Product Feed Enrichment
Add-on for Upgates
04Use of AI

How the add-on uses artificial intelligence

The add-on uses two language models: one proposes missing data, the other checks the proposals. You decide what gets written to your store.

In short

AI proposes, you decide

Nothing you have not confirmed is written to your store. Every batch can be reverted.

Product data only

The models receive product names, descriptions, parameters, categories and images. The add-on does not read orders or customers.

72 hours at most

That is how long product data stays in the service after processing ends.

  1. 1
    Your store data
    name, description, parameters, category, image
  2. 2
    Proposal
    language model 1 (Anthropic)
  3. 3
    Check
    language model 2 (OpenAI)
  4. 4
    Your decision
    accept, edit, reject, confirm the write
  5. 5
    Write to store
    a batch that can be reverted

Who processes the data

ProcessorPurposeWhat it receives
Anthropiclanguage model that proposes missing dataproduct data
OpenAIsecond language model that checks proposalsproduct data and the proposal
Vercelrunning the add-on applicationproduct data
Convexadd-on databaseproduct data, proposals and the write log

What the add-on uses AI for

  • It proposes missing Google Shopping data and product parameters from data your store already has. Whatever cannot be read from it stays empty.
  • Every proposal is checked by a second language model. The add-on marks a proposal as „ověřeno“ (verified), „opraveno kontrolou“ (corrected by the check) or „neověřeno“ (not verified). Only verified proposals can be accepted in bulk.
  • The add-on has no chat and no free-form prompt. It does one predefined thing, so there is no bulk change you did not expect.

Texts written by AI

The Google title and description are written by a language model. They go to the feed without a label telling Google they were created by AI. That is why these two fields are not selected by default; tick them only if you are fine with that.

Review and responsibility

A language model can make mistakes. Read a proposal before accepting it, especially when the check corrected it or could not verify it. Nothing is written to your store until you confirm, and every written batch can be reverted.

Privacy policy · Terms

Last updated 12 November 2026.

Questions about the add-on? Email

How feed enrichment works as a service →