Most product research starts the same way: find a promising product, then justify it.
That order deceives you systematically. Once you have settled on a product, every later step quietly recruits evidence for it — heavy competition becomes "proof of demand", thin margin becomes "we'll make it on volume", low ratings become "room to improve".
Turn it around. Product research is an elimination process. You are not looking for the right product; you are removing the overwhelming majority of impossible options as cheaply as you can.
Four layers, each one an elimination
| Layer | Question | What it removes | Cost |
|---|---|---|---|
| 1. Category | Does this category still have room for a newcomer | A whole category | Lowest |
| 2. Price band | Where is demand underserved | Most of a category | Low |
| 3. Product | Which specific listings are attackable | Nearly all products | Medium |
| 4. You | Whether you can actually do it | Most of what is left | Highest |
The order is the cost order. Eliminating a category at layer one takes one query; eliminating a candidate at layer four may already have cost you two weeks of samples and quotes. Running it backwards means making the cheapest judgments in the most expensive way.
Layer one: eliminate categories before finding products
The most skipped layer, and the one with the highest elimination rate.
The test is not "is this category big" but does this category still leave room for a new listing. The listing age endpoint groups by listing age and returns each band's share of units; the documented unitsRatio for the "over 3 years" band is 0.834 — 83.4% of that category's units go to listings older than three years.
A large total and a share you can reach are two different things. A category doing a million units where incumbents take everything is harder than one doing a hundred thousand with a loose structure.
Four distributions cover it: listing age, seller type, price, and the category's own brand and seller concentration. Worked through in Amazon category analysis. Concentration itself, and why unit share and revenue share diverge, is in Amazon market analysis.
Layer two: find the price gap inside the category
The price distribution endpoint returns product counts and unit share per band. You are not looking for the highest or the lowest, but for where the two disagree:
- Many products, low share of units → the band is crowded with things that do not sell
- Few products, high share of units → demand is here and supply has not arrived
The second is the gap. It usually sits in an unremarkable middle band, because both ends have been tried repeatedly.
Layer three: only now, specific products
By here the range is narrow enough to apply filters. The product research endpoint exposes close to twenty min/max conditions — demand floors, competition ceilings, margin bands, complexity, excluding large brands.
What matters at this layer is not filtering precisely, it is writing the standard down. Three months later you should be able to say what you were gating on, and a colleague should be able to reuse the conditions.
- Chaining three endpoints into a scheduled job: Bulk product research
- Getting candidates into a sheet two people can compare: The product research sheet
- Extension, SaaS or API: Choosing a product research tool
- What to redo when switching to Japan or Europe: product research across marketplaces
Layer four: the part data cannot answer
The first three layers are answerable with data. This one is not:
- Whether you can source it, and at what cost
- Whether there is patent or certification exposure
- Whether your brand, warehousing and cash cycle fit
It goes last because it is the most expensive — and because it is the most expensive, the whole point of the first three layers is to keep the number of candidates reaching it small. A common waste: taking thirty candidates to suppliers when twenty-five should have been eliminated at layer one.
Three ways the order gets reversed
Starting from "what I want to sell." That is not research, it is justification. Preferences are allowed — they belong at layer four, not layer one.
Starting from a bestseller list. Everything on that list has already won; it describes past outcomes, not current gaps. Competitors belong at layer three as calibration, not at layer one as a starting point.
Starting from "high monthly sales." Sales figures are estimates — the marketplace does not publish unit sales, so every monthly number is derived from BSR. Fine for ordering magnitudes, and using it as your only gate walks you straight into the crowded end. Which fields are facts and which are derived: What Amazon sales data can and cannot tell you.
When to stop
Data's job is narrowing thousands of possibilities to dozens. It is not making the final call.
When the differences between remaining candidates are no longer in the data, stop querying. More queries will not make you more certain, only later. What reduces uncertainty at that point is samples, a small first order, and a conversation with a supplier — not running the same question through one more tool.
Questions
Where should a first-timer start? Layer one, still. The most common beginner mistake is starting at layer three or four — having a product idea first and going back for data that supports it.
How long does this take? The first two layers run about fifteen minutes per category, because both are aggregate data. Layer three depends on candidate count. Layer four is what actually consumes time, which is precisely why the earlier layers exist.
Isn't eliminating whole categories arbitrary? It costs you some false negatives, but the costs are asymmetric. Missing a workable category costs opportunity; entering one where positions are locked costs real inventory and ad spend.
Do conclusions carry across marketplaces?
Not directly. Price bands, competitive intensity and seller structure are independent per marketplace, and every endpoint takes a marketplace parameter. One category can yield opposite conclusions in two marketplaces.