Confidence Scoring
Confidence scoring gives you an instant read on how reliable the AI-generated content is for each product field. Every field in the product detail view carries a score, a confidence level, and the reasons behind it — so you know where to focus your review effort.
How It Works
When merchi.ai generates content for a product, it assesses each output field and records:
- Score — a percentage from 0–100
- Level — a plain-English label (High, Medium, Low, or Very Low)
- Reasons — why the score is what it is (e.g. data was provided directly, derived from the product name, or inferred with limited evidence)
- Sources — where the AI drew the information from (e.g. provided data, image analysis, PDF analysis, web data)
This happens automatically for every product, with no configuration required.
Confidence Levels
| Level | Score range | What it means |
|---|---|---|
| High | 80–100% | The AI had strong, consistent evidence across multiple sources |
| Medium | 60–79% | Solid evidence but with some gaps or inference involved |
| Low | 40–59% | Limited evidence; the content may be partially inferred |
| Very Low | 0–39% | Minimal evidence; treat the output as a starting point only |
Where to Find Confidence Scores
Product Detail Modal
Open any product to see confidence scores in two places:
Overall score — shown at the top of the content panel. This is the average across all scored fields. It also shows a count of fields flagged for review (Medium or below). Click the overall score to see a list of which fields need attention.
Per-field badges — each individual field shows its own badge. Click a badge to see:
- The score and level
- The specific reasons the AI assigned that score
- The data sources used for that field
All Products Grid
The product grid shows a confidence badge on each row, using the same colour coding as the detail modal:
- Green — High confidence
- Yellow — Medium confidence
- Orange — Low confidence
- Red — Very Low confidence
How the Overall Score Is Calculated
The overall score is an unweighted average of all individual field scores. For example, if a product has five scored fields with scores of 90, 85, 70, 60, and 45, the overall score is (90 + 85 + 70 + 60 + 45) ÷ 5 = 70%.
Fields that are not scored (such as ETIM classifications, image metadata, or fields that require legal review) are excluded from the calculation.
What Affects a Confidence Score
Scores reflect the quality and completeness of the source data available when the product was processed. Common factors:
- More input data = higher confidence — products with detailed input fields, uploaded PDFs, or rich images score higher than those processed from a product ID alone
- PDF documents — when a product PDF is available, fields derived from it are attributed to PDF analysis and typically score higher
- Image analysis — product images contribute to visual attributes like colour, material, and dimensions
- Provided data — values passed in directly as structured input carry the highest confidence
- Inference — when the AI must infer a value from indirect evidence, the score is lower
Acting on Confidence Scores
Low-scoring fields are candidates for manual review. Common actions:
- Enrich the source data — add missing input fields, upload a product PDF, or provide better images, then re-process the product
- Edit directly — open the field in the product detail modal and update the content manually
- Change workflow status — move low-confidence products to Review status to flag them for a team member
High-confidence fields rarely need attention, but the score is not a guarantee — always apply commercial judgement, particularly for fields with legal or compliance implications.
Frequently Asked Questions
Why does a field show no confidence badge? Some fields are not scored because their content is not AI-generated. ETIM classifications, image metadata, and fields that require manual legal review are excluded.
Can I turn confidence scoring off? No. Confidence scoring runs automatically as part of every generation and cannot be disabled. It does not affect the generated content itself.
Does re-processing a product reset the confidence scores? Yes. Each time a product is processed, the scores reflect the most recent generation only.
Why is my confidence score low even though I provided a lot of data? Check which sources appear in the field-level badge. If the AI did not successfully extract information from a particular source (for example, a low-quality image or a scanned PDF), the score may be lower than expected. Providing clearer source material and re-processing usually improves scores.