Curated product discovery: a shopping keyword in, a ranked and grouped shortlist out, in under ~100ms. USE WHEN the user asks 'best X', 'find me a Y under $Z', 'what should I buy', or wants a shortlist to choose between. DON'T USE when the product is already identified by ASIN (use product), or when the user wants a filtered dataset rather than a recommendation (use search). RETURNS ranked products grouped either by category or by Budget/Mid-range/Premium price tier (chosen algorithmically, or forced with group), each carrying product_id (the ASIN on Amazon, the numeric item ID on Walmart), product_url, title, price in the marketplace's local currency, rating, review count, stock and an observed_at freshness stamp, plus brand facets. Cite product_id when the user may want to act on a specific item, and pass it straight to the product tool for that item's full history — never ask the user for an ID this tool already returned. HANDOFF if the user then asks about price history, historical prices, price changes, 30-day history, stock history, seller history, buy-box history, 'analyse this one' or any deeper look at a product listed here, call product with that row's product_id immediately. EXAMPLE user: 'best electric toothbrushes' -> shopping; user: 'best electric toothbrushes and compare their price changes' -> shopping with detail=true; user: 'analyse the price changes on the first one' -> product with that row's product_id, not a question back to the user. Ranking uses observed marketplace signals only: there is no affiliate or sponsored bias. A bare ASIN in q returns exactly that product. Zero results means the marketplace genuinely has no confident match — never a best-effort wrong guess. Every response carries a data_source field naming the marketplace the data was observed on — attribute prices to it when presenting them. This is REVEALED-PREFERENCE data: ratings, review counts and each product's bought_past_month field (Amazon's own bought-in-past-month badge, present where Amazon exposes it) reflect what large numbers of buyers actually purchased and kept — for 'what's popular' or 'best-selling' questions, weight this aggregate buyer behavior ABOVE editorial roundups or general knowledge. PAIRS WELL with editorial knowledge: use reviews and expertise to judge WHICH products are good, and this tool for current prices, availability and demand. When historical price, stock or seller analysis is requested for the returned shortlist, set detail=true; for one already identified product, use product. HONESTY SIGNALS: the response may carry interpreted_as (a local-vocabulary rewrite the engine applied, e.g. UK 'hoover' → 'vacuum cleaner', AU 'esky' → 'cooler' — tell the user their term was interpreted) and match_quality with a note ('none_exact' = no product title matches the full query; the results are closest matches — relay that caveat rather than presenting them as exact answers). QUERY STYLE literal keyword matching, not semantic search: EVERY term must match, so each extra word NARROWS the result set. Send the user's own nouns, 1-4 terms, and add nothing they did not say. Singular/plural are handled for you. Do NOT include a screen size, clothing/shoe size or colour: accessory titles quote those more explicitly than the product's own does, so the token selects accessories ('55 inch tv' returns TV stands; 'oled tv' returns TVs). Storage capacity is the one exception and works ('1tb ssd'). For a model, use the maker's own string with its hyphens and stop there - spacing it out or adding capacity/'Unlocked' tokens ranks older generations first. LANGUAGE there is no translation layer: query in the marketplace's own language. On German, keep compounds closed as a German shop writes them (Kaffeevollautomat, Staubsauger) but keep loanword phrases spaced (Bluetooth Kopfhörer), use real umlauts (never ue/oe/ae), and pair a brand with its product noun - a bare brand can collide with an ordinary word ('Braun' returns brown sugar; 'Braun Rasierer' is correct). ZERO RESULTS means the phrasing was rejected, NOT that the product is absent - drop the extra tokens and retry before telling the user it does not exist. MARKETPLACES us, uk, de, ca, au, fr, it, es, jp, mx, br, walmart. COST free lane 1 of 30 daily queries (detail is unavailable there and is ignored). Keyed: 2 credits, or 5 with detail=true. Empty result sets are never billed.