Amazon Seller Scraper

Collect public Amazon seller-profile evidence from a bounded list of seller IDs or exact seller-profile URLs. The Actor uses ordinary HTML requests and does not launch a browser, inject a fingerprint, solve challenges, call private APIs, or return fabricated fallback rows.

What it returns

Each dataset item is a successfully verified seller profile with a stable sellerIdentity and recordId, seller name and marketplace, the canonical public profile URL, available public rating and feedback evidence, optional policy and feedback text, and optionally complete product cards from one public storefront page. Missing optional values are omitted rather than written as null, empty strings, or "unknown".

Diagnostics are not billable/countable dataset items. Run status, access blocks, request counts, proxy mode, and bounded errors are written to the OUTPUT record.

Input

{
  "mode": "sellerId",
  "sellerIds": ["A14CZOWI0VEHLG"],
  "country": "in",
  "maxItems": 1,
  "maxFeedbackEntries": 5,
  "scrapeProducts": false,
  "requestTimeoutMs": 30000,
  "proxyConfiguration": { "useApifyProxy": false }
}

• mode: sellerId or sellerUrl.
• sellerIds / sellerUrls: 1–25 inputs. URL mode accepts only credential- and port-free HTTPS /sp URLs on the marketplace selected by country.
• country: in, com, co.uk, de, co.jp, ca, com.au, or com.br.
• maxItems: 1–25 sellers.
• maxFeedbackEntries: 0–20 visible feedback entries per seller.
• scrapeProducts: when true, requests one storefront page per seller.
• maxProducts: 1–50 complete product cards per seller.
• requestDelayMs: 0–3000 milliseconds between sellers.
• requestTimeoutMs: 5000–45000 milliseconds per request.
• proxyConfiguration: optional Apify Proxy configuration; direct access is the default. One sticky proxy session is used for the run, with no rotation on redirects or denial.
• captureSourceEvidence: optional boolean, default false. Store exact complete response bodies as described below.

Access behavior

Amazon can return robot checks, rate limits, regional redirects, or incomplete public pages. The Actor detects those boundaries and fails closed. A zero-row run exits unsuccessfully and records a typed diagnostic in OUTPUT; it does not disguise a challenge as valid seller data. Proxy setup is used only when explicitly requested, and a setup failure is never silently downgraded to direct access.

Output guarantees

• Dataset rows are verified seller profiles only; diagnostics stay in OUTPUT.
• URLs and redirects are restricted to the selected exact marketplace host.
• Responses are capped at 5 MB and redirects at three hops.
• Stable identity is derived from marketplace country plus seller ID.
• fingerprintApplied is always false; this direct HTTP Actor has no browser fingerprint surface.
• Product arrays contain only unique cards with ASIN, title, product URL, and image URL.

Use low limits first. Public HTML and selectors can change, so validate critical downstream decisions against the source page.

Evidence and verification

A name must be visible in a seller-name field. Page titles are not used as seller names. The requested seller, final profile URL, any visible hidden sellerID, profile page-state sellerID, and any canonical profile link must agree. A narrowly supported Amazon India case allows its same-seller www.amazon.com canonical only when explicit spp-page-var-page-state proves both the requested seller and India marketplace A21TJRUUN4KGV; fetched/source profile URLs remain on www.amazon.in. Same-host redirects to another seller or storefront are rejected.

Star ratings require an explicit “out of 5” source descriptor. A lifetime positive-feedback percentage is kept as a percentage and never treated as a star rating. Lifetime rating counts come only from lifetime/total fields, not a 12-month period. Explicit numeric zero is retained when consistent with the source. A zero total-feedback placeholder is omitted when the page contains a complete visible feedback entry, including when feedback output is disabled; absent, boolean, malformed, and unsafe numeric values are omitted. Optional product and feedback inputs are normalized without null/false coercion.

usedProxy, proxyConfigured, and connectionMode describe only Actor-configured Apify Proxy routing. directTransport is false when the Actor explicitly uses a proxy, true only when no inherited/runtime proxying is indicated, and omitted when actual egress cannot be established. Managed environment gateway settings are never disabled or changed. Supported proxy keys are useApifyProxy, apifyProxyGroups, and apifyProxyCountry; unsupported or custom-proxy options fail input validation. Credentials and all query/fragment values are removed from diagnostic errorUrl values; URL and labeled credential content is redacted from error messages.

Only successfully persisted rows are counted. A partial run uses PARTIAL and exits nonzero. A dataset write failure is summarized in OUTPUT. Run npm test to discover every *.test.js under test, npm run check for the aggregate syntax checks and all runnable tests (npm run check:syntax for syntax alone), and npm run validate to validate local dataset/OUTPUT evidence. The validator honors CRAWLEE_STORAGE_DIR.

Local entrypoint tests execute the real SDK and src/main.js, assert Actor.isAtHome() === false and a MemoryStorage storage client before entrypoint import, and persist success and failure evidence. Their Amazon HTTP responses are explicit test fixtures, not current live marketplace observations. No live target success, cloud deployment, or access-control workaround is claimed by those checks.

Opt-in exact response replay

Set captureSourceEvidence: true to retain complete received HTTP response bytes in the run's default key-value store. OUTPUT.sourceEvidenceKeys lists each successfully stored SOURCE_EVIDENCE_001, SOURCE_EVIDENCE_002, etc. record in request order. Each JSON record has sourceUrl, requestUrl, httpStatus, contentType, responseBytes, bodySha256, exact bodyBase64, and the parser settings. Credentials, cookies and response headers are not stored. Raw public response bodies may contain personal information, so enable this only when source retention is wanted.

Capture happens before denial, challenge, and redirect handling. Full denied bodies are retained when received within the 5 MB limit; connection failures, timeouts, and oversized/incomplete bodies cannot provide complete source evidence. Evidence does not imply a successful extraction. An evidence-write failure stops all further source requests, produces a typed diagnostic, and prevents complete success. Previously verified rows can remain a partial result.

Download each listed JSON record from that run's default key-value store. Replay locally without network or credentials:

npm run replay -- /path/to/SOURCE_EVIDENCE_001.json

The replay verifies exact-byte SHA256 and length, then runs the same seller/storefront parser with captured limits. It returns parsed source fields or an explicit rejection code; a denial replay is not successful seller evidence. Compare the returned seller fields with the dataset row; for optional storefront captures compare products with the normalized dataset products (which can additionally include numeric ratingValue). npm run validate verifies listed local evidence records as well as dataset and OUTPUT schemas.

The bounded cloud-validation-input.json keeps the India seller A14CZOWI0VEHLG, one seller, three feedback entries, products off, direct Actor routing, and evidence on. Profile and optional storefront parsing have local fixture coverage. No browser fixture is blocked or required. The current implementation has no authenticated seller-account mode; public seller ID/URL modes and explicitly configured Apify Proxy remain supported. Live cloud success must be established separately.

Captured India profile regression (2026-10-02)

Offline inspection of an actual HTTP 200 India response identified recoverable parsing defects: a same-seller cross-market canonical, truncated anchor attributes, and the legitimate /storefront?me=... route. The parser now handles those source shapes without requesting another marketplace. Exact bare-host public profile links such as amazon.in are normalized to the selected www.amazon.in host; unrelated hosts remain invalid. Explicit star rating, positive-feedback percentage, and lifetime versus shorter-period counts remain separate. This offline replay does not itself establish a successful redeployed cloud run.