Taxonomies rot because models drift and categories accrete. Prother's answer is boring on purpose: seven primary categories, one per tool, plus a controlled tag vocabulary that crosses categories.
The seven
💬 Conversational AI & Chatbots · 🎨 Generative Content Creation · 📝 NLP & Text Utilities · 👁️ Computer Vision · 📊 Data Analytics & Predictive Modeling · ⚙️ Automation & Workflow Orchestration · 🛠️ Developer Frameworks & Infrastructure.
We rebuilt the taxonomy around the seven ways evaluators actually shop for AI: ask, create, write, see, analyze, orchestrate, build. The old buckets described marketing decks; these describe jobs to be done.
Rules that keep it alive:
- One primary category per tool. Multi-category listing feels generous and scores terribly: it dilutes browse pages and confuses comparisons.
- No "Other." An Other bucket is a graveyard. If something genuinely doesn't fit, that's a signal the taxonomy is behind the ecosystem: worth a review, not a bucket.
- Categories are editor-managed, zero-deploy. The admin console can add, rename, and reorder categories; the site reads them from the database at request time.
Tags do the cross-cutting
Where a category answers what is it for, tags answer how does it run: open-source, free-tier, api-available, self-hosted, no-code, enterprise, browser-extension, mobile. Controlled vocabulary: free-text tags become spam within a quarter.
The quarterly drift review
Every quarter: pull the zero-result search queries, the category page bounce rates, and the "uncategorizable" submissions. If three tools failed to fit last month, the taxonomy has a hole. If a category has two tools, it's a tag, not a category.
Taxonomy is a product surface, not metadata. It's how evaluators think, how category pages rank, and how a directory avoids becoming an undifferentiated firehose.