Amazon New Product Research Automation: What AI Should Do

Before deciding what to automate, list the steps. Most research follows a similar sequence, whatever tool you use.

• Idea generation: finding categories and keywords worth checking.
• Demand estimation: search volume, sales estimates and seasonality.
• Competition review: number of sellers, review counts, listing quality and brand concentration.
• Customer insight: what buyers complain about and what they wish existed.
• Cost modeling: landed cost, FBA fees, referral fees, ad cost and margin.
• Supplier evaluation: samples, factory audits, certifications and lead times.
• Decision: go, change the product, or drop it.

The first four steps involve large amounts of data and suit automation. The last three depend on information no tool fully holds.

Third-party suites such as Helium 10, Jungle Scout and SmartScout estimate sales and search volume from their own models. Amazon itself provides Product Opportunity Explorer in Seller Central, which groups search terms into niches and shows search volume trends, click share and return rates where available. Brand Analytics adds actual search term rankings for brand owners.

Treat third-party sales estimates as directional. They are modeled, not reported by Amazon, and different tools often disagree. Product Opportunity Explorer data comes from Amazon but is aggregated into niches whose boundaries may not match your product idea.

Large language models are good at reading and summarizing text at scale. That makes them useful for the customer insight step, which used to take days of manual reading.

Review mining

Export reviews from the top 20 competing ASINs, through a tool or by collecting them manually within the platform’s terms, and ask a model to group complaints by theme. A typical prompt asks for the ten most frequent problems, with a count and a sample quote for each. You get a list such as “lid cracks after a few months”, “too heavy to carry” or “instructions unclear”, which points to product changes.

Keyword clustering

Paste a few hundred search terms from Product Opportunity Explorer and ask the model to cluster them by intent: use case, audience, feature, price signal. Clusters show whether a niche has distinct sub-markets, for example “for travel” versus “for home”, that a single listing cannot serve well.

Listing comparison

Give the model the titles and bullets of the top ten listings and ask what claims they all share and what none of them make. Gaps in messaging are not the same as gaps in the market, but they are a cheap starting point for differentiation.

Automation produces confident output from weak inputs, and product research is full of weak inputs. Watch for these failure points.

• Sales estimates that look precise but come from models with wide error.
• Review summaries that overweight a vocal minority; always check the count behind each theme.
• AI-generated product ideas that ignore manufacturing reality, such as features that would double unit cost.
• Opportunity scores that combine metrics in ways you cannot inspect.
• Seasonal niches that look strong because the data window captured only the peak months.
• Restricted or gated categories that tools do not flag, such as certain topicals or toys requiring testing.

A model can also invent details when asked about a specific ASIN or market it has no data on. Give it the data and ask it to work only from what you supplied.

Three steps resist automation. The first is supplier evaluation. A factory’s real quality shows up in physical samples, communication speed and willingness to fix problems, none of which appear in a spreadsheet. Order samples from at least three suppliers and test them yourself.

The second is brand fit. A product can score well in every tool and still be wrong for your brand, your capital or your team’s skills. A brand built on eco-friendly kitchenware should think twice before adding phone accessories, even if the numbers look good.

The third is the final cost model. Tools calculate FBA fees well, but you need to add realistic ad costs for launch, return rates for the category, storage fees for slow months and the cash tied up in inventory. A product with a healthy margin on paper can lose money in the first six months once launch ad spend is included.

Here is an example workflow for a small team checking a dozen ideas a month.

• Week 1: run Product Opportunity Explorer and a third-party tool on each idea; drop anything with falling search trends or extreme brand concentration.
• Week 1: for the remaining ideas, export reviews of top sellers and run AI clustering on complaints.
• Week 2: a person reads the clusters and decides which complaints a product change could fix at reasonable cost.
• Week 2: build the cost model, including a launch ad budget and a conservative conversion rate.
• Week 3: order samples for the two or three ideas that survive.
• Week 4: test samples and decide.

Automation shortens weeks one and two from days of manual work to hours. It does not shorten weeks three and four, and it should not.

Keep a research log for each idea, recording the data source, date pulled and the conclusion drawn. When a product launches, compare actual results to the research estimate. Over a few launches you learn which tools overestimate in your categories and by how much, and you adjust your thresholds.

It also helps to write down the reason you dropped each idea. Six months later, when a tool flags the same niche again, you can see whether the original reason still applies.