Reading a Product Page: What Every Number on It Actually Means
A number you can't interpret is decoration. Here is the anatomy of the page — what each figure is built from, which parts a machine computed, which parts a person wrote, and the one signal most people read as a footnote when it belongs beside the headline.
Read the preview
What does each number on a product page actually stand for? The overall score is a fixed, published blend of six factors, formula quality weighted most. A concern grade is built differently, partly from what people with that concern reported. And confidence is never a weight — thin data lowers what we claim, not the score. Why letters? The Dermington Gazette turns the page.
Three chapters, six factors, one weighted sum
The headline number is not a judgement rendered and then justified. It is arithmetic, and the arithmetic is published.
Six factors feed it, grouped into three chapters. The Formula carries the largest share, and splits into formula quality — the evidence behind the actives, whether they are present at working doses, and how well the thing is built — and safety and tolerability. The Experience covers real-world results drawn from what users reported, plus usability and feel. The Deal covers value against the product's own category, and brand trust.
Formula quality carries the single heaviest weight. Brand trust carries the lightest, deliberately: a company's record matters, and it should never outweigh what is actually in the bottle.
Each factor scores 0–100 in its own right, gets multiplied by its fixed weight, and the six contributions add up. That is the whole calculation. You can check it by hand from the page, and the fact that you can is the point — a score you cannot reconstruct is a verdict wearing a number's clothes.
What this structure buys is a specific kind of disagreement. If you think a product is underrated, the page lets you find out where you disagree with us: the chemistry assessment, the tolerability deductions, the value ratio, or the weighting itself. That is a far more useful argument than trading opinions about an overall number.
The grid of grades answers a different question
Below the headline sits a set of per-concern grades, and these are not slices of the overall score. They are computed separately, from a different formula, and they routinely disagree with it.
A concern grade combines three things. Half of it is formula-for-the-job: how strong the evidence is that these actives do this specific thing, whether the dose looks adequate from ingredient position, and how well the formula is constructed. Three-tenths is fit — whether this kind of formula suits this skin situation at all, which is where an occlusive cream and oily congested skin part company. The remaining fifth is outcomes: what people who said they had that concern actually reported.
That last component is why a grade can move without the formula changing. When enough reviewers with a stated concern accumulate, the grade for that concern draws on their reports specifically rather than on general sentiment — and it says so on the page.
The consequence worth internalising: a product can be excellent for one concern and a poor choice for another, and both grades are correct. CeraVe's Moisturizing Cream is a useful example precisely because it is uncontroversial — a formula built on ceramides and petrolatum that does one job extremely well and was never designed for another. Reading its best grade as its verdict, and buying it for the wrong concern, is the single most common way to be misled by a page that is telling you the truth.
Grades are letters, never a second 0–100 number. That is a deliberate design choice: two numerals on one card invite people to average them, subtract them, or otherwise perform arithmetic that the underlying data does not support.
The confidence band is not a footnote
Beside the score and beside each grade sits a confidence band — High, Moderate, or Early — and it is the most under-read element on the page.
It reports how much of the assessment rests on measured data rather than on neutral priors. A product with a full ingredient list, a resolved pH, a substantial body of reviews sampled recently, and well-characterised actives earns High. A recent launch with a partial list and no collected review text lands at Early. Same scale, same weights, very different epistemic position.
The rule underneath is one of the system's more important commitments: thin data lowers what we claim, never what a product scores. When a leaf can't be measured, it takes a neutral value and the confidence band drops. It is not penalised. Penalising unknowns would systematically punish new products and small brands for being new and small, which is a bias masquerading as rigour.
So a strong grade at Early confidence and the same grade at High confidence are genuinely different findings, and the page shows the band at equal visual weight wherever the score appears — cards, comparison tables, best-of rows — rather than tucking it underneath. A long-established formula and a product launched last season can carry similar grades on entirely different amounts of evidence, and the band is the only thing telling you which is which.
The practical reading: treat High as a measurement, Moderate as a well-supported estimate, and Early as a formulation opinion with the ingredient list behind it and nothing else yet.
What the flags assert, and what they carefully don't
The flags — fragrance-free, alcohol-free, fungal-acne considerations, pregnancy cautions, comedogenic signals — are the part of the page most often over-read.
Each one asserts something narrow and checkable: this ingredient list does or does not contain items matching a defined rule set. That is all. A fragrance-free flag says no fragrance ingredients were identified in the list; it does not say you will tolerate the product, and it certainly does not say the product is gentle overall.
Some flags are deliberately weak, and the page says so rather than hiding it. The comedogenic signal is the clearest case: the underlying data is ingredient-level, animal-derived and concentration-blind, so it is present as a signal, capped at a small deduction, and stamped with its provenance. It is there because ignoring the question entirely would be its own distortion, and weighted lightly because the evidence cannot carry more.
Pregnancy flags work the other way — always displayed regardless of score effect, because a caution someone needs to see should not be buried by arithmetic.
If you want to run the same checks on something we haven't assessed, or on a product you already own, the ingredient decoder applies the same rule sets to any list you paste in, and returns the same flags with the same caveats attached.
Which parts a machine computed, and which a person wrote
This is the distinction most sites blur, and it is worth being explicit about.
Computed: the overall score, every factor and leaf beneath it, all concern grades, the confidence band, the flags, and the value ratio. None of these involve a human decision per product. They are the same rules applied to every product in the catalog, which is what makes cross-product comparison meaningful at all.
Editorial: the review body, the verdict, the who-should-buy and who-should-skip, the texture description, and the context around what the numbers mean. A person wrote those, and they can carry judgement the arithmetic can't — that a formula is technically fine but unpleasant to use, or that a category leader is coasting.
Neither, exactly: the weights themselves. Those are editorial judgement about what should matter, applied mechanically thereafter, and stated as such rather than presented as objective. What is objective lives below them — the measurements, the sources, and the audit trail.
That audit trail is the load-bearing commitment. Every leaf stores where its value came from: an evidence-table row with a version stamp, an ingredient position, a fit-matrix entry, a dated outcomes collection with its sample size. Any figure on the page should unfold into a citable fact within a few clicks, and every fork should show the weight it carried.
One piece of that deserves naming because it is where people reasonably expect trouble. Review text is processed by a language model, and the contract on that model is narrow: it never judges whether a product is good. It transcribes structured facts already present in the text — "I have rosacea" becomes a stated context, "it pills under makeup" becomes an aspect. Classification, not evaluation, auditable against the quotes it drew from.
Reading the page for your own question
Put it together and the page supports three different readings, depending on what you actually came to find out.
"Is this good?" — read the overall score and its factor breakdown, and check the confidence band before you trust the second decimal of anything.
"Is this good for me?" — ignore the overall and read the grade for your concern. This is the reading most people skip and the one that matters most. A saved skin profile makes the page do it for you: your concern's grade leads on every card and listing you see, so you stop comparing products on a number that was never about your skin. A lightweight gel-cream and a rich occlusive can both be strong products while being answers to opposite questions.
"Which of these two?" — the comparison view puts the factor breakdowns side by side, which is where genuine differences surface. Two products with near-identical overall scores frequently got there by completely different routes, and that difference is usually the deciding one.
The honest limitation of all of it: this is a documentary system. It reads ingredient lists, published evidence and what people wrote, and it computes consistently from those. It does not test products on skin, it cannot know your particular reactivity, and it will occasionally be confidently wrong about a formula that behaves unexpectedly in practice. The tolerability signals exist partly to catch that, and they catch it late.
Which is the right note to end on. The numbers are a well-documented starting point for a decision, not the decision. Anyone — us included — who tells you a single figure settles whether a product suits your face is selling the arithmetic harder than it deserves.
Frequently asked
Why does a product have both an overall score and separate grades?
They answer different questions and are computed from different formulas. The overall score asks whether this is a well-made product in general; a concern grade asks whether it suits a specific skin situation, weighting how well the formula does that job, whether the format fits, and what people with that concern reported.
What does the confidence band actually measure?
How much of the assessment rests on measured data rather than neutral placeholders — ingredient list completeness, whether review text has been collected and how recently, and how well characterised the actives are. It never changes the score itself; it tells you how much weight the score can bear.
Why are grades letters instead of numbers?
Because two numerals on one card invite arithmetic the data doesn't support — averaging a concern grade with an overall score, or subtracting one product's grade from another's. Letters communicate a band without implying precision that isn't there.
Does a missing ingredient concentration count against a product?
No. Unmeasurable leaves take a neutral value and lower the confidence band instead. Treating unknowns as negatives would systematically penalise newer products and smaller brands for disclosure practices rather than for formulation, which would be a bias dressed up as rigour.
Is a language model deciding whether products are good?
No, and the boundary is deliberately narrow. A model transcribes structured facts already present in review text — a stated skin context, a mentioned aspect like pilling or stinging — and never evaluates quality. It is a classification layer, auditable against the specific quotes it drew from.
Each shelf lists every product in the catalogue built on the ingredient, graded live against your profile — the shelf stays current when a single pick would not.
Products named in this piece
Independently reviewed, with a verdict per concern — the real products behind this piece.