The AI Amazon Research Agent: How I Size a Niche Before Committing a Penny of Inventory

What this used to cost

Properly researching one Amazon niche took me the better part of two days. Reverse-engineering competitor ASINs into a keyword set, exporting search volume, cross-checking BSR history so a Prime Day spike did not get mistaken for demand, reading the one-star reviews of the top ten listings to find the gap, then modelling landed cost against the price band the category actually sells at. Two days for a yes or a no on a single product.

So it did not get done properly. Not by me, and not by most brands I work with. It got done on a Tuesday afternoon between other things, on three tabs and a gut feeling, and the decision that came out of it committed six figures of inventory and nine months of cash. Rushing this stage is the most common reason new Amazon businesses fail in their first six months, and it is not because operators are lazy. It is because the honest version of the work costs more time than anyone has, so everyone quietly does the dishonest version.

What the agent actually does

I give it one of three things: a category, a keyword, or a competitor ASIN. That is the whole input. What comes back is a decision brief.

It starts by running the niche through DataDive, which is the research layer I chose because it aggregates Brand Analytics, Keepa and estimate data into one cleaned keyword list rather than making me reconcile three exports by hand. The agent pulls the reverse ASIN keyword set for the top listings, strips variants so the results are not inflated, and ranks by relevance to actual sales rather than raw volume. Volume is the number that seduces people. Relevance is the number that pays.

Then it reads BSR as a trajectory rather than a snapshot. A single BSR reading tells you almost nothing. Three to six months of steady or improving rank tells you there is a market. A vertical spike three weeks ago tells you somebody ran a promotion. The agent also checks subcategory rank alongside main category rank, because subcategory is where you find out who you are actually fighting.

Next it profiles the competitors. Review depth on the top ten, how long each has held position, price band, rating distribution, and the recurring complaint themes in their negative reviews. That last one is where product gaps hide. A category where every top listing gets hammered for the same flaw is a category with a door in it.

Finally it does the arithmetic that most research skips. Achievable price against landed cost, Amazon fees, and the advertising cost of sale needed to break into that keyword set. Then it scores the opportunity and, critically, it writes the case against. Every brief ends with the strongest argument for not doing this. An agent that only ever finds opportunities is a very expensive yes-man.

The context layer that makes it useful

A generic research tool returns data about a market. This agent returns a decision about a market for me specifically, and the difference is entirely in what it remembers. It holds my minimum contribution margin, my typical landed cost structure, the categories I have already lost money in and why, how long I am willing to fund an unprofitable ranking period, and the fact that I would rather own a smaller category outright than rent a slot in a large one.

When I run it for a brand I work with, that memory layer is theirs instead of mine. Their cash position, their manufacturing lead times, the retailers they are already in and therefore cannot undercut, the claims their category is not allowed to make. The agent stops being a research tool at that point and starts behaving like a category manager who has read every brief you have ever written. That is the part you cannot buy off a shelf, and it is the part that took the longest to build.

What the output looks like

One page, four blocks. Verdict at the top with a confidence level and one sentence of reasoning, so I know in five seconds whether to keep reading. Then the demand picture: the keyword set worth ranking for, volume bands, seasonality, and whether the trend line is going up or quietly rolling over. Then the competitive picture: the top listings with review counts, how entrenched they are, price band, and the recurring complaint that represents the gap. Then the money: achievable price, estimated landed cost, contribution margin at that price, and the advertising spend required to break in.

Underneath, the case against. In plain English. Something like: demand is real and growing, but the top three listings average 2,800 reviews and have held rank for eighteen months, so you are buying twelve months of unprofitable advertising before this pays, and your cash cannot fund that alongside the Q4 restock. Walk away.

What this looks like in practice

A wellness brand I work with wanted to extend into an adjacent category their customers kept asking about. The demand was obviously there. On volume alone it looked like an easy yes, and they had already had a quote from a manufacturer.

The brief came back with demand confirmed and a hard case against entering at the size they planned. The top listings were entrenched, the achievable price sat below the level their landed cost needed, and the advertising cost to crack the head terms would have eaten the first year. What it did find was a subcategory two steps sideways where the top listings averaged a few hundred reviews and every negative review complained about the same format problem their existing manufacturing could solve. They launched there instead. The whole redirection came out of one research run, on a decision they were three weeks from making the other way.

Inside the system

How we build this for brands

The research agent rarely runs alone. It sits next to the VOC engine, which mines reviews and support messages across every channel a brand appears on, so the complaint themes in a competitor's one-star reviews arrive alongside the language a brand's own customers use. Between them you get both halves of the picture: where the gap in the category is, and the exact words that will sell into it once you are there.

Downstream, the same output feeds the profit and cash-flow dashboards we build from live Shopify and ad data, so an inventory commitment is checked against actual cash rather than an optimistic forecast, and a reporting agent surfaces the moment a launch starts drifting from its assumptions. Part of this runs live for portfolio brands today; the full system is what we deploy when we take a brand on.

Frequently asked questions