What a Ninety-Five Percent Rating Hides
Rating percentage read literally
A rating percentage is an average of submitted scores, and an average is a summary rather than a record. Read literally, ninety-five per cent means that across every score the platform counted, the typical submission sat close to the top of the scale. That statement is accurate and nearly content-free, because it says nothing about how many submissions there were, what they were attached to, or when they arrived.
The arithmetic behind the figure matters more than the figure itself, because one percentage can rest on two very different kinds of evidence. A page showing ninety-five per cent from four thousand entries and a page showing the same figure from thirty entries look identical in a list. The first rests on a spread that volume has tested; the second can move several points on one unhappy submission. The platform does not print the count beside the percentage in every view, so the count has to be found separately, and where it cannot be found the figure is better treated as a claim than as a measurement.
What the number describes is also narrower than it appears. Scores attach to orders rather than to parts, so a purchase that arrived quickly and matched its photographs leaves the same mark whether the stitching held or not. Categories with visible variation, such as a printed panel or a plated surface, produce more complaints per order than categories with fewer visible parts, and one workshop can therefore carry different figures across two product lines.
What the figure is attached to matters as much as the arithmetic behind it, and dates are the cheapest way to check both. A storefront, a listing, a seller account and a product line are counted differently, and a number cannot be moved from one to another without changing its meaning. Reading the three newest entries and the three oldest, then writing both date ranges beside the figure, answers the volume question, the age question and the composition question at once. Where a page does not say which of the four is being averaged, the honest position is that the figure has no defined subject.
Why the metric misleads
Averages mislead in three specific ways, and each one leaves a visible symptom on the page rather than hiding inside the number. The first is thin volume, where a handful of submissions carries the whole figure and a single entry can shift it by several points. The second is age, where an old spread keeps a current number standing long after the supply behind it changed. The third is composition, where the entries behind the number were attached to products the page no longer lists.
Thin volume shows up as a page whose reviews all sit inside a short window. Two hundred entries dated across six weeks describe one production run rather than a line. Buying from that page is a bet on the run, and a repeat order a season later is a bet on a run nobody has scored yet. The percentage stays where it was while the thing it measures has been replaced.
Age produces the opposite symptom: reviews spread across years with most of the text at the older end. A page whose last twenty entries predate the last two quarters is quoting a version of the product that may not be the version listed now. Platform rules on review visibility differ, so part of this is arithmetic rather than concealment, and the desk records it as a reading rather than an accusation.
Composition is the hardest of the three to see, and none of the three failure modes requires a bad intention. Entries attached to a discontinued line stay on the account, and the average keeps them, so a storefront figure can be built from a product the page has stopped selling. Reading the newest entries and the oldest entries separately, rather than the average of both, is the only way to notice that the composition changed underneath a stable number. Averages decay, samples stay thin and product lines end; each of those is ordinary. The mistake that costs money is treating a figure built for one purpose as though it settled a different question, such as whether the next parcel will match its photographs.
Weighted signal practice
The desk's alternative is a weighted reading, and the weights are published here so a reader can argue with them. Seven signals carry the hundred points set out in the table below. Each one was chosen for two reasons: a reader can check it from public pages without buying anything, and it moves when the condition behind the rating moves instead of staying flat while the supply changes underneath.
Review text carries the largest single share, because a sentence naming a part can be compared against a threshold while a score cannot. A review saying that a strap pulled away at the seam is evidence about a part; a review saying the order was great is evidence about a mood, and the two should not carry equal weight. Photo-bearing entries take a smaller share, since an attached image shows the item that arrived rather than the impression it made. Age of the oldest visible entry matters because a page with no history has no track record to average; it has only a current run.
Recency is weighted separately from total volume, which is the departure from the platform figure. A dispute phrase appearing in the last two review windows counts as heavier evidence than an older average, because the supply behind a page can change within a season and an average cannot. Repeat-buyer wording carries a modest share as the closest public proxy for continuity.
Response pattern closes the set, and applying the weights takes only a few minutes on a page that publishes its entries in order. A storefront that answers a critical entry with a specific account of what went wrong reads differently from one that answers with a template, and both read differently from silence. The signal is not about politeness; it is about whether the public record contains a correction or only a repetition. Score the signals you can see, mark the rest as unread, and treat the total as a reading rather than a grade. Where two signals are missing entirely, no combined figure is recorded, because a sum built from five of seven parts invites more confidence than the evidence carries.
Outcome comparison
Worked side by side, the two readings separate on the cases that matter. Take three storefront pages that each display ninety-five per cent. The first shows four thousand entries dated across four years, a third of them naming a part. The second shows two hundred entries inside a single season. The third shows one hundred and twenty entries, most attached to a line the page has stopped listing.
Under the weighted reading, the first page keeps most of its points: volume, spread and recency all hold. The second loses its history points and keeps the merits of a current run, which is a smaller and more specific claim than the figure suggests. The third loses composition points outright, because the entries behind its figure describe products a buyer cannot order from that page today.
What changes in practice is the question the number is allowed to answer. On the first page it supports a general expectation about a line. On the second it supports a statement about one run. On the third it supports nothing about the listing as it stands. Applied that way, one percentage leads to three different levels of caution, and the difference is visible before any message is sent.
Two limits belong with this method, and the comparison also explains why a lower percentage is sometimes the stronger page. The weights are the desk's own construction rather than a platform formula, and no page here claims to reproduce how the platform computes its figure. The reading is public-information-only: it sorts pages into those with a checkable track record and those without one, and says nothing about what a particular order will deliver. A storefront showing ninety-one per cent across a long spread, with recent entries naming parts, carries more checkable history than one showing ninety-five across a single season. Where the two appear side by side in a list, ordering by percentage reverses ordering by evidence, and that reversal is the practical reason this article exists.
Signals behind a rating percentage, and the weight each carries
Editorial weighting. The desk defined these seven signals after reading 1,120 public review entries drawn from 40 storefront pages, counting how many carried a statement about a named part rather than a score alone. The points are whole-number shares of a hundred and are this desk's construction, not a platform formula; the platform's own arithmetic is not published in full. No entry in the reading was used to evaluate a named seller, and none of the pages read is identified here. Last checked 2026-09-22.
| Item | Reading |
|---|---|
| Entries that name a part | 20 points — a sentence about a part can be compared; a score cannot |
| Review count at the page level | 18 points — a share is unstable below a few hundred entries |
| Dispute wording in the last two windows | 16 points — recent text outweighs an older average |
| Photo-bearing entries | 14 points — an attached image shows the item rather than the mood |
| Age of the oldest visible entry | 12 points — a page with no history has no track record to average |
| Repeat-buyer wording | 10 points — the closest public proxy for continuity of supply |
| Response pattern to critical entries | 10 points — a specific account and a template read differently |
Sources and checks
- Public community thread — Public review entries read across 40 storefront pages; counted by whether they named a partn=1120
- Editor analysis — The seven signals and their weights; constructed by the desk from the reading aboveno sample
Last checked · next review due · status current.