Most competitor research starts with price, images, and ranking. All three are useful, and all three are already public β€” which means everyone doing the same research sees the same picture and reaches the same conclusion.

Customer reviews are the opposite. They are public too, but almost nobody reads them systematically. A few hundred reviews carry more product-level information than any other source on the listing: what actually breaks, when it breaks, and which promise the seller could not deliver.

This article walks through the workflow we use to turn that raw text into a decision.

What you are actually dealing with

A typical category with two meaningful competitors will have 300 to 500 text reviews between them. In the pet water fountain category, for example, our latest analysis covered 478 reviews across two products β€” 236 on one, 242 on the other.

Reading that by hand takes hours. At roughly three hours per thousand reviews, you are looking at a full working day for a single category, before you have categorized anything, compared anything, or counted anything.

There is also a floor worth respecting: below about 100 text reviews for a product, patterns stop being reliable. 40% of 50 reviews and 40% of 5,000 reviews are not the same finding β€” the first one is closer to a coin flip.

The workflow

1. Pull the reviews. The official API path is effectively closed for full review sets. Tools like SellerSprite work for sampling; anything closer to complete requires your own collection setup, and the hard part is not the collection but the rate limiting β€” raise the frequency and you meet the CAPTCHA wall.

2. Label by pain point. A few hundred reviews is past the point where a human can hold the structure in their head. Hand them to a model with a fixed taxonomy β€” durability, fit, sound, connectivity, support, and so on β€” and let one review carry more than one label.

The one mistake that costs the most time here: letting the model invent its own categories each run. Two runs, two different category sets, and no way to compare them. Fix the taxonomy first. Everything downstream depends on it.

3. Cluster and compare. Once labels exist, the rest is counting: how many reviews mention each pain point, what share of the total each represents, and how the two products line up side by side.

This is where the useful findings show up. In the wired earbuds case, Amazon Basics ($9.99, 174 reviews) and LUDOS ($12.99, 215 reviews) both sat at exactly 3.5 stars with near-identical negative ratios (30.5% and 31.6%). Side by side they looked interchangeable.

They were not. Amazon Basics' top complaint was short lifespan β€” 23.3% of its pain-point mentions, against 1.3% for LUDOS. LUDOS' top complaint was the earbuds falling out β€” 26.8%, against 4.1% for Basics. Same score, opposite failure modes.

4. Classify the failure. Group each pain point by type β€” promised feature not delivered, basic experience collapsed, quality control over time, service that does not back the product. The classification is what turns a list of complaints into a decision about what to build and what to avoid.

Three things that quietly ruin the analysis

Rating and review text are not the same pool. The headline star rating is a weighted aggregate of all ratings, including people who never wrote a word. The text reviews you can analyze are a separate and much smaller set. The two will not reconcile, and that is normal. Settle the definition before anyone argues about the number.

Incentivized and fake negatives need their own bucket. Real complaints carry specific usage detail β€” how long it was used, how it failed. Fake ones tend to be vague and describe problems that do not match the product. It becomes obvious after a few dozen reviews, but you have to look for it deliberately.

The model will misclassify. Expect it. Spot-check a sample against the raw text before acting on any conclusion. We have had a batch come out with the entire direction wrong, and only caught it by pulling the underlying reviews.

What the output is worth

Here is the same method applied to the pet water fountain category. Two premium fountains, 478 reviews, category average 3.2 stars.

Neareal (236 reviews, $29.99, 3.5 stars) carries a 23.3% negative ratio. Its dominant complaint is the pump failing within weeks β€” 17.5% of its pain-point mentions. Its marketing leans on stainless steel construction, and reviewers pushed back on that too, noting that only the top is steel.

PETLIBRO Dockstream 2 (242 reviews, $19.98, 3.0 stars) carries a 35.5% negative ratio. Its dominant complaints are different: the unit stops working within a year (10.3%), support that does not respond (10.3%), and water leaking from the base (9.9%).

On dimension scores out of 10, they land at 2.5 and 2.3. Neither product scores above 4 on any single dimension except water flow.

That is the finding. The category is not divided into a good product and a bad one. It is a category where nobody has solved hygiene and pump reliability, and where the more feature-loaded product added new failure modes instead of fixing the old ones β€” app connectivity, server outages, subscription friction on advertised features.

For someone entering that category, the decision writes itself: the gap is not more features. It is a fountain that comes apart completely, has no unreachable crevices, and ships with a pump that is either over-engineered or warrantied without argument.

What this method does not do

A high share of mentions does not mean a large share of impact. A pain point that drives few returns can still dominate the review text. Cross-check the clusters against return data and cost before committing β€” analysis marks where the traps are, but pricing and target audience decide whether a trap matters.

Reviews also lag the market. A problem trending upward in review text over the last two months is a leading signal; one that dominated two years ago and has since disappeared may already be fixed.

Where this goes

Everything above is mechanical once the taxonomy is fixed, and it is the part we automated. Choicesages collects the reviews, clusters the pain points, scores each product across dimensions, and produces a full PDF report β€” the same structure used in the examples here, across 14 sections and two languages.

Reports are per-category and cross-brand, which is what makes the comparison meaningful in the first place. If you want to see the format on a category you are evaluating, there are public samples on the reports page.