Skip to main content

​
Content Classification

Use this guide after your product lines are created and verticals are mapped. Classification assigns your existing content library to the correct product lines so that content generation, automation, and reporting filter correctly from the start. If you’re setting up a brand-new website with no existing content, you can skip this guide — new content is scoped at creation time.

​
What You’ll Need

  • Multi-product mode enabled
  • At least one product line created with a name and description
  • Existing content in WisePilot (assets, CTAs, offers, ad creatives, stories, or conversion units)

​
The Starting Prompt

Always run a dry run first. Classification applies in bulk — reviewing proposals before applying prevents mis-tagging. 
Propose a classification plan for all content on [website name] — dry run first, don't apply yet.
The dry run is free to re-run. Use it to spot patterns before committing. If the proposals look wrong, refine your product line descriptions and run it again — better descriptions produce better classification.

​
What Claude Will Do

Claude runs propose_classification, which scans 6 entity types across your content library:
Entity typeWhat’s scanned
Content assetsBlog posts, landing pages, commercial pages
CTAsCall-to-action variants and their conversion language
OffersOffer names and descriptions
Ad creativesAd headlines, descriptions, and asset names
StoriesCase studies and testimonials
Conversion unitsForm names, calendar names, checkout page titles

​
How Classification Works

Classification is deterministic string matching — not LLM inference. The system matches entity text against product line names, descriptions, and vertical names. This makes it fast, predictable, and auditable. For each entity, the proposal includes:
  • Suggested product line — the best match
  • Confidence level — high, medium, or low
  • Match reason — which term triggered the match (e.g., matched “ERP” in product description)
High-confidence proposals are typically safe to apply in bulk. Review medium and low-confidence items individually.

​
Reviewing Proposals

After the dry run, Claude will present a summary grouped by entity type. Look for:
  • Correct bulk assignments — high-confidence items you can approve as a group
  • Ambiguous items — content that could belong to multiple products (see edge cases below)
  • Unmatched items — content that didn’t match any product line (this is fine — see below)
Ask follow-up questions like: 
Show me all medium-confidence proposals for blog posts.

Which CTAs matched multiple product lines?

​
Applying the Classification

Once you’re satisfied with the proposals, apply them by entity type: 
Apply the classification for blog posts and landing pages.

Apply the classification for CTAs and offers.
Claude will use batch_tag_product_line to write the product line assignment to each entity. You can apply entity types one at a time and pause between batches to review results in the UI.
Applied classifications can be changed at any time by editing an individual entity or running a new classification pass. Batch tagging is not destructive — it only sets the product line field on unscoped entities unless you explicitly ask it to override existing tags.

​
Common Edge Cases

Content that matches multiple products — When a blog post or CTA references language from two different product lines, classification will flag it as ambiguous rather than guessing. These items surface in the medium-confidence list. Decide which product line owns the content, then manually tag it or tell Claude: “Assign the ambiguous [entity type] items to [product name].” Content that doesn’t match any product — Unmatched content stays unscoped. This is intentional and correct for cross-product content like brand-level blog posts, general FAQs, or company announcements. Unscoped content remains available to all product lines during generation — it won’t be filtered out. Product descriptions that are too generic — If your product line descriptions use vague language (e.g., “Our main software product”), the matching algorithm has little to work with. Before re-running, update your product descriptions with specific terminology, product names, feature language, and industry terms your content actually uses.
Do not run classification on a website where product lines are still being set up. Proposals are generated against the product configuration at the moment of the scan — if you add or rename a product line afterward, re-run the dry run before applying.