Why we rebuilt the category system from scratch

The dataset we imported carried 1,376 distinct categories for 4,921 tools. That is roughly 3.6 tools per category. Dozens held exactly one listing. "Meeting summaries", "meeting notes" and "meeting transcription" existed side by side as separate top-level concepts.

A taxonomy that fine-grained is not a taxonomy. It is a tag cloud with ambitions, and it fails at the one job it has: helping someone who does not yet know what the tool is called find the right shelf.

What was broken

• No browsable surface. A category page with one tool teaches you nothing about your options.
• Split popularity. The same demand was spread across four near-identical labels, so nothing looked popular.
• Unstable naming. Casing and pluralisation were inconsistent, which produced duplicate-looking pages.
• Invisible to search. Thousands of thin pages compete with each other rather than ranking.

The rebuild

We mapped every source tag to one of 26 main categories, chosen to match how people describe the job rather than how vendors describe the product. The original tag survives on each listing as a subcategory, so nothing was thrown away.

The remaining 18 groups cover marketing, design, sales, support, education, health, security, ecommerce, video and the rest. You can see all of them on the categories page.

The judgement calls

Three decisions were genuinely contested, and it is worth being open about them.

Where an assistant belongs. A chatbot that writes marketing copy could sit in either group. We assigned by primary interface: if the product is a conversation, it is an assistant; if it is a document or campaign workflow, it is a writing tool.

Agents as a category or a property. A growing share of new listings call themselves agents. We treat that as a capability, not a shelf — an agent that books meetings belongs in productivity, not in a catch-all "agents" bucket that would repeat the original mistake.

Keeping "Other AI Tools". A residual group is unfashionable but honest. Forcing a genuinely novel tool into an adjacent category is worse for the person browsing than admitting it does not fit yet.

We deduplicated at the same time

Consolidation exposed a second problem: the same product listed more than once. Matching on name plus domain found 45 duplicate groups covering 66 redundant records — two Grammarly entries, two Jasper entries, and so on.

We merged rather than deleted, keeping the longest description and the highest vote and rating values, so no community signal was lost. The directory total went from 4,987 to 4,921, and every category count, featured list and sitemap entry was regenerated to match.

A count that shrinks after a data pass is usually a sign of honest maintenance, not shrinking coverage.

What we watch now

1. No category below 20 listings unless there is real search demand for that phrase.
2. New listings map on import, so the tag sprawl cannot rebuild itself.
3. Duplicate detection runs on every data pass, not once.
4. Category counts are generated, never hand-written, so the numbers on this page cannot drift from the database.

The full process, including how ratings and popularity signals are handled, is written up in our methodology.