Search for the best Amazon product research tool and you get listicles. The problem with them is not only that you cannot verify them — it is that they compare the wrong axis. Nearly every product research tool draws on similar sources with similar semantics, so the real difference lies elsewhere.
The difference is this: how much of the judgment does the tool make for you?
Three shapes, with the judgment in different places
| Browser extension | SaaS database | Data API | |
|---|---|---|---|
| How you use it | Whatever page you are on | Filter a database by criteria | Send your own requests |
| Who judges | The tool (it hands you a score) | Half — presets are adjustable | Entirely you |
| Suits | Checking one product ad hoc | Routine bulk screening | Fixed pipelines, scheduled jobs |
| Ceiling | One page at a time | The filters the platform exposes | Any condition you can express |
| Reproducible | No | Partly | Yes — same parameters, same result |
That last row gets overlooked, and it decides whether anyone else can check your conclusion. The number you saw in an extension cannot be reproduced a week later, even by you.
An "opportunity score" is a set of undisclosed weights
Most research tools attach a composite score to each product — opportunity score, potential, competition index. Different names, same construction: demand, competition and margin metrics combined into one number under some set of weights.
The number is not the problem. The problem is that the weights are usually not published.
And the weights decide what it recommends. Weighting toward "low competition" pushes you into niches; weighting toward "high demand" pushes you into crowded categories. When two tools rank the same category completely differently, it is usually not because the data differs — it is because the weights differ.
You can check this yourself: run the same category through two tools and compare the ordering. If the ordering diverges while the underlying fields (price, BSR, rating count) agree, the disagreement lives in the weights.
Someone else's weights are not your strategy. A seller with an established brand and a seller just starting out have completely different tolerance for competition, and both are shown the same score.
The alternative: write the criteria down as your own
Drop the score and you have to define the conditions yourself. The product research endpoint exposes close to twenty min/max filter pairs, which covers most strategies:
| What you are gating | Parameters |
|---|---|
| Demand floor | minUnits / maxUnits, minRevenue / maxRevenue, minBsr / maxBsr |
| Competition ceiling | maxSellers, minRatings / maxRatings, minLqs / maxLqs |
| Margin band | minPrice / maxPrice, minProfit / maxProfit |
| Complexity | minVariations / maxVariations, minWeights / maxWeights |
| Exclusions | excludeBrands, excludeSellers, excludeKeywords |
| Badge filters | badgeAC, badgeBS, badgeNR, fulfillment |
The value here is not sharper filtering. It is that your standard becomes explicit. Three months later you know what you were gating on; a colleague can reuse the conditions directly; and adjusting strategy means changing one threshold rather than switching tools.
Exclusions like excludeBrands and excludeSellers earn their keep especially — slots held by large brands carry little signal for most sellers, and excluding them beats skipping past them by hand in the results.
An acceptance checklist you can run
Whichever shape you land on, these five questions surface the real differences. Pick a category you know well, run the same conditions through your candidates, and work down the list.
1. Where does "monthly sales" come from? The marketplace does not publish unit sales, so every monthly figure is derived from BSR. A tool should say so. Presenting an estimate as if it were a fact is a signal. What Amazon sales data can and cannot tell you covers this in full.
2. Which category level is it using? Category paths have several levels, and top-level rank differs enormously from subcategory rank — yet interfaces often just say "rank." When one product ranks differently in two tools, check this first.
3. Which marketplaces are covered? Working in the US implies nothing about the marketplace you actually sell in. Verify this separately rather than trusting the landing page.
4. When was the data measured? Data without a timestamp cannot support a claim about a trend. If the interface does not show one, ask whether the export carries it.
5. Can you export it, and can you integrate it? After filtering down to 200 candidates, the next step is always exporting into your own sheet for a second pass. A tool you can only read on screen ends here.
Point 5 is usually where tools give way to an API. The real signal is not that the tool is bad — it is that you have started repeating yourself: the same filters weekly, the same fields exported, the same formula applied. At that point, Bulk product research: three endpoints chained into a candidate list covers turning it into a scheduled job.
When you do not need a tool at all
Two cases:
When there are few candidates. With three to five products to judge, reading the pages plus a research sheet is enough; setting up a tool costs more time than it saves.
When the deciding factor is not in the data. Whether your supply chain can source it, whether there is patent exposure, whether it fits your brand — these often matter more than any metric, and no tool answers them.
Data's job is narrowing candidates from thousands to dozens, not making the final call for you.
Questions
Are free research tools enough? It depends where you get stuck. For a rough read on a single product they usually are. Once you reach bulk screening and periodic re-checks, the limits show up in exports and call volume rather than in the data.
How much disagreement between tools is normal? Public fields like price, BSR and rating count should agree; disagreement there is a semantics problem, most often category level. Units and revenue are estimates, so a close order of magnitude is fine and exact agreement is not the goal.
Can I just use one tool? Yes, as long as you know its semantics. The value of cross-checking two is not averaging them — it is that a discrepancy makes you investigate, and the cause you find is usually more useful than either number.
The extension shows different data than the API returns. Align three things first: marketplace, category level, and time of measurement. Only a difference that survives all three is a semantics difference.