AI Transparency

Agentible runs four agents against your store. Two of them generate text with AI; two of them only measure. This list is rendered from the same registry the product runs, so it cannot drift from what actually ships:

• AI readiness (Ready to be bought): Whether an AI shopping assistant can read your products and tell a shopper what they are, what they cost and whether they can be bought. This is the score most of your work moves.
• SEO (How you look in search): The few lines a search engine shows before anyone reaches your store: the headline, the summary line under it, and the written description on your photos. Scored on its own, separately from your Agentible score.
• GEO (Whether AI can reach you): Whether an AI assistant is allowed to open your storefront at all, can read the words on the page, and can tell that what it found belongs to your brand. Turning away the readers that only gather training material never counts against you here.
• AEO (Whether you answer questions): Whether your store holds the answers shoppers want before they buy: returns, delivery, and product copy with real numbers in it rather than a mood. Scored on its own, separately from your Agentible score.

The AI readiness and SEO agents generate text (product descriptions, conversational attributes, search titles, meta descriptions). The GEO and AEO agents call no AI model at all: they are deterministic checks of your storefront.

All generated text is written by Claude, a model built by Anthropic. Anthropic is listed on our sub-processor page with what it receives and where it processes. We never train a model on your data, and Anthropic's API terms do not train on it either.

No generated text reaches your storefront without your review and approval. Every proposed fix shows the exact before and after, is labelled as written by AI before you decide, and can be reverted for free after publishing. You can also choose export-only mode and download fixes as a file instead of publishing anything. Where our compliance checks find a claim in generated copy that needs a human eye, the copy is held for your review rather than published.

Under Article 50(2) of the EU AI Act, AI-generated content must be identifiable by machines, not just people. Every AI-generated fix Agentible serves carries that marking: the API responses and export files say which content is AI-generated and which was computed by deterministic rules, along with the generation's provenance (a fingerprint of the prompt version, the exact model that answered, and the software build). Rule-computed fixes, such as schema.org markup, are deliberately not marked as AI-generated, because they are not.

We assessed every agent against the AI Act's risk categories (assessment dated 2026-08-27, reviewed whenever an agent is added or changed). None of Agentible's agents is a high-risk AI system under Annex III: generating and checking product copy for an online store is not biometrics, credit, employment, essential services, or any other listed high-risk area. The obligations that do apply to us are the transparency duties, and the marking described above is how we meet them.

Agentible is a downstream provider building on Anthropic's general-purpose model; Anthropic carries the model-level obligations. You, the merchant, review and publish the text on your own store, and your approval is a real editorial decision, not a formality: that is by design.

For quality and traceability we keep a record of each agent run and its outcome, keyed to your store: which agent, when, which prompt version and model, and whether you approved or skipped the result. These records deliberately contain no product text, no prompt text and no model output. If your store is erased, the link between those records and your store is severed. The full detail of every table we hold is on the privacy policy, and the processing terms are in the DPA.