Synchronize the shared Amazon transport, including search interstitial handling. Review access limits and billing behavior are unchanged.
2026-11-13 — Restore previous listing keywords
Restore title, description and SEO text to the November 5 baseline at the owner’s request. Preserve current runtime fixes, input/output contracts, billing, prices and defaults.
Changelog
2026-11-12 — spending safeguards and scheduled product price
Clamp each charge batch to the remaining event budget before calling the SDK, including when concurrent fetches finish after another job spends part of the cap.
The optional product row changes from $0.003 to $0.005 on 2026-11-27 at 00:00 UTC to cover observed source costs; review rows stay $0.005. Current pricing remains active until then. The quick-start example uses the cheaper default proxy route.
Skip product fetches when no requested event fits in the remaining spending budget, including zero-budget runs and queued products after exhaustion.
2026-11-08 — Public review parsing and explicit access failures
Read titles from Amazon's current reviewTitle markup as well as the legacy review cards. A captured UK page verifies five real titles, preserving review IDs, text and all output fields.
Stop a target when Amazon explicitly requires sign-in instead of retrying that authentication boundary as a captcha. The target produces an error and no product/review charge; ordinary transient source failures still retry.
Duplicate ASIN/URL inputs no longer make a run appear successful when every unique product failed.
UK public samples remain available in verified cases. US review access remains restricted; this is not a US recovery or full-history release. Input defaults, prices and billing units are unchanged.
2026-11-07 — Verified UK first-call example
Verified real review text and IDs from two UK product URLs. The input form now prefills the UK iPad URL, a three-review cap, and text-only reviews. The published example uses the same bounded input.
Made the regional limit explicit: the tested US desktop and mobile pages require sign-in, so US review access has not recovered. This remains product-page sampling, not full history or a guaranteed count.
Runtime source, API defaults, filters, row fields, pricing and billing logic are unchanged. Only prefills, help text and documentation changed; the maintenance notice is left to platform checks.
2026-11-06 — Billing failure accounting
A billing-service exception now stops further billing and marks the run failed, without delivering the affected rows or claiming unconfirmed charges in SUMMARY. Previously the exception path reported the requested counts as charged. Prices are unchanged.
This accounting repair does not resolve the public Amazon source blocks observed in the current owner-account tests. Blocked pages still fail explicitly and are not billed.
1.5 — 2026-10-19
Fetching: a passing proxy error on a residential session now just rotates the IP; only a 407 (no residential access) drops the run back to the cheap pool. Before, one such error disabled the residential fallback for the rest of the run. Same inputs, rows and prices. Also up to 8 tries per product page (was 5).
2026-10-16 — Keyword Planner and locale documentation
Validated search terminology against Google Ads Keyword Planner and Store shelves. Added localized usage guidance and bounded examples; retained existing English phrases. Clarified locale limitations where measured. No runtime, output fields, defaults, or pricing changed.
1.4 — listing: blind-test fixes, 2026-10-12
No behaviour, output or price changed. Six Sonnet testers read the card and the input form for six real requests: 5 of 6 picked this actor and 6 of 6 wrote a valid first call. What they guessed at is now written down: minRating is called a star CEILING in its first words, the review row keys are listed in the event text, "8-13 reviews per product" replaces "up to 13", the residential proxy JSON is spelled out, the marketplace field says there is no language filter, and the form, README and event text all say there is no only-new mode (dedupe on review_id) and that a rating-only sweep is maxReviewsPerProduct: 1 (review rows cannot be switched off). Feature gaps the testers named, for a later release: an only-new mode and a product-row-only switch.
1.4 — listing and input form reworded, 2026-10-11
No behaviour, output or price changed; the same inputs return the same rows.
Title and description now say what one row is and what the ceiling is: one row per review with its fields, a product row with rating, ratings count and star histogram, the three free filters, and the honest limit — up to 13 reviews per product from the public page, no login.
The input form is reordered: includeProductRow (on by default, $0.003 per product) moved up next to the filters so the one default that changes the bill is in view; nothing renamed or removed.
Every input now says what to pass, with an example, its default and what it costs: minRating says it keeps reviews at or below the number (a complaints feed); maxReviewsPerProduct says why its maximum is 13; domain lists the 20 marketplaces with an example; proxyConfiguration says what the run does on a captcha or a sign-in-walled reviews card.
The pricing text gives the measured 8–13 reviews per product, lists what is never charged, that there is no start fee, and that a run stops at the maximum cost you set.
README: a section mapping common requests to the exact input, the output and empty-result contract, and the words the non-English marketplaces use for reviews (Bewertungen, Rezensionen, avis, opiniones, recensioni, レビュー, avaliações).
1.4 — 2026-09-12
Fix: runs came back with a product row and zero reviews, again. Amazon began serving some residential sessions (roughly one to three in ten) a complete product page whose reviews card reads "Sign in to see customer reviews." — title, price, rating and details all present, only the reviews withheld. The page passed the block check and the run reported no_reviews. That page is now recognised as a soft block for a reviews fetch: the session rotates to a fresh IP and retries, exactly as for a captcha, and the next IP almost always gets the full page. Product-detail fetches (the sibling actors) are unaffected, since the page is complete for them. Counted in the SUMMARY's http.blocked. No change to inputs, output rows or pricing.
1.3 — 2026-09-07
Fix: runs came back with a product row and zero reviews. Amazon started answering the browser identity the actor sent (a Chrome/139 user agent) with either a captcha or a product page whose below-the-fold section was empty — title and price present, reviews, details and description missing — so a run delivered the product row and reported no_reviews. The actor now identifies as a current Chrome (152), which gets the full page again on every marketplace tested, and the empty-below-the-fold page is recognised as a soft block: it rotates the IP and retries like a captcha instead of being billed as a product with no reviews. No change to inputs, output rows or pricing.
2026-08-30
Reliability: a run that cannot act on its input — a URL from the wrong site, a name nothing resolves to, an empty target list, a number outside the documented range — now ends cleanly with the reason on the run card instead of as a failed run. Nothing is charged in either case; a blocked or throttled source still fails loudly.
Store listing: moved into the categories buyers actually browse (the old Business category has no Store page).
1.2 — 2026-08-29
Automatic residential fallback. Amazon refuses whole datacentre ranges for minutes at a time, which occasionally ended a run with amazon page failed after 5 tries (last: 503) even though the same input had just worked. A run now starts on the cheap proxies you chose and escalates only when it is pushed back: after the same page is refused twice in a row the sessions re-open on a residential proxy and stay there for the rest of the run, so nothing repeats attempts that were never going to work. Nothing to configure, no change to pricing, and a run that is never refused never leaves the cheap pool. Reported as http.escalated in the run's SUMMARY.
1.1 — 2026-08-29
Amazon began answering 503 to the previous browser signature on 2026-08-29; the fetcher now presents a current Chrome build (verified on residential IPs).
1.0 — 2026-08-26
First release.
Customer reviews for any Amazon product in any of 20 marketplaces: star rating, review body, reviewer, the country the review came from, date, Verified Purchase flag, helpful votes and the variation bought.
A product row per ASIN with the overall rating, total rating count, brand, availability, image and feature bullets.
Input by ASIN or product URL; a URL's marketplace is used automatically, bare ASINs use the marketplace setting. An ASIN and its URL count as one product and are fetched once.
Three filters that run before billing: minRating (complaints only), verifiedOnly, and requireText (drop rating-only reviews). Filtered reviews are never charged.
Billing: pay per delivered review row and per product row. Status rows, filtered reviews, products with no visible reviews and failed products are all free.
Reliability: Amazon's captcha page is detected and the session rotates to a fresh IP instead of being parsed as an empty product.
Note on scope: this reads the reviews Amazon shows publicly on the product page — roughly 8–13 per product. Amazon requires a signed-in session for its full review archive, which this actor deliberately does not attempt. Price is returned when Amazon serves it, but it is frequently withheld from unauthenticated clients, so treat price as best-effort.