How We Built a 6,000-Tool AI Directory (And What We Learned)
Building a directory of AI tools sounds simple. Collect tools, add descriptions, publish a website. We thought the same thing. Two years later, we're at 6,199 tools and counting, and the unglamorous reality is that almost nothing about this project went the way we expected.
Here's an honest account of how TryAnAI got built, what the data has shown us, and what we'd do differently.
Why Directories Fail (And Why We Did It Anyway)
Most AI directories are abandoned within six months. You can spot them easily: tools from 2023 still listed as "new," dead links that haven't been cleaned up, categories that made sense 18 months ago but are obsolete now.
The graveyard of failed AI directories is actually one of the reasons we built this one. We were tired of landing on lists that hadn't been maintained. The value in a tool directory isn't the initial collection — it's the ongoing curation. Anyone can scrape ProductHunt. Not everyone sticks around to verify that the tools still work, that the pricing is accurate, that the categories make sense as the market evolves.
Our thesis was simple: build infrastructure that makes ongoing maintenance tractable, not just an initial scrape job.
The Technical Reality
We started with a basic Next.js app and a Postgres database. That part was fine. What got complicated fast was the data pipeline.
AI tools change constantly. A tool that launched with a freemium model switches to paid-only three months later. A video generator gets acquired and the standalone product disappears. A niche coding tool adds image generation and now belongs in three categories instead of one.
To keep up, we built:
- Automated discovery: Scrapers that monitor AI tool directories, sitemaps, ProductHunt launches, GitHub trending repos, and 300+ RSS feeds
- Enrichment pipeline: Each discovered tool gets processed through an AI model that extracts pricing, use cases, and categories from the tool's own website
- Deduplication: More complex than it sounds. The same tool often appears under slightly different names across different sources
- Status checking: Regular verification that links still work and tools are still active
The enrichment step is where we've had the most friction. Language models are good at extracting structured data from marketing copy, but marketing copy is deliberately vague about pricing. "Free to start" could mean a 7-day trial or a genuinely unlimited free tier. We've had to build a lot of disambiguation logic.
What 6,000 Tools Tells You
Once you've indexed this many tools, patterns start to emerge that you can't see from the outside.
The category distribution is lopsided. AI writing tools and AI image generators account for a disproportionate share of the market — roughly 35% of all tools we've indexed fall into these two buckets. That's both a market signal (people want this) and a quality problem (the good tools are buried under dozens of near-identical alternatives).
The pricing landscape is chaotic. About 40% of tools offer some kind of free tier. But "free" covers everything from truly unlimited free access to a trial that requires a credit card and expires in 48 hours. We've tried to normalize this into consistent categories (Free, Freemium, Paid, Free Trial), but the edge cases are endless.
Tool longevity is shorter than you'd think. We've seen roughly 15% of the tools we initially indexed go dark or get absorbed into other products within 18 months. The AI tool graveyard is real — and it's one reason why our status-checking infrastructure matters more than the initial discovery.
The niche tools are often better than the famous ones. A specialized AI tool for legal contract review built by a team of lawyers will usually outperform a general-purpose AI assistant trying to do the same thing. We've leaned into this with our category structure — the more specific the category, the more useful the recommendations tend to be.
The Curation Decisions Nobody Sees
Every directory makes choices about what to include and exclude. Here are some of ours:
We don't list tools that are in stealth or invite-only. If you can't actually try it, it doesn't belong in a directory of tools you can use. This cuts out a lot of hype-driven launches.
We don't list tools that are obviously wrappers with no value-add. The number of products that are just a ChatGPT API call with a thin UI and a $30/month price tag is... depressing. If the tool's entire value proposition is "access to GPT-4 with a nicer interface," we generally skip it.
We do list tools that are in early access if they're genuinely accessible. The line between "beta" and "launched" is blurry in AI, and we'd rather include a rough-but-functional early-access tool than wait until a formal announcement.
What We'd Do Differently
Start with better category taxonomy. Our initial category structure reflected the AI tool market of 2023, which looked very different from today. We've had to do a significant restructuring twice, which is painful when you have thousands of tools already filed under old categories.
Build the enrichment pipeline earlier. We started with manual data entry for the first few hundred tools, which created inconsistency that took months to clean up. The automated pipeline should have been day one infrastructure.
Track tool updates more aggressively. We get a lot of reader messages along the lines of "this tool's pricing changed" or "this tool shut down." We've improved our monitoring, but we're still partially dependent on users flagging stale data.
Where This Goes
The directory itself is a foundation. The more interesting work is on top of it: tool comparisons that test specific use cases head-to-head, curated category lists that reflect actual user needs rather than marketing categories, and eventually personalized recommendations based on what a user is actually trying to accomplish.
The AI tool market isn't slowing down. New tools launch every week. Some of them are genuinely useful. A lot of them aren't. Finding the useful ones without spending hours testing them yourself is the actual problem worth solving.
That's what we're building toward.