these segments to target consumers by psychological profile, and the targeting will lift performance. The lift, according to the research consensus, is approximately zero. But the infrastructure persists because the lift is not how the infrastructure is evaluated internally.
The advertiser who purchases psychographic segments evaluates the purchase on the same metrics that drive all digital advertising: CTR, conversion rate, ROAS as reported by the platform. The research showing that the underlying targeting does not work does not appear in these metrics. The metrics measure outcomes — clicks, purchases — not the contribution of the targeting layer to those outcomes. A campaign can achieve its performance targets while the targeting layer contributes nothing, because the outcomes are driven by other factors: price, product quality, existing brand affinity, timing.
The targeting layer is evaluated by its absence in a controlled test, not by its presence in a live campaign. Most advertisers do not run those tests. The vendors who sell the targeting do not run them either.
Here is the structural problem, stated plainly:
The advertising industry pays a continuous tax — in data costs, platform fees, segmentation labor, and middleman margin — to operate a targeting system whose end-to-end effectiveness, by the most rigorous available evidence, is indistinguishable from zero.
This does not mean digital advertising doesn't work. It means psychological targeting specifically — matching message to inferred personality — does not demonstrably improve what would have happened otherwise. The attention economy has built an enormous amount of infrastructure on the assumption that it does.
Perla et al. estimate that prior optimistic findings — the 50% purchase lifts, the 40% click increases — reflect data leakage and evaluation flaws in approximately 90% of published studies. The rigorous studies, the ones that properly isolated the targeting effect, found negligible results. The field has been publishing its own version of the truth for a decade.
The industry continues to pay the tax because the tax is structural. It is embedded in contracts, platform fees, data licensing agreements, and agency scopes of work that were written when the optimistic research findings were still considered valid. Unwinding it would require a coordinated acknowledgment that the targeting layer does not justify its cost — an acknowledgment that no single participant has an individual incentive to make first.
The machine runs. The tax levies. The effect stays parallel to nothing.