Jump into the Future of Marketing with the AI That Knows Your Audience •

Get access to Elaris
● Audience Intelligence · Validation Study

The psychology behind the audience.

Behavioural data tells you what people did. It can't tell you why — or who's next. This study validates a psychographic layer that answers both: 29 measured personality traits that sharpen how you understand any audience, and measurably improve what you recommend to them.

0
People profiled
0
Audiences psychologically distinct
0
Profile reliability (test–retest r)
0
Lift in recommendation accuracy

Evaluated on 497,161 people · 29 psychometric trait dimensions · 4,510 affinity groups · 4 consumer domains — brands, games, movies & TV, and interests.

01 · The asset

A personality profile for every person — measured, not guessed.

Each person carries 29 psychographic trait scores spanning the dimensions of personality: imagination, achievement, sociability, morality, sensation-seeking, anxiety and more. Scores are produced with Item-Response Theory (IRT) — the same psychometric standard behind clinical and academic personality assessment — so they behave like real measurements, not engagement guesses.

Measurement-grade

IRT-scored traits

29 continuous dimensions per person, scored on a calibrated latent scale via Bayesian (EAP) estimation — with 0.83 mean score reliability (internal precision) and 0.80 test–retest reliability (stability over time).

Multi-dimensional

Genuinely independent

The 29 traits show clean factor structure with zero highly-collinear pairs — each dimension carries its own information.

Rich linkage

Connected to affinities

Each person links to a median of 5 affinity groups (brands, titles, interests), enabling audience-level psychology at scale.

This report answers two commercial questions with that asset. Can psychology sharpen how you understand an audience — beyond what demographics already tell you? And can it improve what you recommend to them — beyond what behaviour already predicts? First, the foundation both answers rest on: proving the signal is real.

02 · The foundation

Audiences occupy tight, distinct regions of psychological space.

If psychographics were noise, an audience's members would scatter randomly. They don't. Members of the same group cluster together far more tightly than chance, re-measuring the same audience returns the same profile — and the pattern repeats, independently, across every domain we tested.

99.3%

of audiences are more psychologically concentrated than a random group of the same size (statistically significant after multiple-testing correction).

3.0×

tighter than random: the typical audience clusters three times more closely around its own psychological centre than a like-sized random sample.

0.92

median test–retest reliability. Split an audience in half at random and both halves return essentially the same trait profile.

Reliability is exceptional across every content domain
Split-half profile reliability (r) — measure the audience twice, get the same answer. 1.0 = perfect.

Reliability is the foundation of any targeting product: a signal you can't reproduce can't be sold. These profiles reproduce at r ≈ 0.98–0.99 across brands, games, movies & TV, and interests.

And this isn't a one-category trick — the validation was re-run independently on four domains. Pick one:

Entities analysed
232

distinct brands profiled at audience scale

Psychologically distinct
99%

pass a strict separability test vs. a random null

Split-half reliability (r)
0.985

measure the audience twice, get the same profile

03 · Use case: Audience insight

The brief you couldn't write from demographics.

Group entities by personality alone — no genre labels, no category tags — and recognisable human archetypes fall out. That's the insight product: a psychological identity for every audience that's vivid enough to write creative against, and provably not repackaged age-and-gender.

The Achiever
Traits
Achievement Duty & obligation Imagination
Audiences
New BalanceReebokLevi'sColumbiaYETI
The Sensation-Seeker
Traits
Imagination Sensation-seeking Forgiveness
Audiences
PlayStationXboxVansCall of Duty
The Connector
Traits
Sociability Assertiveness Anxiety
Audiences
The GodfatherGoodfellasSuitsForrest Gump
The Harmoniser
Traits
Morality Harmoniousness Sensation-seeking
Audiences
The OfficeBridgertonStranger ThingsGardenscapes

These clusters were built blind — yet scored against the true category labels they never saw, they recover real structure with strong agreement (casual-puzzle game audiences, for example, land in one psychological cluster with 69% purity). The traits aren't just statistically separable; they're semantically meaningful — the difference between a number and an insight you can plan and brief around.

The hardest test: is any of this new information? We trained three models to predict audience membership — demographics only, psychographics only, and both combined. Psychographics alone out-predict demographics, and stacking them on top lifts accuracy again.

Predicting audience membership — model accuracy (AUC)
0.50 = coin flip · higher is better. Psychographics alone beat demographics alone, and the combined model is best.
Demographics only Psychographics only Both combined
0.72

Psychographics-only accuracy vs. a 0.50 shuffled-data baseline — the signal is unambiguously real.

98.4%

of audiences gain meaningful predictive signal from psychographics over a random null.

+15%

relative accuracy lift from adding psychographics on top of a demographics-only model (0.62 → 0.76).

Method: L2-regularised logistic regression, 5-fold cross-validation, with controls matched on age & gender so the comparison is apples-to-apples. Traits are residualised on age, gender and geography — so this is signal demographics genuinely cannot reproduce.

Insight at every scale

Niche & broad both win — differently

Small, niche audiences have the sharpest psychological identity (highest coherence) — ideal for positioning and creative. Large audiences carry the strongest incremental lift over demographics — ideal for targeting. There's a distinct insight proposition at every audience size.

Sub-segmentation that's real

Personas hold up

Within audiences, persona splits explain 15% of psychological variance — confirming that sub-segments capture genuine, not cosmetic, diversity you can activate against.

04 · Use case: Recommendations & discovery

Personalise by personality, and prediction jumps double digits.

Psychology earns its keep twice in a recommender. It finds the next audience before the behaviour exists — and once behaviour does exist, it tells the model whose behaviour to listen to, beating behaviour-only baselines in every domain.

Psychological similarity predicts who people engage with together
Spearman ρ between two entities' trait-profile similarity and how often real people engage with both (behavioural ground truth).
People who are psychologically alike engage with the same things — without us ever observing that behaviour. Trait similarity recovers real co-engagement at ρ up to 0.60, strongest in high-involvement categories like games. That is the core of look-alike expansion and cold-start recommendation: a way to find the next audience before the clicks exist.

And when behaviour is available, psychology makes it work harder. Weighting a recommender's neighbours by psychological similarity to each user beats standard behaviour-only collaborative filtering in every domain — by +11% to +36%.

Recommendation accuracy lift from psychographic personalisation
Improvement in MRR (Mean Reciprocal Rank) of a trait-personalised recommender vs. a strong behaviour-only baseline.
A behaviour-only recommender gives every user the same answer for the same starting point. The psychographic model gives a different, better recommendation to people with different personalities — capturing the audience heterogeneity that population averages wash away. Biggest gains land in discovery-heavy domains: +36% for movies & TV, +27% for games.

Comparison: trait-kernel personalised item-CF vs. standard item-CF, evaluated on held-out next-item prediction across all four domains.