# Make your products easier for ChatGPT to recommend

> ChatGPT reasons about what a product is, not about exact keywords. Convrail scores every feed item from 0 to 100 on validity and completeness, lists the actions that gain the most points, and can draft a title and description written for a language model. Drafts are suggestions: nothing changes in your store until you accept them.

Canonical: https://convrail.com/feed-optimization/

## The completeness score, explained

The score is deterministic and the same for every store. A valid row earns 50 points; each recommended field adds a fixed bonus. Actions are sorted by the points they would gain, so the list starts with what matters most.

*How the 0 to 100 completeness score is computed*

| Criterion | Points | Why it matters |
| --- | --- | --- |
| Row passes validation | 50 | An invalid row is never delivered, so nothing else counts until it is fixed |
| GTIN or MPN present | 10 | Identified products are matched to known catalogs far more reliably |
| Product category | 8 | Tells the model where the product belongs in a taxonomy |
| Description of 200 characters or more | 6 | The model reasons on meaning; a one-line description gives it nothing to reason on |
| Sale price when on promotion | 6 | Lets the assistant quote the price the shopper will actually pay |
| Additional images | 5 | More views, better product cards |
| Rating and review count | 5 | A strong trust signal when the assistant compares options |
| Variants described (variant_dict) | 5 | Sizes, colors and materials become answerable questions |
| Return window in days | 5 | A concrete policy the assistant can state |

## AI-drafted titles and descriptions, reviewed by you

The rewriting assistant produces a title under 150 characters, never in capital letters, and a plain-text description between 200 and 1,000 characters that states what the product is, its materials or specifications, who it is for, its main uses and its constraints (sizes, compatibility). Each draft comes with a short rationale.

- Drafts respect the feed specification by construction: the same rules the validator enforces are part of the prompt, and the output is validated again before it is shown.
- Nothing is written to your store automatically. You read the draft, edit it if you want, and apply it yourself.
- The model provider is configurable. Without an API key, the assistant runs in a deterministic mode that only restructures your existing text.
- Brand and category are passed as context so the draft does not invent attributes the product does not have.

## Context hints for the assistant

Beyond individual products, Convrail generates hints about your catalog as a whole: which attributes are missing across a category, which descriptions are near duplicates, which products lack an identifier. The goal is to spend your editing time where it changes what the assistant can say about your store.

## Frequently asked questions

### Does Convrail change my product pages?

No. Scores and drafts are shown in the app. You decide what to apply, and you apply it in your store. The feed reflects your store on the next run.

### Which AI model writes the drafts?

The provider is configurable by the operator. Prompts encode the feed rules (150-character titles, no capital letters, plain text) and the output is validated before display, whatever the model.

### Is my catalog used to train a model?

Convrail sends only the fields needed for a draft (title, description, brand, category) to the configured provider, under that provider’s API terms. Your catalog is not used by Convrail to train anything.

### Why does a valid product still score 50?

Because validity is the floor, not the ceiling. The other 50 points reward the recommended fields that make a listing easier to match and describe: identifiers, category, a real description, ratings, variants and policies.