> ## Documentation Index > Fetch the complete documentation index at: https://docs.merchkit.com/llms.txt > Use this file to discover all available pages before exploring further. # Enrichment Quality: What 'Good' Looks Like > Learn what quality AI-generated content looks like and how to establish benchmarks for your enriched catalog. # Enrichment Quality: What 'Good' Looks Like Your catalog is now enriched with AI-generated content. But how do you know if it's *good*? This article helps you set benchmarks, spot problems, and improve quality before you push enriched data to your sales channels. ## Quality Benchmarks by Attribute Type Different attribute types have different quality markers. Here's what to look for: ### Product Descriptions **Signs of good quality:** * Length is appropriate for your channel (e.g., 150 words for marketplace listings, 300+ for your own site). * Tone matches your brand voice (is it technical, casual, luxury, eco-focused?). * Key details are included: materials, dimensions, use cases, benefits. * No filler or generic text like "this product is great" without specifics. * Grammar and spelling are correct. * Reads naturally — not obviously machine-generated. **Signs of poor quality:** * Too short (5 words) or too long (1,000+ words) for the intended channel. * Generic boilerplate: "This item is a high-quality product designed for everyday use." * Hallucinated details not found in your source data. * Missing key information (what material is it, how big is it?). * Awkward phrasing or grammar errors. * Reads like a tech manual when you need friendly, conversational tone. \[SCREENSHOT: Side-by-side comparison showing a poor description ("This is a chair. It has four legs. You can sit on it.") next to a good description ("Mid-century modern dining chair with solid walnut frame and hand-stitched upholstery. Pairs well with industrial tables. Available in navy, charcoal, and natural linen.")] ### Categorical Values (Categories, Collections, Tags) **Signs of good quality:** * Every value matches your acceptable values list exactly. * Values are consistent — similar products get the same tags. * Values are specific, not vague ("Sofas" is better than "Furniture"). * The selection makes sense for discovery (if it's a "Bedroom Set," it's also correctly tagged "Bedroom"). **Signs of poor quality:** * Values don't exist in your acceptable values list. * Inconsistent tags across similar products ("Sofa," "Couch," "Sectional" for the same product type). * Too broad or too vague ("Home" when you meant "Home Office Furniture"). * Missing obvious categories that should be filled in. ### Numeric Values (Price, Stock, Dimensions, Weight) **Signs of good quality:** * Values are within plausible ranges for your product type. * Units are correct and consistent (all weights in lbs, or all in kg — not mixed). * Precision matches your needs (dimensions to 0.1", weight to 0.5 lbs). * Values align with your source data (if your vendor says 48", the AI should say 48", not 50"). **Signs of poor quality:** * Implausible values ("Chair height: 2,000 inches"). * Mismatched units or missing units entirely. * Hallucinated numbers not in your source data. * Rounding errors or inconsistency (one product says "24"" and the next says "2 feet" for the same measurement). ### Images **Signs of good quality:** * Images are clear and well-lit. * The product is the main focus (not a crowded shelf or busy room). * Image dimensions match your channel requirements. * Images are relevant to the product and accurate. * No watermarks or logos from competitor sites. **Signs of poor quality:** * Blurry, dark, or low-resolution images. * Wrong product entirely (wrong color, style, or item). * Images are too small or don't meet your channel's minimum dimension requirements. * Watermarked or copyrighted images that can't be used. * Generic stock photos that don't match your product. ## Consistency Across Similar Products One of the biggest quality markers is **consistency**. Sample 10 products in the same category: * Do they all have descriptions of roughly the same length? * Are they all using the same tone of voice? * Are category tags applied consistently (all sofas tagged "Sofas," not half tagged "Sofas" and half tagged "Seating")? * Are numeric values using the same units and precision? If you see wild variation, your prompts or acceptable values need tightening. ## Before and After: The Impact of Refinement Let's say your initial product descriptions were vague. Here's what improvement looks like: **Before (Poor Prompt, Generic Sources):** > "This sofa is a comfortable piece of furniture for your living room. It comes in different colors and sizes. Great for relaxing." **After (Refined Prompt, Specific Sources Added):** > "Contemporary mid-century sofa with solid hardwood frame, high-density foam cushioning, and stain-resistant fabric. 84" wide, perfect for large living rooms. Available in charcoal gray, navy, and natural linen. Pairs beautifully with modern or transitional decor. Ships fully assembled." The difference: the second version includes specific dimensions, materials, available colors, care details, and styling guidance. This came from: * A more detailed prompt that asked for materials, dimensions, and care tips. * Adding the manufacturer's spec sheet as a source. * Tightening acceptable values for colors so the AI picks from valid options. ## How to Improve Your Enrichment Quality If your quality isn't where you want it, take these steps in order: ### 1. Refine Your Prompts Go to **Workspace → Attributes** and review your prompts for each attribute: * Are you asking for the right details? (Example: "Include materials, dimensions, and care instructions" is much better than "write a description.") * Are you specifying tone and style? ("Write in a friendly, conversational tone" vs. "Write professionally.") * Are you setting length expectations? ("150–200 words" vs. open-ended). * Are you pointing to specific sources? ("Use the manufacturer's spec sheet and customer reviews, not competitor descriptions.") ### 2. Add Better Sources Go to **Workspace → Attributes** and check the Product Data context. Are you using all available sources? * Manufacturer specs and datasheets? * High-quality images (not just low-res thumbnails)? * Customer reviews and testimonials? * Vendor descriptions (though avoid just copying them)? More sources = better context = better AI output. ### 3. Tighten Your Acceptable Values For categorical attributes, your acceptable values list is the guardrail. If the list is too long, vague, or inconsistent, the AI will struggle: * Remove near-duplicates ("Sofa," "Couch," "Sectional Sofa" — pick one canonical term). * Be specific ("Women's Footwear" instead of just "Shoes"). * Add helpful descriptions next to each value so the AI knows when to use it. ### 4. Be More Specific About Product Data Context In the attribute settings, the "Product Data" field lets you tell the AI what fields to look at. Instead of letting it use everything: * Specify: "Use the 'Specifications' section only, not competitor reviews." * Or: "Use the product images and the manufacturer's material list. Ignore reviews with fewer than 4 stars." ### 5. Test and Iterate After making changes, use "Generate All Attributes" on a small batch (5–10 products) and review. Does the output improve? If so, scale to your full catalog. If not, refine further and test again. \[SCREENSHOT: Before/After side-by-side showing enrichment results, with annotations highlighting specific improvements like "Now includes material," "Tone is friendlier," "Dimensions are specific."] ## Red Flags: When Output Starts to Fail Watch for these signs that your enrichment is degrading: * **Hallucination:** The AI is making up details not in your source data (e.g., claiming a product has features it doesn't). * **Repetition:** Descriptions are getting copy-paste boilerplate across multiple products. * **Mismatches:** Categorical values that don't match your acceptable values list. * **Dropout:** Some products stop generating values entirely. If you see these, pause enrichment and revisit your prompts, sources, and acceptable values. Small fixes often resolve them. ## You've Hit the Core Onboarding Milestone If you've configured your attributes, enriched your catalog, and reviewed the output — **congratulations, you've hit the core onboarding milestone.** Your catalog is now rich, structured, and ready for the next step. From here, you have two paths: **Connect an Integration** — If you imported your catalog from a platform like Shopify or WooCommerce, you can sync your enriched data back. Head to [How Integrations Work in Merchkit](/product/integrations-and-field-mapping/how-integrations-work) in Section 4. **Add a Channel** — Once your enrichment is solid, you can optimize content for specific sales channels (Amazon, Wayfair, your own site). Go to Section 5 to learn about channels and start generating channel-optimized listings. Both paths are available to you now. Choose based on your immediate needs.