Let me cut to the chase: I’ve tested like 10+ of those “AI product research” tools, and 90% of them are straight garbage. They just spit out a random list of keywords and ASINs, throw in a pretty graph, and call it “research” — zero actionable insight, zero actual strategy. I wasted so much time (and money on AI tokens) messing with them that I finally said “screw it” and built my own workflow with Claude Code. Took a ton of trial and error, but now it actually helps us find real supply-demand gaps — not the fake “low competition” nonsense everyone peddles.

Full disclosure: I’m not selling anything here. No course, no tool, no affiliate links. Just sharing the methodology because I wish someone had done this for me 6 months ago. You can replicate this with whatever data source you already use — Keepa, Jungle Scout, Helium 10, even manual CSV exports. No fancy tools required, just a little patience and a willingness to stop letting AI do all the thinking for you.

The core problem with most “AI” product research (rant incoming)

Here’s the thing: Most of these AI tools treat you like an idiot. They give you a list of potential products, but they never answer the questions that actually matter for your business:

Why the hell is this niche worth entering? (Not just “low competition” — actual reason)

Which feature combinations are peoplebegging for but no one’s selling?

What are customers actually complaining about with existing products?

So we built our own workflow around a few simple principles — no fluff, just stuff that actually works.

Principle 1: One unified data source (no mixing tools, I beg you)

If you’re pulling data from 3 different tools (Jungle Scout for search volume, Keepa for BSR, Helium 10 for reviews), you’re gonna get inconsistent trends and AI hallucinations out the wazoo. Trust me — I tried that, and I ended up chasing a “gap” that didn’t even exist because the data was conflicting.

Pick one source for your core analysis. We use a combo of:

Keepa API (or just manual CSV exports if you’re cheap like me) for BSR, price history, and rankings

Basic product page scraping (we use Playwright, super easy to set up) for titles, bullet points, images, and reviews

No mixing tools unless you’re 100% sure you understand their biases. Consistency = no AI lies.

Principle 2: Full Top 100 tagging — no shortcuts (yes, it’s a pain)

Most people look at the top 10 or 20 ASINs and call it a day. That’s lazy, and it’s why you end up with the same saturated products everyone else is selling.

We tag every single product in the Top 100 of a category. Not just “headphones” — we tag them by:

Form factor (over-ear, on-ear, earbuds — duh)

Key features (noise canceling, waterproof, battery life — the stuff people actually care about)

Use case (sports, office, travel — who’s buying this?)

Connectivity (Bluetooth, wired, hybrid — basic but critical)

You can do this manually in a spreadsheet (painful, I know) or use AI with a structured prompt. The key is 100% coverage — no sampling, no shortcuts. We use regex to auto-tag from titles (~70-80% accuracy), then manually verify the rest using product detail pages. It takes time, but it’s the only way to get reliable data for cross-analysis.

Principle 3: Cross-analysis to find real supply-demand gaps (AI’s actual superpower)

This is where AI actually earns its keep — not just spitting out keywords, but finding patterns you’d miss. Once you have all 100 products tagged, you can ask the AI to do pairwise cross-analysis — like, “Show me all combinations of [form factor] + [use case] and calculate supply (how many products exist) vs demand (search volume or estimated sales).”

It’ll spit out a matrix, and the gaps (high demand, low supply) are your opportunities. But here’s the catch: not every blank is a real gap. Sometimes the combination doesn’t exist because it’s technically impossible (like a waterproof wireless charger that fits in your pocket) or because no one actually searches for it. So we force the AI to explain the reason for every gap it flags. If it can’t give a plausible reason? We ignore it. No exceptions.

Principle 4: Hard rules to prevent AI sloppiness (trust me, you need this)

Without rules, AI will cut corners. It’ll say “here are 20 representative products” instead of giving you all 100. It’ll skip the boring but necessary validation steps. It’ll make up data if you let it. I learned this the hard way when AI invented a “high search volume” keyword that didn’t exist — wasted 2 weeks chasing it.

We built a list of 15 hard rules into our system. A few that saved us the most headaches:

Always output full Top 100, no abbreviations — no “etc.”, no “similar products”, no shortcuts. If I’m gonna put in the work, I want the full picture.

Every data table must have a “key insight” paragraph — don’t just show me numbers, tell me what they mean. “50% of products have <10hr battery life” → “Opportunity to launch a long-battery option for frequent travelers.”

For each supply gap, include a reason — why is this combination missing? Technical barrier? No demand? Did everyone just overlook it?

Include an executive summary — 3-5 bullet points with “data → implication → action”. I don’t have time to read 10 pages of fluff when I’m making a decision.

Every conclusion must be traceable to a specific data source — no AI guessing. If it says “search volume is 10k/month”, make it show the keyword and the tool it came from (Keepa, JS, etc.).

These rules aren’t optional. They turn a useless AI toy into something you can actually use to make real business decisions.

What the final output looks like (no fancy stuff, just useful)

We ask the AI to deliver three things — nothing more, nothing less. It keeps our team on the same page:

A detailed markdown report — Executive summary, market structure, competitor analysis, supply-demand matrix, and prioritized opportunities. Each opportunity includes a rough product spec (dimensions, features, target price), why it’s better than existing options, and estimated monthly sales potential. No fluff, just straight to the point.

An Excel file — Raw data with multiple sheets (Top 100 with tags, search term volumes, review analysis). So my team can audit the AI’s work — because let’s be real, AI makes mistakes.

A simple HTML dashboard — For quick visual overview: KPI cards, product distribution charts, gap analysis, and go/no-go scores. My manager hates reading long reports, so this lets him get the gist in 2 minutes.

This way, the person doing deep work (me) has the markdown, the manager has the Excel, and the leadership team gets the dashboard. Everyone gets what they need, no drama.

A few things we learned the hard way (save yourself the pain)

AI will hallucinate if you let it. Always ask for source citations. If it says “search volume is high”, make it show the keyword and the tool. I can’t stress this enough — I wasted so much time on AI lies.

Manual verification is still necessary. We trust AI for pattern recognition, not for final decisions. Any product idea we like goes through a human sanity check — does this actually make sense? Would I buy it?

You don’t need expensive tools. We started with Keepa exports and manual scraping. It’s slower, but it works. The AI just speeds up the analysis — you don’t need to drop $500/month on tools to make this work.

The output is only as good as your tags. Spend time on getting the attribute tagging right. Garbage in, garbage out — if your tags are lazy, your gaps will be fake.

Final thought (no fluff, just real talk)

I’m not claiming this is the perfect system — it’s still evolving. We tweak the rules every month, and we still make mistakes. But it’s moved us from “randomly guessing products and hoping for the best” to “making data-informed decisions that actually pay off.”

If you’re tired of AI tools that just spit out keyword lists and waste your time, try building your own workflow around these principles. It’s not easy, but it’s worth it.

Curious if others have tried similar approaches — or if you’ve found better ways to handle attribute tagging or supply-demand analysis. Would love to hear what’s working (or not) for you. Drop a comment below — I’ll reply to every one (no bots, promise).