How the risk flags work
Updated 2026-08-15 · Risk-term list version 2026-08-15c
The risk flags are a term-by-term assisted check against a versioned, dated, publicly downloadable list. They are not a compliance ruling, not legal advice, and not a stand-in for platform review.
This page covers three things: how matching works (which determines what it misses), where the list comes from (which determines whether to trust it), and how it is updated (which determines whether it goes stale).
If you read one section, read the one on what it does not do.
What it does
- Scans what you wrote: listing copy, voiceover scripts, captions, and the brief you send a creator
- For every match: why it is risky, what to say instead, and where this kind of phrasing usually shows up
- Scans Chinese and English separately, and points to the counterpart term in the other language
- Stamps the list version and date so you know how old the list is
What it does not do
- Does not stand in for platform review and does not predict its outcome — a match or a miss changes nothing about how the platform will rule
- Is not legal advice and does not replace your counsel
- Makes no semantic judgment: a synonym walks straight past it, and no rule engine closes that gap
- Reads text only — not images, not video frames, not audio
- Does not cover every market or category: only shapewear and functional intimates are built out
The third one bears repeating: a synonym walks straight past it. Rewrite "burn fat" as "calorie expenditure" and it stays quiet, while human and model review still read the promise. The list is a prompt to notice your own promise, not a gate you can clear.
How matching works
| Language | Rule | Why | Example |
|---|---|---|---|
| Chinese | Substring | Chinese has no word boundaries, so any position should match | 「立刻瘦」matches inside 「穿上立刻瘦五斤」 |
| English | Word boundaries plus inflection expansion | Raw substring fails in both directions: false positives (cure inside secure) and false negatives (list says "prevents", seller writes "prevention") | `prevents` catches prevent / prevention / preventing / prevented, and `cure` does not hit `secure` |
| Multi-word | Spaces, hyphens and underscores between words | Sellers write vet-approved as often as vet approved | `medical grade` catches medical-grade |
The scanning logic is identical to the product’s: the free checker on this site and the risk card in the exported plan run the same code against the same list, so the site and the product never disagree.
Where the list comes from
- Published platform advertising policies. The pages we opened and verified are listed under Sources on each page, with the "last updated" each page states. Where no official page exists, the page says so — we do not invent URLs or paraphrase rules we have not read.
- Chinese advertising law provisions on absolutes and false claims — the origin of the Chinese listing conventions, and the explanation for why the copy reads the way it does.
- Phrasings we actually found in product copy samples. Every entry is something someone really wrote, not an imagined banned word.
The list has not been reviewed by a lawyer, which section 2 of the draft terms states along with why not yet.
How levels are assigned
| Level | Criterion | Typical example |
|---|---|---|
| High risk | The sentence promises something a garment cannot do, or drags the product into a certification regime, and has no reasonable risk-free reading | “lose inches”, “medical grade”, “clinically proven” |
| Needs review | The same sentence has two readings and one of them is fine; context decides | “instantly” (ships instantly is fine), “a size smaller” (sizing advice is fine) |
There is deliberately no third "low risk" tier. Three tiers invite reading the bottom one as "fine", which is the one judgement we must never make.
Organised by family, not alphabetically
The list is grouped into seven families because fixing one line usually introduces another from the same family. Real rework happens inside a family.
| Family | Holds | Shows up in |
|---|---|---|
| Weight and body-change claims | 9 | Listing title and bullets, The first line of the voiceover |
| Fat-burning, sweat and metabolism | 9 | Listing bullets, Text burned into the main image |
| Detox and body-system claims | 4 | Listing bullets, Replies in the comments |
| Medical, recovery and device claims | 12 | The spec block on the listing, Packaging and hang-tags |
| Endorsement and evidence claims | 4 | The trust block on the listing, What the creator says on camera |
| Immediacy and absolutes | 5 | The first line of the voiceover, On-screen captions |
| The sizing grey area | 2 | Size-chart notes, What the creator says on camera |
| Drug-style claims (treating and repairing) | 0 | Listing bullets, Text burned into the main image |
| Structure or function claims | 0 | The ingredient-explanation block on the listing, The "how it works" section of a creator video |
| Regulatory status and grade illusions | 0 | Listing title, The trust block on the listing |
| Endorsement and evidence | 0 | The trust block on the listing, What the creator says |
| Absolute safety claims | 0 | Listing title, Listing bullets |
| Immediacy and absolutes | 0 | The first line of the voiceover, On-screen captions |
What false positives and negatives look like
| Type | Example | How we handle it |
|---|---|---|
| False positive | “ships instantly” matches “instantly” | Marked needs-review rather than high risk, with the trap spelled out on the term page |
| False positive | “if you are between sizes, take a size smaller” read as an outcome claim | Same — the sizing family is always needs-review |
| False negative | “burn fat” rewritten as “calorie expenditure” | A rule engine cannot close this. Human list expansion or, later, semantic classification. Stated in the terms rather than hidden |
| False negative | Text inside an image; words spoken but not captioned | Out of scope. Text only |
| False negative | The English list does not scan Chinese copy, and vice versa | Cross-language mapping currently gives the counterpart term but does not yet backstop the other language |
False positives are worse than false negatives, because they make a seller delete usable copy and stop trusting the tool. So the bar for certainty is high: anything ambiguous is needs-review, and the call stays with you.
The second category (beauty) has a different status
Shapewear rests mainly on platform advertising policy pages we opened line by line, which is why it has per-term pages, source links and verification dates. Beauty is different: it rests on regulation, and we could not open the official pages from this machine in this pass. Publishing dozens of authoritative-looking regulatory explainers in that state is exactly what this site criticises elsewhere.
So the beauty list is wired into the product (the checker reads it, and a link produces a plan), while the site carries only one status page that concentrates the caveats. The unblock conditions are on that page too.
The update process
- Every change bumps the version (currently 2026-08-15c) and lands in /changelog.
- Every artefact — page, exported plan, risk card, open dataset — stamps the version it was built from.
- A new term must pass two tests: it matches inside a sentence a real seller would write, and a full page of compliant copy produces zero hits.
- No fixed cadence. It moves when policy moves; padding it to look busy helps nobody.
FAQ
Why not use a model for semantic judgement?
This version deliberately calls no model, for explainability and determinism: the same copy always produces the same result, and every hit traces to a line in the list. Semantic classification is the obvious next step, but it introduces a new problem — when a model says "this is risky", we cannot say why or what to write instead, and those two are the actual value here.
How much does the list cover?
Not everything, and it should not pretend to. It covers the most common, most avoidable and most expensive class — the few dozen sentences that keep reappearing in real listings.
Can I contribute terms?
There is no submission path today (the site is static and collects nothing). The dataset is CC BY, so fork it and maintain your own.
Why open the term list but not the angle weights?
Because the moat is not the term list. Publishing it only helps: it gets cited, checked and corrected. Angle weights, hook formulas and the selling-point lexicon are the category know-how, and those stay closed.