Every product team wants qualitative feedback and most pay a research agency for it. Meanwhile a few thousand customers have already written down exactly what they think, in public, for free.
Marketplace reviews are the cheapest customer research available and the most consistently ignored.
Why they get ignored
Because reading them does not scale. Four hundred reviews across nine listings is a day nobody has, and skimming produces confirmation bias — you notice the complaints you already suspected.
Pulling them programmatically and summarising by rating band removes both problems.
The structure that makes them useful
Group by star rating first. This is the whole method. One and two-star reviews are where the product problems are. Five-star reviews are where the purchase drivers are. Three and four-star reviews are mostly logistics complaints wearing a product costume, and they will mislead you if you pool everything together.
Summarise the recurring complaints in the low band. Recurring is the operative word — one furious review about a damaged box is a courier problem, forty reviews mentioning the cap leaking is a packaging defect.
Summarise the recurring praise in the high band. This is the more commercially useful half and the one teams skip. What buyers actually valued is rarely what the brief said the product was for, and that gap is your positioning.
Quote, do not only paraphrase
Pull a few representative lines verbatim into the output. A summary that says "buyers mentioned scent" is worth nothing next to three actual sentences in a customer's own words. Verbatim quotes survive the journey into a slide; paraphrases get softened at every step until they say nothing.
And record how many reviews were read. A pattern across 300 reviews is a finding; the same pattern across 11 is an anecdote, and the reader deserves to know which they are looking at.
Running it
Submit the product URLs, poll until complete, then group and summarise. Lazada reviews are available through a dedicated endpoint at 6 credits per request item.
One practical warning: Lazada endpoints each validate their own URL pattern, and a URL accepted by one is often rejected by another. If a submission is refused, read the error and try the pattern it names rather than assuming the endpoint is broken. This trips up more first integrations than any other detail.
Where it gets genuinely valuable
Run it on a competitor's product rather than your own.
Their one-star reviews are a list of unmet needs in a category you are already in, written by people who bought the alternative. That is a product brief someone else paid to generate, and it is sitting in public.
Pair it with the panel to pick targets: rank the category by GMV, take the top twenty SKUs, and mine the reviews of the ones that are not yours. You are then reading complaints about the products that actually matter rather than the ones you happened to think of.
Recipe 2.7 in the agent skill has the prompt.