▣ DATA BRIEF · BEAUTY & SOCIAL MEDIA
☀ Key Stats
◆ Pooled analysis of sponsored Instagram beauty posts linked a one-standard-deviation shift toward lighter skin-tone representation to about 5% higher engagement.
◆ When researchers instead compared posts within the same creator, the direction reversed: darker-than-usual skin-tone representation was associated with about 1.8% higher engagement per within-creator standard deviation.
◆ The study analyzed 232,088 sponsored Instagram beauty posts from 1,527 creators during 2024 and 2025.
◆ Between-creator and within-creator estimates pointed in opposite directions on 4 of 11 visual attributes.
◆ The pooled result itself pointed against the within-creator direction on 2 of 11 attributes: skin tone and close-up versus wide framing.
◆ Creative guidance based on the pooled estimates missed an estimated 30.7% of the achievable engagement gain under the study’s magnitude-based validation rule.
◆ Pooled analysis made text overlays appear associated with about a 15% engagement penalty per standard deviation, compared with about 7% for a within-creator standard-deviation change.
◆ Held-out validation used 35,336 posts from 449 creators not used to estimate the original creative rules.
Continue reading ↓
Pooled analysis of sponsored Instagram beauty posts linked lighter skin-tone representation to about 5% higher engagement per standard deviation, but the direction reversed when researchers compared posts within the same creators, according to a new research paper.
The study analyzed 232,088 posts from 1,527 creators and argues that the apparent lighter-skin advantage in pooled data partly reflects differences between creators and their audiences rather than what happens when an individual creator changes the visual representation in a post.
Within a creator’s own content stream, posts featuring darker skin tones relative to that creator’s usual baseline were instead associated with higher engagement.
The finding is observational and does not establish that changing skin-tone representation itself causes engagement to rise or fall.
◎ HEADLINE NUMBER
~5% higher
Pooled engagement associated with a one-standard-deviation shift toward lighter skin-tone representation.
The 5% result changed when creators were compared with themselves
The researchers’ main argument is that an average calculated across many creators can combine two very different patterns.
The first is what they call audience preference: creators with different typical visual styles can attract differently composed audiences and also differ in followers and engagement.
The second is the creative effect: how a creator’s existing audience responds when an individual post looks different from that creator’s usual content.
Those two questions do not necessarily produce the same answer.
POOLED RESULT
~5% higher
Engagement associated with moving one standard deviation toward lighter skin-tone representation across the pooled dataset.
WITHIN-CREATOR RESULT
Direction reversed
Darker-than-usual representation was associated with higher engagement within the same creator’s posts.
The within-creator coefficient corresponds to roughly 1.8% higher engagement per within-creator standard deviation toward darker representation.
Important distinction: The 5% and 1.8% figures use different standard deviations, so they should not be read as directly comparable effect sizes. The notable finding is that the direction reverses.
Creators with lighter baseline representation tended to have larger audiences
The researchers say the cross-creator relationship helps explain why the pooled result points toward lighter representation.
In the study’s panel, creators whose baseline imagery scored toward lighter skin tones tended to have larger followings.
That pattern contributes to a positive relationship between lighter representation and engagement when posts from all creators are combined.
But changing an individual post cannot change the creator’s pre-existing audience composition or the history that produced that audience.
The authors therefore argue that the pooled relationship is not the appropriate statistic for telling an already-selected creator which direction to change a post.
↳ WHY THE DISTINCTION MATTERS
A descriptive pattern could become creative advice
The distinction matters because visual-content analytics are not always used only to describe past performance.
Brands, agencies and platforms can use relationships between image characteristics and engagement to recommend how future content should look.
If a pooled coefficient linking lighter representation with higher engagement were treated as a creative instruction, the analysis could appear to support increasing lighter-skin representation.
The within-creator estimate in this dataset points the other way.
The authors emphasize that the two statistics measure different phenomena rather than one being mathematically incorrect.
The issue is whether a statistic describing differences across creators is being used to answer a question about changing one creator’s content.
Four of 11 visual attributes had opposing signals
Skin tone was not the only visual dimension where the two parts of the analysis disagreed.
The researchers scored each image along 11 visual style axes using CLIP, including color, lighting, composition, framing and representation.
The between-creator and within-creator estimates had opposite signs on 4 of the 11 attributes.
↔ OPPOSING BETWEEN- AND WITHIN-CREATOR SIGNALS
4 of 11
Warm ↔ Cool
Female ↔ Male Model
Light ↔ Dark Skin Tone
Selfie ↔ Posed
For skin tone and close-up versus wide framing, the pooled estimate itself also failed to recover the direction indicated by the within-creator analysis.
On close-up versus wide framing, the pooled coefficient was statistically indistinguishable from zero, while the within-creator analysis indicated a significant shift toward wider framing.
On several other attributes, the pooled and within-creator estimates agreed about direction but differed substantially in magnitude.
↳ CREATIVE ATTRIBUTE · TEXT OVERLAYS
Text overlays looked twice as damaging in pooled data
Text overlays provide an example where the direction did not reverse but the apparent size of the effect changed sharply.
The pooled analysis suggested that text overlays were associated with about a 15% reduction in engagement per standard deviation.
The corresponding within-creator estimate was closer to a 7% reduction per within-creator standard deviation.
~−15%
Pooled text-overlay association per SD
~−7%
Within-creator association per within-creator SD
The authors attribute much of the difference to a broader pattern in which creators who frequently use text overlays also tend to have smaller audiences.
A creative team relying on the pooled estimate could therefore overestimate how much engagement an individual creator might gain simply by reducing text.
The difference mattered on creators the model had not seen
The researchers also tested whether separating the two signals improved creative recommendations on held-out creators.
The main training sample contained 1,068 creators and 161,382 posts, while 459 creators and 70,706 posts were initially reserved for validation.
The final held-out validation panel contained 35,336 posts from 449 creators.
Under the researchers’ magnitude-based rule, the pooled coefficients were associated with a 7.94% engagement gain per standard deviation of the combined direction score.
The rule based on the study’s within-creator estimates was associated with an 11.47% gain.
◎ HELD-OUT VALIDATION
30.7% missed
Estimated share of the achievable engagement gain forgone when the pooled magnitude rule was used instead of the within-creator rule.
A simpler rule based only on whether each visual attribute should increase or decrease produced a smaller but still substantial estimated gap of 18.2%.
A separate off-policy statistical test also ranked the within-creator rule above the pooled rule.
What 30.7% means: Engagement itself did not fall by 30.7%. The number describes the authors’ estimate of the share of possible improvement missed by one prescription rule relative to another.
The study covered creators from 1,000 to more than 1 million followers
The dataset covers sponsored beauty-category posts published during a 24-month period spanning 2024 and 2025.
The median creator had 18,325 followers.
The sample included 524 micro creators with 1,000 to 10,000 followers and 71 mega creators with audiences above 1 million.
Creators contributed an average of 152 sponsored posts, while the median was 162.
Engagement was highly uneven: the median post received 572 likes, compared with an average of 5,999.
StatsJournalist has previously examined the growing role of social-media influencers and the audiences they attract.
✦ Why it matters ✦
The skin-tone result illustrates how a socially consequential pattern in a large dataset can change meaning depending on what is being compared.
Across creators, lighter representation was associated with higher engagement in the pooled model.
Within the same creators, however, the relationship went in the opposite direction.
That does not show that the broader social or cultural forces associated with skin tone are absent.
Instead, it suggests that differences already embedded in creator audiences can become mixed with the effect of an individual creative choice when data from many accounts are pooled together.
This matters if analytics tools take those pooled relationships and turn them into recommendations for brands and creators.
A descriptive association can be statistically valid for explaining differences across accounts while still being inappropriate as an instruction about how the next post should look.
The findings therefore raise a broader question for algorithmic marketing systems: whether the data pattern being optimized actually measures the decision the system is being asked to make.
ⓘ How to read the findings
The paper was posted to arXiv in August 2026 and should be treated as preliminary research. A peer-reviewed journal publication of this specific study has not been verified.
The study uses 232,088 sponsored Instagram beauty posts from 1,527 creators during 2024–2025. The findings should not automatically be generalized to ordinary Instagram posts, other industries, other platforms or different periods.
The headline 5% figure is a pooled association. It does not establish that lighter skin itself causes engagement to increase, nor does it mean Instagram’s platform algorithm explicitly rewards lighter skin by 5%.
The pooled skin-tone coefficient was reported at p < 0.05. The table does not provide an exact p-value, so it would be inappropriate to describe the result as only narrowly or barely statistically significant.
The within-creator skin-tone coefficient was statistically stronger at p < 0.001 and pointed toward darker-than-baseline representation. That result remains observational rather than the result of a randomized experiment.
Interpreting the within-creator estimates causally requires the assumption that residual differences between posts are sufficiently independent of unmeasured factors after accounting for creator effects, other visual attributes and observed controls.
The authors acknowledge that campaign objectives, product types, sponsorship terms or other unmeasured post-level factors could influence both visual decisions and engagement.
The skin-tone variable is a CLIP-derived visual axis created from image embeddings and language prompts. It is not a manually measured biological skin-tone classification.
The researchers manually inspected posts at the extremes of each visual dimension for face validity and report rank correlations above 0.92 when the attributes were scored using a different CLIP encoder.
The pooled 5% and within-creator 1.8% figures are calculated using different standard deviations. Their magnitudes should not be directly compared; the more informative finding is the reversal in direction.
Engagement is measured using likes. The study does not use purchases, conversions, comments, shares, brand lift or long-term follower growth as its primary outcome.
The headline skin-tone result should not be interpreted as a statement about the attractiveness, desirability or inherent value of any skin tone. It describes statistical patterns in one dataset of sponsored beauty content.
The 30.7% figure does not mean engagement fell by 30.7%. It measures how much of the estimated achievable engagement improvement was forgone when the pooled magnitude-based prescription rule was used instead of the within-creator rule in held-out validation.
