Skip to main content

Reviewing AI-Generated Content

After you run enrichment, your results appear directly in the Products table. This is where you assess whether the AI-generated values are accurate, consistent, and ready for use. You don’t need to check every single product—a smart spot-check approach is faster and more effective.

Where Your Results Live

Once enrichment completes, head back to your Products view. The attributes you just enriched now show values in their respective columns. [SCREENSHOT: Products table with newly populated attribute columns showing AI-generated values]

The Review Workflow

Your enrichment results live in the same Products table where you work every day. This means:
  • Review in context. Inspect current values in the table, or use generation Job Details for the values a batch wrote.
  • Real-time feedback. As you review, you can immediately edit values (see Editing and Overriding AI Output for details).
  • Easy to scale. Once you’re confident in the quality, you can run enrichment on larger batches knowing your review process is simple.

Rationale, Citations, and Confidence Rating

When available, rationale explains how the model arrived at a value, confidence describes the model’s assessment, and a source reference points to supporting context. Use these to decide what to inspect; none is an independent accuracy check.

Check a generated value against its source

Suppose an AI-written description says a chair is solid oak. A citation gives you a place to check that claim; it does not prove the claim is correct. Before starting, generate a value with an AI prompt attribute that has the relevant supplier information as context. See Configuring an attribute for AI. You need permission to generate attributes; editing a result also requires permission to edit product values.
  1. In Products, find the generated description or detail and click its cell to open the information popover.
  2. Read Rationale and Confidence when present. These are the model’s explanation and confidence, not an independent quality check.
  3. Look for Source. Its badge identifies a Data Source, Attribute, Image or Web reference.
  4. Click a URL citation to open it in a new tab. For an attribute or image reference, inspect that information in Merchkit. Hover over a shortened badge to read its full reference.
  5. Compare the actual supplier information with the generated value. Correct unsupported claims using Editing AI output, or improve the context and regenerate a small sample.
Inferred vs. extracted values. A model may infer a value from a parent product, related attributes or other context rather than quote it directly. Read the rationale when available and compare it with the actual evidence. Do not assume that every inference is labeled or always receives a particular confidence rating. Using confidence during review. Treat a Low or Medium rating as a reason to inspect the source more closely, and spot-check High ratings too. Review the recorded confidence in cell information or generation Job Details. These signals help prioritize attention; they do not establish correctness.

When a source is missing

Source citations are optional, model-reported primary references for newly prompt-generated values. They are not an exhaustive list of everything the model considered.
  • Older prompt-generated values need regeneration to gain this information; adding a source does not annotate existing values automatically.
  • Manual value edits clear the AI provenance, including source, rationale and confidence.
  • Taxonomy, image, video, formula and other non-prompt generators do not populate this Source field.
  • The reference may be an attribute name or the aggregate Product Images, rather than a link to one exact image.
  • CSV exports and public API responses do not include this source metadata.
An absent citation is not evidence that a value is wrong, and a citation is not evidence that it is right. Review the underlying information before publishing.

Review what a generation job wrote

Use this when you want to inspect a batch’s results, rather than only today’s cell values.
  1. Open Jobs in the workspace.
  2. Find the attribute-generation job and click Details.
  3. Review the affected entities and attribute columns in the paginated table. These are the values recorded when the job wrote them; later edits can differ.
  4. Inspect confidence where recorded, and hover over its badge for rationale when available. Use pagination to review more affected entities.
Details is available for attribute-generation jobs, not every job type. Older results depend on retained audit records. A job spanning an archive transition can temporarily show a partial set; retry later if a recent archival transition is suspected. This view does not add rollback coverage. Reference-generation values can also carry rationale, confidence and failure information in their cell popovers. Review that context when a reference is empty or the selected image is unexpected; it is separate from the prompt-only Source citation.

Reviewing at Scale

If you’ve enriched hundreds or thousands of products, don’t try to check every one. Instead, use these techniques to review efficiently:

Use Views to Filter Recently Enriched Products

Create a view that shows only the products you just enriched:
  • Filter by a date range (products updated today)
  • Filter by category or vendor to review similar products together
  • Combine filters to create a focused sample
This narrows your review set to a manageable number. [SCREENSHOT: Products view with filter applied showing only recently enriched products]

Sort by Specific Attributes

Click the column header of an attribute you just enriched to sort by that column. This groups similar values together, making patterns and outliers easier to spot. [SCREENSHOT: Products table sorted by a newly enriched attribute, showing all values of the same type grouped together]

Scan Across Rows

Once you’ve filtered and sorted, scan horizontally across the table to see how each product’s enriched data looks in context with other product information (name, category, price, etc.). This helps you judge whether the AI-generated value makes sense for that specific product.

What to Look For

When spot-checking, evaluate on three dimensions:

1. Accuracy

Does the value match the product?
  • Example: If you enriched “Suggested Use Case” for a camping tent, does the AI suggest appropriate activities (camping, hiking) rather than unrelated uses?
  • Red flag: Values that don’t match the product’s actual purpose or nature.

2. Format

Does the value follow the structure you asked for?
  • Length: Is it the right length (one sentence, 3 bullet points, etc.)?
  • Tone: Does it match your brand voice (professional, casual, technical)?
  • Structure: If you asked for a specific format (e.g., “Price range: X–X–Y”), is it formatted correctly?
  • Red flag: Values that are too long, too short, wrong tone, or don’t follow your template.

3. Consistency

Do similar products have similar output patterns?
  • Example: If you enriched “Size Range” for apparel, do all shirts have sizes like “XS–XXL,” and all jackets also follow that same pattern?
  • Red flag: One product says “Fits sizes 2–20,” another says “Small to Extra Large,” and a third says “All sizes.” Same attribute, three different formats.

Spot-Checking Strategy

Rather than reviewing every product, pick a representative sample:
  1. Choose products across categories. If you sell apparel, home goods, and electronics, pick a few products from each category.
  2. Sample different vendors. If you work with multiple suppliers, review products from 2–3 vendors.
  3. Pick different price ranges. If your catalog spans budget to premium, include products from each tier.
  4. Include edge cases. If some of your products are unusual (limited editions, custom items, very old stock), include one or two of those.
This approach catches issues across your product diversity without needing to review every row.

Tracing Back to Configuration

If output looks wrong, the issue usually lies in one of four places:
  1. Prompt clarity. Your prompt might be too vague, contradictory, or missing context. Re-read it and imagine explaining it to a colleague—if it’s unclear to you, the AI will struggle too.
  2. Product data context. The AI might be missing the information it needs. Check that you’ve selected the right Product Data fields when configuring the attribute.
  3. Image Source. If your attribute relies on product images, verify that the images are clear, relevant, and actually show the product (not just packaging or logos).
  4. Acceptable Values (if set). If you’ve limited the AI to specific values, verify those values are appropriate for the actual products you’re enriching.
To troubleshoot, go back to the Attributes page (Workspace → Attributes), find the attribute, and review its configuration. Once you’ve made changes, use Re-running Enrichment to test your improvements on a small batch before scaling.

Common Issues

Output is too generic or vague
  • Example: “Description” shows “A product that serves its intended purpose” for all products.
  • Likely cause: Your prompt is too broad or Product Data context is missing. Try adding more specific guidance (e.g., “Include the material, primary use, and key features”) and ensure you’ve selected descriptive Product Data fields.
Values don’t match the expected format
  • Example: You asked for “Price: X–X–Y” but got “Ranging from X dollars to Y dollars.”
  • Likely cause: Your prompt template isn’t strict enough, or Acceptable Values aren’t set. Consider refining your prompt with an explicit example, or use Acceptable Values to enforce a specific pattern.
Some products have blank values while others don’t
  • Example: 50 products enriched successfully, but 20 are empty.
  • Likely cause: The AI requires a dependent attribute to generate the value, and some of your products are missing that dependency. Check your attribute configuration and verify all required upstream attributes have values.
Values are inconsistent across similar products
  • Example: One “Fabric Type” says “Cotton,” another says “100% cotton,” another says “Natural fiber—cotton.”
  • Likely cause: Your prompt doesn’t enforce consistent formatting, or Acceptable Values aren’t used. Consider using a strict Acceptable Values list (Cotton, Polyester, Wool, Blend) to ensure uniformity.

Next Steps

  1. Fix individual values: If you spot errors in specific cells, head to Editing and Overriding AI Output to correct them.
  2. Improve your prompt: If you see systematic issues, update your attribute configuration and then check out Re-running Enrichment to re-generate with the updated logic.
  3. Scale up: Once you’re confident in quality, run enrichment on larger batches using the same approach.