Forem: 武乐丹

I Built an AI Tools Directory — 10 Lessons on What Actually Works

武乐丹 — Sun, 24 May 2026 04:10:14 +0000

Building an AI tools directory sounds straightforward. Scrape some data, build a UI, call it a day. After spending months building toolsdepth.com, here are the lessons that took me from "it works" to "people actually use it".

1. Categories matter more than you think

People browse AI tools by use case, not by AI model. "Code assistant" is useful; "Powered by GPT-4" is not.

2. Quality over quantity (at least for the landing page)

Having 500 tools is great for SEO. But the first 20 tools users see — those determine whether they stay or bounce.

3. Real screenshots beat mockups

Users can smell AI-generated demo screenshots. Actual screenshots of the tool interface build trust.

4. Pricing transparency is a competitive advantage

Most directory sites bury pricing. Showing "Free / $20/mo / Custom" in the listing itself increases click-through by a lot.

5. The filter UX is your biggest technical challenge

Categories + pricing + features + platform + rating = a complex filter UI that needs to be fast. If users cannot narrow down in two clicks, they leave.

6. New tools drive repeat visits

The "what is new" section is the most visited page. AI tools launch every week — showing freshness matters.

7. Reviews are the hardest thing to bootstrap

Users do not write reviews for directories without traffic. Seed reviews yourself (disclosed as editorial) until you reach critical mass.

8. SEO takes 3-6 months

Do not expect organic traffic week one. Focus on getting listed on other directories, write on Dev.to, answer Quora questions. The long game pays off.

9. Mobile-first is non-negotiable

Over 60% of our traffic comes from mobile. If your tool listings do not work on a phone screen, you lose more than half your audience.

10. The business model is still TBD

Affiliate links? Sponsored listings? Job board? Premium placement? Still figuring this out. If you have suggestions, leave a comment!

Would love to hear from others who have built similar directories — what worked for you?

Check it out at toolsdepth.com

I Built an AI Tools Directory. These 10 Lessons Hurt the Most.

武乐丹 — Sun, 24 May 2026 04:06:00 +0000

I Built an AI Tools Directory. These 10 Lessons Hurt the Most.

What nobody tells you about building a content site in the AI age.

Six months ago, I launched an AI tools directory. I thought the code would be the hard part. Build a scraper, spin up a database, design a clean UI. Weekend project.

Wrong. The things that decide whether a directory lives or dies have almost nothing to do with technology.

Here are the 10 lessons that cost me months of mistakes.

1. Categories Are Your Product — Not the Tools

I spent my first month obsessing over tool count. The real question I should have asked: how do users actually think about AI tools?

Nobody wakes up wanting "a GPT-4 wrapper." They want to write better emails, code faster, find design inspiration. They browse by use case, not by model.

When I rebuilt the site around workflow categories — Writing, Coding, Design, Research, Productivity — engagement surged. Time on site jumped 40%. Return visits doubled.

Your information architecture is the product. Get that wrong, and nothing else saves you.

2. The First 20 Tools Decide Everything

You may have 500 tools. Users see twenty. That's the game.

When I hand-curated the first twenty listings, bounce rate dropped from 78% to 54%. One change. Twenty-four points.

Curate your first screen like your business depends on it.

3. Screenshots > Mockups, Always

I replaced every generic image with real product screenshots. Click-through jumped ~30%. Users said "wow, this actually shows what it looks like."

Show the real thing. Every time.

4. Pricing Transparency Wins Trust

I hid pricing behind "Contact Sales" at first. Big mistake. When I switched to clear labels — Free, $20/mo, Custom — on every listing, trust improved across the board.

A directory that shows real prices gets bookmarked.

5. The Filter UX Will Break You

20 categories × 4 pricing tiers × 10 feature tags × 3 platforms × 5 ratings = 12,000 filter combinations. Every one needs to feel instant.

I rewrote it three times. If users can't narrow things in two clicks, they disappear.

6. "New" Is the Most Powerful Category

The second most visited page wasn't "Best AI Writing Tools." It was "Newly Added Tools."

AI moves absurdly fast. Users come back to see what's fresh. I added a "This Week in AI Tools" section and repeat traffic climbed.

Build for freshness, not just permanent collections.

7. Reviews Are Brutal to Bootstrap

Nobody writes reviews for a site with no traffic. Classic chicken-and-egg.

What worked: I wrote editorial reviews myself (labeled "Editor's Pick"). I contacted tool makers for official descriptions. I was transparent about everything.

After 3 months, organic reviews started trickling in. Seed your content. Be honest about it.

8. SEO Takes 3-6 Months. No Shortcuts.

Month 1-2: zero traffic. Month 3: trickle. Month 4: measurable. Month 5: meaningful. Month 6: server bills covered.

Start on day one. Measure on month six.

9. Mobile-First Is Survival

60% of visitors were on mobile with a terrible experience. I rebuilt mobile-first. Mobile bounce rate dropped from 82% to 61%.

If your site works better on a laptop, you're losing most of your audience.

10. The Business Model Is Still Open

I haven't cracked monetization yet. Affiliate revenue is inconsistent. Sponsored listings risk trust.

Right now I'm optimizing for traffic and trust. Build value first. Figure out money later.

What's the hardest lesson you've learned building something? Drop it in the comments.

I curate AI tools at toolsdepth.com — 200+ tools, updated weekly.

How I Built a Review Site with 800+ Articles Using AI

武乐丹 — Sat, 23 May 2026 04:35:47 +0000

How I Built a Review Site with 800+ Articles Using AI

The stack, the workflow, and what actually worked

A few months ago, I wanted to build a review site for Chinese consumer brands — products like GaN chargers, USB-C hubs, smart home devices, and laptops that are popular in Asia but don't get much coverage in English-language tech blogs.

The goal was simple: produce useful, data-driven reviews at scale. No clickbait, no affiliate-first garbage. Just honest comparisons with real specs and real user feedback.

Here's how I built it, what the workflow looks like, and what I learned along the way.

1. The Stack

The site runs on a minimal stack:

Next.js (static export) — fast builds, great DX
Decap CMS — Git-based CMS so editors don't need a database
Vercel — free hosting, instant rollbacks
GitHub — content and code in one repo

No backend to manage. Every article is a Markdown file in the repo. Decap CMS gives the content team a nice UI on top, but the source of truth is Git.

2. Why AI for Content?

I didn't want to build yet another AI-generated content mill. The approach was different:

Research first: AI gathers product specs, pricing, and real user reviews from multiple sources
Human structure: Each article follows a template (overview → specs → performance → real user feedback → verdict)
Data verification: Pricing and specs are checked against official sources before publishing
Images are real: No AI-generated product images. Every photo comes from official brand sites or verified listings

The AI handles the heavy lifting — research, formatting, translation of Chinese reviews — while humans control the quality bar.

3. The Content Pipeline

Here's the actual workflow for each article:

Product selection — Identify trending products on JD.com, Taobao, and Tmall
Spec collection — Pull official specs from brand sites and verified product pages
User review aggregation — Collect real buyer feedback (good and bad) from verified purchasers
Article generation — Structure everything into a consistent, readable format
Review pass — Check all specs against official sources, verify pricing, remove any hallucinated claims
Image sourcing — Download official product images from brand sites (no guessing CDN URLs)
Publish — Commit to Git, Vercel deploys automatically

4. What Actually Worked

✓ Real data beats SEO tricks

The articles that perform best aren't the ones with keyword-stuffed titles. They're the ones with actual benchmarks and real user experiences. A USB-C cable buying guide with measured charging speeds and compatibility testing gets more engagement than any generic listicle.

✓ Consistency matters more than perfection

Publishing 3-5 articles daily (focused, well-researched ones) built organic traffic faster than trying to write one perfect article per week. Search engines reward freshness.

✓ User reviews are gold

Chinese e-commerce platforms have incredibly detailed review systems, often with photos. Translating and aggregating authentic user feedback gives articles depth that pure spec sheets can't match.

5. What Didn't Work

✗ Pure AI generation

Early tests with full AI generation produced articles that looked good but lacked depth. They'd say "great product" without explaining why. The fix was adding real user quotes and verified test data.

✗ Guessing CDN image URLs

We tried building an automated image pipeline using pattern-matched CDN URLs. It failed constantly. The solution was going back to sourcing images manually from official brand sites.

✗ Over-optimizing for search

The first batch of articles tried too hard to match search patterns. They read like SEO sludge. The fix was writing for humans first and treating keywords as a secondary concern.

6. Numbers After 800 Articles

800+ published articles across 19 categories
100% real product images (every article has at least one genuine photo)
Organic traffic growing steadily since launch
AdSense approved and running

Not explosive growth, but steady, sustainable progress.

7. Key Takeaways

AI is a multiplier, not a replacement — The best results come from AI handling research and structure while humans handle verification
Content quality is a flywheel — Good content attracts better readers, which attracts better engagement, which signals quality to search engines
Don't skip the boring parts — Checking specs against official sources, sourcing real images, and verifying user reviews takes time but builds trust
Ship fast, iterate faster — Get the first 50 articles up, then improve based on what the data tells you

If you're building something similar or have questions about the workflow, drop a comment below. Happy to share more details about specific parts of the pipeline.