The Anatomy of a Makeup Score

I spent a week with these analyses. Eight or nine hours a day for six days, reading them in batches of fifty, taking notes on a yellow pad next to my laptop, occasionally getting up to make coffee and forgetting whether I had already made coffee. The reason I am telling you this is that the headline finding of this report — that the single biggest problem in modern makeup is application technique, not product choice — is one I did not expect to find, did not particularly want to find, and resisted finding for the first two days because it contradicts the way almost every beauty publication, retailer, influencer, and brand-funded media organization talks about the practice of putting things on your face.

The beauty industry, considered economically, is a product industry. It sells things. It sells primers and foundations and concealers and powders and creams and serums and pigments, and the entire downstream ecosystem of magazines and YouTube channels and TikTok feeds and Sephora reviews exists, in part, to help you decide which products to buy. This is fine. I have no objection to it. But it produces a strong, structural bias in the way the topic of "what makes a makeup look good" gets discussed. The bias is toward nouns. It is toward the things you can buy. It is toward "this Pat McGrath palette" or "that Charlotte Tilbury wand," because those are the units of commerce and therefore the units of conversation.

What the data in this corpus shows, almost without contest, is that the bigger problem is verbs. It is how the products are being applied. It is the blending, the angle of the brush, the order of operations, the restraint or absence of restraint. The product, as far as the model can tell from a photograph, is mostly a downstream issue. The technique is upstream. And nobody is selling technique, which is, I think, why nobody is talking about it.

This report walks through the evidence for that claim. I tried to keep it honest. The data is what it is. There are real limitations, which I will be specific about. But by the end of it I hope you will understand why the most useful sentence in beauty advice — the one I have ended up muttering at my screen at four in the morning, alone in a small office in Casablanca — is not "buy a better foundation." It is "blend the foundation you already have for thirty more seconds."

Section 1What This Report Is, And What It Is Not

The methodology underlying this analysis is identical, in its main contours, to the one described in Report 01: photographs uploaded by users are submitted to a multimodal vision-language model under a fixed scoring rubric, the model returns a JSON object containing a top-line score (0–100), a verbal rating, three to five key strengths, two to three improvement recommendations, and a verbal assessment. Temperature is set around 0.4 to balance evaluation stability against the natural diversity of real faces. The makeup rubric and the outfit rubric share architecture but evaluate different surfaces: the makeup rubric weighs skin finish, eye look, lip work, brow grooming, and overall color cohesion, where the outfit rubric weighs fit, color, occasion, styling, and details.

The 917 makeup analyses in this report were submitted between roughly February and May 2026. A small but useful detail: the corpus was split into two extracts during data preparation, separated by whether the model produced a full visual inventory of the photograph (the "image description" field). One hundred sixteen analyses (12.6% of the total) contained a rich visual inventory; the remaining 801 did not. The reason is, mostly, prompt-version drift on my end. Earlier in the year the makeup pipeline emitted image descriptions only inconsistently. I have since regularized this, but the corpus for this report was assembled from analyses spanning the older inconsistent regime. Where image descriptions are available, I use them. Where they are not, I rely on the structured strength and improvement fields, which were always present.

I want to be specific about who is in this dataset, because demographic shape matters. Of the 917 submissions, 844 (92.0%) returned a gender inference of "female," 31 (3.4%) returned "male," and 42 (4.6%) returned unspecified, unknown, or null. Every single submission in this corpus came from an unauthenticated user — that is, a guest, not a logged-in account holder. This is in part because the makeup analysis flow on OutfitScore has been free and unauthenticated for most of its history, and in part because the kind of person who tries an AI makeup evaluator tends to do so impulsively, before any sign-up flow can catch them. Both of these facts shape the corpus.

I do not have repeated-submitter identification. I cannot tell you, from this data, whether the same user submitted twenty photographs or whether 917 distinct people submitted one each. My intuition, from staring at the image descriptions where they exist, is that the corpus represents perhaps 600 to 800 distinct individuals, with a long-tailed minority of repeat submitters skewing the count modestly upward. This is a hypothesis I cannot prove with the available data and that I will revisit in a future longitudinal study.

The corpus is, like the outfit corpus before it, voluntary and self-selecting. The people who upload photographs of their faces to an AI evaluator are not a random sample of the world's makeup wearers. They skew young (visually, by inspection of the image descriptions in the rich-metadata subset), they skew urban, they skew comfortable with their faces being processed by software, and they skew, by an enormous margin, female. Generalizing from "what scored well in this corpus" to "what is universally a good makeup look" requires every caveat I can think to attach. I will attach them where they matter.

Section 2The Score Landscape, Charted Honestly

The makeup corpus scores higher, on average, than the outfit corpus. The mean is 66.1 (vs 62.8 for outfits), the median is 70 (vs 64), and the standard deviation is 16.8 (vs 14.2). At face value this is surprising, since you might expect makeup, with its smaller surface area and tighter set of variables, to be a less forgiving domain than the entire body. Looking at the distribution carefully, the higher mean is real, but it is also misleading.

The higher mean is partly an artifact of the rubric. Makeup evaluations have a smaller surface area than full-body outfit evaluations and therefore fewer independent failure points; a well-applied lip and a clean base can carry a look that, if the rubric measured posture or proportion, would have additional failure modes. There is also a survivorship effect. The people who submit makeup photographs to an AI tool are, on average, people who have already invested fifteen to forty-five minutes that morning in producing a look they consider photographable. The pre-submission filter is real. The model is not scoring random faces in the wild; it is scoring faces whose owners thought were worth scoring.

Both factors are real, and both should make us suspicious of any direct comparison between the makeup mean and the outfit mean. The comparison that is honest, and that this report leans on, is between high-scoring and low-scoring submissions within the makeup corpus. The corpus is internally consistent, the rubric is fixed, and the differences between the top tier and the bottom tier carry information regardless of the absolute level of the mean.

One more thing about the distribution. The cluster of submissions scoring zero through nine — twenty-three analyses, plus a smaller number of low scores in the teens and twenties — is almost entirely composed of what the engine calls "no makeup detected." That is to say: the user submitted a photograph, the model looked at it, decided that no cosmetic intervention was actually visible, and returned a low score by default. Whether this is the right behavior for the engine is a question I will revisit. For now, the relevant fact is that approximately one in seven submissions to a tool that exists to evaluate makeup contains no detectable makeup, and that this fact is not an error in the data. It is, as far as I can tell, the most interesting psychological finding in the entire corpus, and I am going to give it its own report.

Section 3The Five Things That Are Almost Always Wrong

The recommendation field returned 2,352 individual improvement tips across the 917 analyses, averaging just over 2.5 tips per submission. Of these, 1,868 were lexically unique strings. The lexical uniqueness is misleading, in the same way it was misleading in the outfit corpus: many of these tips are saying the same thing in different ways. The phrase "consider more precise blending at the contour" and the phrase "blend the contour more carefully" are, for any practical purpose, the same recommendation.

To recover the underlying themes, I built a keyword taxonomy by hand. I started with a list of obvious candidate themes derived from the highest-frequency exact-text recommendations, then iterated. I read perhaps 400 of the 2,352 tips during this process, which is enough to be confident in the structure of the top five themes and somewhat less confident in the long tail. Where a single tip touches multiple themes — "blend the contour more carefully and warm up the lip color" touches both application and lips — it was counted toward each. The taxonomy is documented; readers who want the precise keyword lists can email me.

Before walking through the five top themes one at a time, I want to note something about the shape of the chart. There is no smooth decay from the top theme to the bottom. There is a single dominant category — application — followed by a tight cluster of four product categories (lips, brows, foundation, blush) at roughly half its size, followed by a long tail. This is unusual. In the outfit corpus the top theme (fit and proportion) was also dominant, but only by a small margin over the next two. In the makeup corpus, application sits a clear head and shoulders above everything else, and it is not a product category at all. It is a meta-category that spans every product category. It is, in some sense, the substrate on which all other recommendations are written.

1. Application Technique — 672 tips · 28.6%

The phrases that recur in this cluster, in roughly descending order of frequency, are: blending, blend, technique, application, seamless, uneven, precision, crease, smoother transition, more deliberate placement. The recommendations themselves divide into two rough sub-clusters.

The first sub-cluster is what I would call boundary control: tips about where one application begins and another ends. "Blend the foundation more carefully into the jaw to avoid a visible line." "Soften the edge of the blush at the temple." "Diffuse the eyeshadow at the outer corner so it does not look stamped." The visual problem these tips describe is a real one and well-documented: the human eye is unusually sensitive to discontinuity in skin tone, and a sharp edge between two pigmented surfaces reads as artificial in a way that beauty editors and high-end makeup artists spend most of their careers learning to suppress. The makeup artist Bobbi Brown has written extensively about this; her general thesis, in Bobbi Brown Makeup Manual (2008) and elsewhere, is that the difference between professional-looking makeup and amateur-looking makeup is almost entirely a matter of how the boundaries between products are handled. The data in this corpus supports her thesis with depressing precision.

The second sub-cluster is what I would call order of operations: tips about the sequence in which products were applied. "The mascara appears to have been applied before the foundation set, leaving flakes on the cheek." "The contour appears to have been applied over a still-wet base, producing a muddy transition." "The lip liner appears to have been applied after the lipstick rather than before, leaving a visible double edge." The order in which makeup is applied affects how the products interact, how they dry, and how they sit on the face. Many of the tips in this sub-cluster are not really about the products themselves but about the implicit choreography of getting them onto the face in the right order with the right drying time between layers.

What unites both sub-clusters is that none of the recommendations in this category can be solved by purchasing a different product. They are all recommendations about the relationship between the user's hand and the user's face. They are the cosmetics-industry equivalent of being told that the reason your golf swing is bad is not because of the clubs.

2. Lips — 411 tips · 17.5%

The lip cluster is more varied than the application cluster but contains a few clear sub-themes. The most frequent specific recommendations involve color choice ("the lip color competes with rather than complements the eye look"), finish ("a slightly more pigmented matte finish would balance the otherwise dewy face"), and boundary work ("define the lip line to prevent feathering into surrounding skin"). The lip is one of the few features in makeup where the recommendations span both product choice and technique with roughly equal weight, which makes lip tips somewhat more useful, in practical terms, than tips in the application category, because some of them can be solved by buying a different product.

What is striking, on inspection of the high-scoring subset (more on this shortly), is that the lip approach in top-scoring looks is overwhelmingly understated. The phrase "subtle lip color" appears 18 times among strengths and "appropriate lip color" appears 10 times. The phrase "bold lip" appears mostly in the improvement column, often paired with notes about it overwhelming the rest of the face. The data, in other words, is rewarding restraint at the lip — which is consistent with what the experimental psychology literature, particularly the work of Etcoff, Stock, Haley, Vickery, and House (PLoS ONE, 2011), found about the perception of makeup: looks that were subtle received higher ratings on competence and trustworthiness than looks that were dramatic, even when both were rated similarly attractive.

I want to be careful about over-reading the Etcoff finding. The 2011 paper had a relatively small sample (149 raters, 25 models) and a specific experimental setup. But the directional finding — that subtler makeup reads as more competent — has been replicated enough times in subsequent literature, including in Mulhern, Fieldman, Hussey, Lévêque, and Pineau's earlier work (International Journal of Cosmetic Science, 2003), to be treated as relatively robust. The corpus in this report is consistent with the pattern. The high-scoring lip is, almost always, a quieter lip than the beauty marketing apparatus would have you believe.

3. Brows — 334 tips · 14.2%

Brows are, by my reading, the single most interesting category in this corpus, because they appear with high frequency in both the strength column and the improvement column. The exact phrases "well-groomed eyebrows" and its close variants account for more than 150 distinct strength mentions across the corpus, making brow grooming the second most frequently observed strength after "even skin tone." At the same time, brow-related recommendations appear in 14.2% of all improvement tips, making them the third most common deficiency category. The same feature is simultaneously the most reliable source of strength and the third most common source of weakness, which is, when you think about it, a remarkable property.

The explanation, I believe, is that brows are asymmetric in their visibility. A well-groomed brow is barely noticed. The eye reads the face as harmonious and moves on. A badly-groomed brow, on the other hand, is impossible to miss. The asymmetry exists because the brow is one of the few facial features whose visual function is structural: it frames the eye, anchors the upper face, and in a literal evolutionary sense communicates emotion via raise and angle. The 2003 paper "The Role of Eyebrows in Face Recognition" by Sadr, Jarudi, and Sinha at MIT made the structural argument with eye-tracking evidence; people look at the brow before almost any other feature when assessing a face.

What this means for the data is that brow recommendations in the corpus split into two sub-clusters with very different practical implications. The first sub-cluster is about shape: brows that are over-plucked, under-defined, or assembled into an angle that does not match the underlying bone structure. These recommendations are difficult to act on quickly because brow hair grows back slowly (an over-plucked brow takes weeks to recover and may not recover at all). The second sub-cluster is about density and fill: brows that are visible but sparse, or that have been filled in with a pencil too dark or too unblended to read as natural. These recommendations are immediately actionable. A lighter pencil, applied in feathered upward strokes rather than horizontal lines, resolves the second sub-cluster in most cases without any structural intervention at all.

4. Foundation, Base, and Coverage — 286 tips · 12.2%

The foundation cluster overlaps substantially with the application cluster, since the most common base-related complaint — visible boundary at the jaw — is also a blending issue. But the cluster contains enough distinct recommendations about match (the foundation does not match the underlying skin tone) and finish (the foundation reads too matte for the rest of the look, or too dewy for the occasion) to deserve its own category.

The match problem is interesting because it is, in principle, solvable at the retail level. Every major foundation brand now sells in a wider shade range than was available even five years ago. The improvement on that axis has been real and is well-documented; Allure and other beauty publications have tracked the shade-range expansion of brands like Fenty, Lancôme, and Maybelline since 2017. And yet match-related recommendations still appear in roughly 4% of all tips in this corpus. The reason, I suspect, is that the difficulty of matching foundation is not primarily about availability. It is about the difficulty of self-assessing one's own undertone. The corpus contains a number of analyses where the foundation match would have been perfect if the user had been a half-shade warmer; instead the user appears to have chosen a half-shade cooler, presumably because they read their own face as cooler than it is. This is a known phenomenon in cosmetics retail and is one of the reasons in-store shade-matching with a trained advisor still outperforms online shade-matching tools.

5. Blush — 256 tips · 10.9%

Blush recommendations split roughly evenly between placement ("the blush sits too low on the cheek, dragging the face downward" or "the blush is too high, reading as theatrical") and intensity ("too pigmented for the rest of the look" or "barely visible against the foundation"). What is notable about the blush category is that it punches above its weight in terms of corrective impact. A blush placement issue is a small motor movement of the brush, perhaps two centimeters in either direction. A blush intensity issue is one fewer or one additional pass of the brush. The cost of getting blush right is among the lowest of any single intervention in this corpus, and yet it shows up as a real deficiency in roughly one in nine recommendations.

If I had to guess at the cause, I would say it is that blush is one of the most personally variable cosmetics. The right placement is dependent on the underlying face shape (rounder faces benefit from a higher, more diagonal placement; longer faces benefit from a horizontal placement that visually broadens). The right intensity is dependent on the rest of the look (a dewy base with a heavy lip can carry a stronger blush; a matte base with a quiet lip needs a softer blush). There are no universal rules. Beauty media tends to teach blush as if there were, and the corpus suggests that this teaching is not landing.

Section 4What the High Scorers Are Doing Right

I extracted the 144 analyses scoring 80 or above (15.7% of the corpus) and re-read the strengths and improvement fields for each. The contrast with the bottom tier — the 65 analyses scoring below 40, or 7.1% of the corpus — is consistent and, I think, important.

Top-tier strengths cluster around three traits, in roughly this order:

The phrase that recurs in this list, often implicitly, is restraint. "Natural skin texture preservation" is restraint: choosing not to over-conceal, not to over-mattify, not to over-erase the actual lived-in surface of the face. "Appropriate intensity for daily wear" is restraint: choosing a level of pigmentation that does not require the rest of the look to bend toward it. "Subtle lip color" is restraint. "Balanced facial symmetry preserved" is, again, a recognition that good makeup does not impose a new face onto the underlying one.

This is consistent with a long literature in beauty studies. The cultural critic Jia Tolentino, in her 2019 essay "Always Be Optimizing" (collected in Trick Mirror), made the argument that the dominant aesthetic of the 2010s was the cosmetically constructed "Instagram face": cheekbone sharpened, lip plumped, skin filtered, eye lifted. The look she described was, in scoring terms, an over-extended one — every available variable pushed toward its visual maximum. The cultural backlash against the Instagram face, which I will discuss at greater length in the next report in this series, has produced an aesthetic — variously called "clean girl," "no-makeup makeup," "soft glam," "natural enhancement," all the phrases I read repeatedly in the corpus — that is in essence the same face, but with every variable dialled back to something closer to neutral. The data suggests that the model rewards this dial-back not because it is the prettiest or the most striking look, but because it is the look that reads as competent in the way Etcoff and colleagues identified.

One more observation from the high-scoring subset. The recommendations that do appear in top-tier looks, when they appear at all, are overwhelmingly about fine-grained adjustments: a slightly warmer lip, a marginally darker brow, a touch more highlighter at the inner corner of the eye. The recommendations in the bottom tier are almost the opposite: basic missing components. No visible base. Brows that have not been touched in months. Blush that has been applied as a stripe rather than as a contour. This produces the same two-phase model I described in the outfit corpus: completion at the bottom (getting all the expected elements onto the face), editing at the top (refining the elements that are already there). The interventions that move a 45 to a 65 are different in kind from the interventions that move a 70 to an 85.

Section 5The Application Question, At Length

I want to spend a section just on the application finding, because I think it is the most important thing in this report and because I want to be specific about what it does and does not imply.

The finding, restated: of 2,352 improvement recommendations across the corpus, 672 — that is, more than one in four — implicate application technique rather than product selection. This is a larger share than any single product category. It is larger than lips, brows, and foundation combined would be if you counted only the recommendations that were purely about those products. The size of the application category is, in my reading, the most consequential single fact in the entire dataset.

What does it imply? It implies, first, that the structural emphasis of the beauty industry on products as the unit of intervention is mismatched to the actual structure of the problem. If you walk into a Sephora and ask what would most improve your makeup, you will be sold a product. If you read a beauty magazine, the advice will be product-shaped. If you watch a TikTok beauty influencer, the video will be about a specific named item. None of this is wrong, exactly. There are real differences between products and the differences matter. But the data in this corpus suggests that for the median user, the marginal value of a product upgrade is substantially smaller than the marginal value of a technique upgrade.

It implies, second, that the cultural conversation around makeup has a structural blind spot in the direction of verbs. We have a vast and well-developed vocabulary for talking about products: brand names, formulations, finishes, coverage levels, undertones. We have a much smaller and less well-developed vocabulary for talking about how to apply products. This is partly because verbs are harder to write about; you cannot photograph a verb the way you can photograph a tube of lipstick. It is partly because there is no obvious business model in selling technique; nobody can patent "blend for thirty more seconds." But the absence of a business model does not change the empirical fact that technique is where most of the gains are.

It implies, third, that the rise of professional makeup tutorial content on YouTube and TikTok — content where you watch someone actually do the application, in real time, with the camera close enough to see the brush strokes — is, in cultural-technological terms, more significant than the rise of any individual product. The tutorial format teaches the verb. The product photograph teaches the noun. The verb is what is missing.

I want to flag what I am not claiming. I am not claiming that products do not matter. They do. A foundation that does not match your undertone will produce a worse result than one that does, however skillfully it is applied. A lipstick whose pigment ages badly over the course of a day will look worse at five p.m. than one whose pigment is stable. I am not claiming that the beauty industry is uniformly wrong to focus on products. I am claiming that the marginal allocation of attention is off. For the median submitter to this corpus, an additional thirty seconds of blending would produce a larger improvement in their score than an additional fifty dollars spent at Sephora. The fifty dollars at Sephora is a familiar move. The thirty seconds of blending is unfamiliar and the data suggests it is rare.

Section 6The Gender Gap, Briefly

Mean score for female submissions: 67.3 (n = 841). Mean score for male submissions: 58.1 (n = 31). The gap is 9.2 points, which is more than double the gender gap I documented in the outfit corpus, and which deserves at least a paragraph of comment.

The most charitable reading of the gap, and I think also the correct one, is that the makeup rubric was developed against a primarily female reference distribution. The model has been trained on, and the rubric was tuned against, what looks "good" in a context where the unmarked subject is female. When a male face is submitted, the rubric does not know what to do with it, and the closest it can get to a coherent verdict is to treat the makeup as if it were on a female face and score it against that reference. Since male makeup — when it is worn at all — typically operates under different visual conventions (no visible lip color, no eye shadow, a deliberately undetectable base, lighter or absent brows), the model penalizes male submissions for failing to meet criteria that were not designed for them.

This is, I want to be specific, a real limitation of the rubric, not a finding about male makeup. It is, in software terms, a bug. I have begun work on a gender-aware variant of the makeup rubric that will produce separate scoring trees for male, female, and unspecified subjects. The 9.2-point gap is, in my reading, mostly an artifact of the current rubric's design, and not a fact about the world. I will report on this in a future update once the gender-aware variant has been deployed and a comparable corpus has accumulated.

Section 7A Hierarchy of Fixes

What follows is my best attempt at a ranked, evidence-backed list of interventions, ordered by impact-per-effort, that the data in this corpus supports. The ranking is mine, and it is judgment-based; I have weighted the recommendation frequency against my heuristic estimate of effort cost, the way I did for the outfit hierarchy in Report 01. As with any such ranking, the order is approximate. The point is the rough shape, not the exact ordinal positions.

Tier 1 — Blend the foundation jaw boundary

The single highest-leverage intervention in the entire corpus. The visible-line-at-the-jaw recommendation appears in roughly one in six analyses, and the fix costs nothing in money and approximately thirty seconds in time. If you read this report and act on exactly one thing, this is the thing. Take whatever sponge or brush you currently use, and spend an additional thirty seconds working the boundary between your jawline and your neck after you finish your foundation. Watch yourself in a mirror under natural light. The line should disappear.

Tier 2 — Address the brow density problem

If your brows are sparse or under-defined, a lighter brow pencil applied in short feathered upward strokes — rather than a darker pencil applied in horizontal sweeps — resolves most of the brow density issues in the corpus. The "well-groomed brow" is one of the three most reliable strengths in top-tier looks, and the cost of getting there is, for most users, a one-time purchase of a single product. If your brows are over-plucked, the intervention is slower and partly biological; growth serums and patience are the available options.

Tier 3 — Dial the lip color back one notch

The data is consistent: subtler lips score higher. If you currently wear a deep matte lip and your scores are plateauing, try a half-shade lighter or a more sheer finish. The product cost is zero; you almost certainly already own something one notch quieter than your default. The cost is purely cognitive: deciding to wear less of what you already have.

Tier 4 — Re-shade-match the foundation

If you have not re-shade-matched your foundation in the last twelve months, your skin tone has almost certainly drifted (seasonally, with sun exposure, with age) and the match you bought is no longer right. A trip to a counter that does in-person shade-matching, ideally with a trained advisor, is one of the highest-impact interventions available, and is well-documented in the cosmetics-retail literature. Do this once a year.

Tier 5 — Dress for the actual face you have

The most cognitively demanding intervention, and the one that is hardest to write about, is to apply makeup to the face that is in the mirror rather than to the face you are trying to construct. Bobbi Brown has been writing about this since the late 1990s. Pat McGrath has been talking about it in interviews for almost as long. The principle is the same in both: makeup is at its best when it enhances the underlying face, and it is at its worst when it tries to overwrite the underlying face. Most of the looks in the bottom tier of this corpus are bottom-tier because of an attempt to overwrite. Most of the looks in the top tier are top-tier because of an attempt to enhance. The difference is psychological more than technical, and it is the same difference, in essence, that I identified in Report 02: the difference between satisficing and maximizing.

Section 8What This Report Cannot Tell You

I want to close, as in the previous reports in this series, with a careful accounting of the limits of what this corpus can support.

First, the corpus is voluntary and self-selecting. Users who upload photographs of their faces to an AI evaluator are not a random sample of the world's makeup wearers. They skew young, they skew urban, they skew comfortable with software-mediated self-evaluation. They are also, in this particular corpus, all unauthenticated guests, which means I have no way to track whether the same person submitted multiple times. The corpus represents some number of distinct people between, by my rough estimate, 600 and 800. I cannot be more precise than that with the available data.

Second, the scoring engine is a multimodal large language model, not a panel of human judges. The model has been tuned against a rubric I designed, and the rubric reflects my own choices about what to weigh and how. Other rubrics would produce different scores. I am not aware of systematic studies comparing AI makeup evaluation to inter-rater agreement among human makeup artists, and such a study would be a meaningful next step. The closest analogous work I am aware of in the broader beauty literature — Etcoff and colleagues' 2011 PLoS ONE paper, Korichi, Pelle-de-Queral, Gazano, and Aubert's 2008 paper in the Journal of Cosmetic Science on makeup as camouflage versus seduction — used human raters under controlled conditions, not multimodal models, and the rating dimensions were different. The comparison would have to be done carefully.

Third, the corpus skews 96.5% female. Findings about male makeup in this report are based on 31 submissions and should be treated as preliminary at best. The gender-aware rubric I mentioned earlier is in development and will, I hope, produce a more honest evaluation when the corpus has accumulated more male submissions to work with.

Fourth, the thematic clustering of recommendations is keyword-based, developed iteratively against the corpus, and sensitive to the specific keyword choices I made. A different taxonomy would produce different numbers. The top three or four themes are robust to reasonable variations in the taxonomy. The long tail is less robust. The full keyword lists are available on request.

Fifth, the corpus covers a roughly four-month window in early-to-mid 2026. Seasonal effects on makeup choice are real — heavier base in summer for sweat resistance, more matte finishes in winter — and the data here cannot capture them at the annual scale.

And sixth, the corpus does not include any longitudinal information. I cannot tell you, from this data, whether users who score 50 on their first submission improve over time, or whether they plateau, or whether they leave the platform. The analogous questions in the outfit corpus, which I discussed in Report 02, remain unanswered for the makeup corpus and will need a follow-up study with stable user identification to address.

Section 9What This Report Is Trying to Say

If I had to condense this report into a single paragraph, for someone who has neither the time nor the inclination to read the previous twenty-five pages, here is what I would say. The single most common deficiency in modern makeup, by an enormous margin and across nine hundred and seventeen real analyses, is not a deficiency of product. It is a deficiency of technique. The cheapest, highest-leverage intervention available to any reader of this report is to spend an additional thirty seconds blending the foundation they already own at the boundary of their jaw. The second-cheapest is to use a lighter brow pencil with a different stroke. The third-cheapest is to dial back, by one notch, the intensity of whatever they are currently wearing. None of these interventions require buying anything. All of them are documented in this corpus as the most common deficiencies among low-scorers and the most common strengths among high-scorers. The cultural conversation around makeup has, for reasons that have to do with the structure of the cosmetics retail industry rather than with the structure of the cosmetics themselves, organized itself around products and around the purchase decisions that products imply. The data in this corpus suggests that this organization is wrong. The products matter. They matter less than the verbs.

I think the most interesting question this report raises, and the one I will pick up in the next two reports in this series, is not really a question about makeup. It is a question about what happens when an aesthetic culture organizes itself around something other than its actual point of leverage. The cosmetics industry has organized itself around products because products are what you can sell. The data suggests that the actual point of leverage is technique. That gap — between what the industry is incentivized to sell and what would actually most help its customers — is, in my reading, both the substantive finding of this report and the subject of every follow-up I plan to write.