Fetch reviews for a target product from the merchant's review API (if available) or from the product page HTML; collect a representative sample across rating levels, not just the top-displayed reviews.
Deduplicate reviews by normalized text to remove duplicates that can arise when syndicated reviews appear across merchants.
Cluster reviews by topic (e.g., build quality, battery life, customer service, shipping) using keyword extraction or a short classification prompt; compute sentiment distribution per topic.
Produce a structured summary: overall sentiment, key positive themes with example quotes, key negative themes with example quotes, and notable one-off issues.
Flag aspects that are dealbreakers for the user's stated use case: e.g., if the user needs the product for outdoor use, surface any reviews mentioning weatherproofing failures.
Include metadata: number of reviews analyzed, date range, average rating, and whether the review corpus is verified purchases only.
Known gotchas
Review counts and ratings from merchant pages can be inflated by incentivized or fake reviews; prefer verified purchase reviews and apply skepticism to suspiciously uniform 5-star distributions.
Reviews are often written for a product family; a negative review may describe a previous model's flaw that has been fixed — check review dates relative to product release dates.
Summarizing reviews with an LLM can hallucinate specific quotes; always ground summary quotes in the actual fetched review text and include a source link.
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp