Free tool

Scan a live shop for restricted green claims

Give us a URL. We read up to ten pages the way a visitor would, and report every restricted claim with the sentence it came from.

We read up to 10 pages of HTML text, follow your robots.txt, and never execute your scripts.

Market for the exposure estimate Annual turnover (€) Deeper rewrites (1–2 min)

Free, no account. Rate limited to 5 site scans per hour per address.

These tools need JavaScript to talk to the scanner. Everything else on the site — the 92 restricted terms, the example report and the country guides — works without it.

What you get back

A report you could hand to a lawyer

Not a traffic light with nothing behind it. Every finding carries the sentence, the provision and a worked rewrite.

The sentence, not the word

Findings quote the surrounding text and link to the page they came from, so you can judge whether the match is a real claim or an incidental mention.

Severity that reacts to evidence

A claim standing alone and the same claim next to an FSC licence number are not the same risk. The engine reads the sentence for certifications, standards and measured figures.

Country guides

8 national regimes: authority, statute, ceiling, case law.

→

Exposure illustration

Enter a turnover and a market and the report shows a range built from that market's own ceiling — a euro figure in six of the 8 markets, and in Sweden and Finland no euro ceiling at all, only 4% of turnover. The Netherlands reaches 10% for exactly the practices this scanner finds. Clearly labelled as an illustration, not a prediction.

Limits

What the scanner will not do

Being clear about this matters more than looking capable. The scanner reads text. It does not render your JavaScript, open your PDFs, watch your videos, or look at what a photograph depicts.

Pattern matching also cannot tell a claim from a discussion of a claim. This page uses the phrase “carbon neutral” several times; matched on its own, it would light up. A second pass over the findings marks most of that — but it is a judgement made by a model, so the report prints its verdict next to the finding instead of removing anything. We measured the model removing more real claims than false ones, which is why it may only mark.

What it does well is the mechanical half of the job: finding every place a restricted term appears across a site — 7-20 seconds for ten pages, measured on 3 real shops on 2026-08-30 — so a human spends their time on the judgement calls instead of the search.

Common questions

How many pages do you read?

Up to ten. We start at the URL you give us, then follow internal links, preferring paths that look like product, collection, about or sustainability pages — that is where claims usually live. The report lists every page we read and every one we skipped.

Do you execute JavaScript?

No. We read the HTML your server returns. If your product copy is injected client-side, we will not see it, and the report says so rather than pretending the page was clean.

Do you respect robots.txt?

Yes. We fetch your robots.txt first and skip anything disallowed for * or for CosmeticlaimBot. Skipped pages appear in the report with the reason, and the report tells you how many your own robots.txt kept us out of.

Can I scan a site I do not own?

Technically yes — we only read public pages, at the rate a person browsing would. It is the same information any visitor sees. Whether that is appropriate is your call; competitive teardowns are a normal part of compliance work, and 65% precision means you should read the findings before you send them to anyone.

Why did a perfectly reasonable page get flagged?

A lexicon cannot read intent — a glossary, a policy page or a quotation containing "carbon neutral" matches the same pattern as a product claim. A review model runs afterwards and marks the ones it reads as commentary, but it is not allowed to remove them, so the judgement stays with you.

Is anything stored?

Partly, and you should know exactly how. The matching happens on our own server, in memory, and nothing is written to disk. But when a match is found, that one sentence — up to 320 characters, or 400 when we draft a rewrite — is sent to a language model behind a private gateway we operate. It keeps request counts, not request bodies, and nothing is used for training. Text with no match is never sent.

Restricted words

The full lexicon, with the reason and a rewrite for each.

The directive

What changed, when it applies, and which practices are now blacklisted.

Not ready to scan a whole site?

Paste a single product description and see what comes back.

Open the copy checker See an example report