Drop in a product page. See whether the copy reads as AI, and whether an AI shopping agent can match it.
The real watermark is invisible. Reading it needs the provider's private key, so no public tool can see it, including this one. This checks what you can change instead.
A real text watermark is a hidden signal the model bakes in while it writes, by nudging which words it picks in a secret, key-seeded pattern. The words still read normally. The pattern is the signal, not a reading of your writing style.
On Gemini output, Google adjusts token probabilities at generation time so a detector holding the key can measure the pattern and score, with confidence, that the text was machine-made. It needs length, and fades under heavy rewriting.
Anthropic embeds an imperceptible watermark directly in Claude's text, plus C2PA provenance metadata on images. It proves the text was processed by Claude, not who authored it, and survives copy and paste while degrading with heavy editing.
The EU AI Act, Article 50, requires providers to mark AI output in a machine-readable, detectable format from August 2026. The law says "as far as technically feasible," which is exactly why text watermarks are the weakest of the bunch.
Detecting a real watermark needs the provider's private key. No public tool, including this one, can confirm a Claude or Gemini watermark just by looking at text. Anyone claiming to is guessing.
First, whether your copy reads like generic AI, from its vocabulary and rhythm. Second, and the one that moves revenue, whether an AI shopping agent can match your product, by scoring the page against the Shopify product taxonomy for its category. The catalog is the API of commerce, so the second score is the one to fix first.