Amazon Product Data for AI Agents: Listings, Offers and Sellers as JSON

Amazon product data for AI agents means giving an assistant a way to ask "what does this listing cost right now, and who is selling it?" and get structured JSON back instead of a web page. You hand it an ASIN or a seller ID, it returns price, rating, stock, features and sellers, over REST or as MCP tools. This guide shows one real request and response, and prices every call in points. For exact fields, request shape and prices, the Amazon product source page is the reference.

The short version

• —What you get. Five Amazon sources: product by ASIN, every offer on a listing, a seller profile, a seller's storefront list, and keyword search. Each is one request that returns one JSON object.
• —Who it is for. Teams that sell on Amazon, resell what others sell there, or build agents and tools around the catalog, and need a repeatable answer instead of opening pages by hand.
• —What it costs. 5 points per successful product, offers or seller lookup ($5 per 1,000). Search and storefront lists cost 1 point per returned record. A lookup that finds nothing is not billed.
• —How an agent reaches it. The same sources are MCP tools (run_amazon_product and its siblings) and REST functions, at the same price.
• —What to expect. Each response is a snapshot read at request time, and some fields come back null because Amazon does not publish them for every listing.

What the five Amazon sources return

"Amazon data" is really a handful of different questions that share a website. Each maps to one source with one price per request.

• —What is this product? Title, price, list price, images, rating, review count, features, details table, category and the buy-box seller. That is the Amazon product source, keyed by ASIN.
• —Who is selling it, and at what price? Every offer on the listing, with condition, shipping source, Prime flag, delivery estimate and seller. That is the Amazon product offers source, also keyed by ASIN.
• —Who is this seller? Storefront name, rating windows, business name and address when published. Keyed by seller ID.
• —What else does this seller sell? The storefront's product list, keyed by seller ID.
• —Which ASINs match a keyword? Product search, so an agent can start from a search term instead of identifiers you already hold.

All five take an optional domain parameter. It defaults to amazon.com and accepts any of the 19 marketplaces the source pages list. The same ASIN can come back with a different price, stock status and even a different buy-box seller depending on the country's site, so the domain is part of the question.

How a lookup works

1. —You choose the identifier and the marketplace. Products use an ASIN, sellers a seller ID, searches a query keyword, and any request can carry domain.
2. —The source resolves the record at request time. Price and stock are only as old as the call.
3. —The response is one JSON object. Every Amazon source returns the same envelope: a session_id to trace the run, a kind naming the source, the identifier and domain you asked for, a short human summary in Markdown, a machine object with the structured fields, and total_points.
4. —You are billed only if the record resolves. For ASIN and seller-ID lookups, an identifier that cannot be found is not billed. For search and storefront lists, billing is per record actually returned.

One real request and its response

The request follows the shape on the source page: a POST to the run-amazon-product function with an asin and an optional domain. The source page's own example uses a different ASIN, so this one uses the ASIN of the captured sample, B004YAVF8I, to keep request and response matching. Endpoint and parameter reference lives in the documentation.

curl -X POST https://api.neuralverge.ai/functions/v1/run-amazon-product \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "asin": "B004YAVF8I", "domain": "amazon.com" }'

This is the captured response from the source page. One long block is shortened: the details table in the real response has more entries than shown, and the line marked "..." stands in for the ones dropped. Nothing else was changed.

{
  "session_id": "11111111-1111-1111-1111-111111111111",
  "kind": "amazon_product",
  "asin": "B004YAVF8I",
  "domain": "amazon.com",
  "human": "# Amazon product\n\n- **ASIN:** B004YAVF8I\n- **Domain:** amazon.com\n- **Proxy traffic:** 674.4 KB\n- **Title:** Logitech M185 Compact Ambidextrous 2.4 GHz Wireless Mouse - Swift Grey\n- **Price:** $14.9\n- **Rating:** 4.4 (45164 reviews)\n- **Availability:** In Stock",
  "machine": {
    "product": {
      "id": "B004YAVF8I",
      "title": "Logitech M185 Compact Ambidextrous 2.4 GHz Wireless Mouse - Swift Grey",
      "brand": null,
      "byline": null,
      "price": 14.9,
      "currency": "$",
      "list_price": 17.99,
      "image": "https://m.media-amazon.com/images/I/51WN5aXZWIL.jpg",
      "images": [
        "https://m.media-amazon.com/images/I/31U-r5U5hoL.jpg",
        "https://m.media-amazon.com/images/I/4131CzsV8vL.jpg"
      ],
      "rating": 4.4,
      "review_count": 45164,
      "availability": "In Stock",
      "is_in_stock": true,
      "features": [
        "Compact Mouse: With a comfortable and contoured shape, this Logitech ambidextrous wireless mouse feels great in either right or left hand and is far superior to a touchpad",
        "Durable and Reliable: This USB wireless mouse features a line-by-line scroll wheel, up to 1 year of battery life (2) thanks to a smart sleep mode function, and comes with the included AA battery"
      ],
      "details": {
        "Connectivity Technology": "USB",
        "Additional Features": "Wireless",
        "Movement Detection": "Optical",
        "Button Quantity": "1",
        "Hand Orientation": "Ambidextrous",
        "Color": "Swift Grey",
        "Brand": "Logitech",
        "Model Name": "910-002225",
        "Manufacturer": "Logitech, Inc",
        "UPC": "097855112705 097855074126",
        "Warranty Description": "3 year limited",
        "Best Sellers Rank": "#1 in Computer Mice",
        "ASIN": "B004YAVF8I",
        "Customer Reviews": "4.4 out of 5 stars (45,164 ratings)",
        "Item Weight": "30 g",
        "...": "remaining detail entries omitted"
      },
      "breadcrumbs": [
        "Electronics",
        "Computers & Accessories"
      ],
      "category": "Mice",
      "seller": {
        "id": null,
        "name": "Amazon Resale"
      },
      "ships_from": "Amazon.com",
      "sold_by": "Amazon.com",
      "variations": [
        {
          "asin": "B004YAVF8I",
          "name": null
        },
        {
          "asin": "B0G4S2NJ31",
          "name": null
        }
      ]
    },
    "proxy_traffic_bytes": 690578
  },
  "total_points": 5
}

Reading the response

A few things in that JSON shape the code, or the agent prompt, that consumes it.

• —Prices are numbers, currency is a separate field. price is 14.9 and currency is "$". The human summary prints $14.9, so read machine, not the human string.
• —list_price lets you compute a discount. The list price is 17.99 against a current 14.9. You compute the difference; the source does not hand you a percentage.
• —Nulls are real. brand and byline are null here, and the brand appears only in the details table ("Brand": "Logitech"). Fall back to details rather than assume a top-level field is filled.
• —The buy-box seller has a name but no id here. seller is { "id": null, "name": "Amazon Resale" }, while sold_by reads Amazon.com. Do not use seller.id as a required join key.
• —variations lists sibling ASINs. B0G4S2NJ31 is another ASIN in the same product family. It is a lead for another lookup, not a finished record.
• —total_points is the bill. The call cost 5 points; cost is stated in points, not in the proxy_traffic_bytes figure.

Offers: every seller on a listing

The offers source returns one object per offer. The captured sample, for ASIN B087Z733CM, came back with a single offer, copied from the source page:

{
  "asin": "B087Z733CM",
  "price": 39.99,
  "currency": "$",
  "savings_percent": 39,
  "condition": null,
  "condition_note": null,
  "ships_from": "Amazon.com",
  "sold_by": "Amazon.com",
  "is_prime": true,
  "delivery": "FREE delivery Tuesday, September 29",
  "seller": { "id": null, "name": "Amazon.com", "rating": null, "rating_count": null, "positive_percent": null }
}

The envelope also carried "total_items": 1 and "total_points": 5. On a listing with several sellers the offers array holds more entries, each with its own condition and seller block. This sample is a single first-party offer, so it shows the field names but not the used-market spread you will see on busier listings. Seller rating fields are often empty, as here, because the offers view does not always publish them.

Seller and list sources

The seller source takes a seller ID. The captured sample is Amazon Resale, A2L77EE7U53NWQ, which returned a name, a business name (Amazon.com Services, Inc), a business address (410 Terry Avenue North) and a privacy-policy link, with null for every rating window and for phone. Business details appear only when the seller chose to publish them. A seller with a long history gives you rating_30d, rating_90d, rating_365d and rating_lifetime to compare.

The seller products and product search sources list results at 1 point per returned record. max_items caps the result size and start_page pages deeper. The source pages give a default of 10 for storefronts, 25 for search and a cap of 250 per request, so max_items is also your spending ceiling for the call.

Amazon product data for AI agents over REST and MCP

The same sources are reachable both ways, with the same arguments and the same point cost, because both routes run the same source. Over REST you POST to the function, as in the curl call above. Over MCP an agent calls the tool by name, with underscores where REST paths use hyphens:

{ "tool": "run_amazon_product", "arguments": { "asin": "B004YAVF8I", "domain": "amazon.com" } }

The four tools for this family are run_amazon_product, run_amazon_product_offers, run_amazon_seller and run_amazon_seller_products; product search follows the same naming. For connecting the hosted MCP server to Claude, see our MCP setup guide.

A few habits help when an agent is the caller:

• —Let it chain the lookups. A search finds ASINs, a product call fills in details, an offers call finds the winning seller, a seller call vets that seller.
• —Tell it nulls are normal. Instruct the agent to report a missing field as missing, not to infer it.
• —Keep the session_id. Store it with whatever the agent concludes so a surprising value can be traced back to the run that produced it.

Cost and limits

Every price below comes from the source pages, and the point value appears in each response as total_points.

The cheap, wide step is search: 25 results cost 25 points. The deeper step is the product lookup at 5 points per ASIN. Searching a keyword and pulling full details for ten shortlisted ASINs costs the 25 search points plus 50.

A worked example

This one is illustrative. Say Acme Oy, a Finnish distributor, wants to know whether a competing reseller is undercutting its price on one product. An agent runs the offers lookup on the ASIN (5 points), reads the winning offer and its seller, then runs the seller lookup on that seller ID (5 points) for business details and rating windows. That is 10 points for the comparison. Run daily, it is a fixed, predictable number rather than a project. For the scheduling side of repeated checks, see the guide to price monitoring over an API.

Limits to plan around

• —Snapshots, not streams. Price, stock and the buy-box seller can change within minutes. If you track a price, store the response with your own timestamp.
• —Not every field is filled. Seller id, brand, ratings, phone and condition notes are null in the captured samples.
• —Regional restrictions. A product with regional restrictions may not resolve on every domain.
• —Suspended or delisted identifiers do not resolve. That is a clean miss, and a miss is not billed.
• —Public listing information only. The sources return what the page publishes, stamped with the identifier it was read from and a traceable session id.

How to decide

Questions worth answering before you give an agent or a pipeline an Amazon data source, ours or someone else's:

1. —Which marketplaces do you need? Confirm they are among the supported 19 and test an ASIN on each.
2. —How fresh must the data be? If minutes matter, plan frequent re-runs and cost them in points.
3. —Which fields must be present? Run a sample of your real ASINs and count the nulls in the fields you depend on.
4. —Do you need history? A lookup gives you now. A time series is yours to store.
5. —What does a miss do to your workflow? A missing ASIN returns no product and no charge; treat that as a result, not an error.
6. —Which job is it? The best fits are price and stock monitoring, buy-box and reseller tracking, minimum advertised price checks, catalog enrichment, cross-market comparison and seller vetting before you source from a storefront.
7. —Who calls it? A developer scheduling jobs can use REST; an assistant answering questions inside a conversation can use MCP. Both are available.

Frequently asked questions

What does an Amazon lookup by ASIN cost?

The Amazon product source costs 5 points per successful lookup, which the source page lists as $5 per 1,000. The offers source and the seller profile source are also 5 points each. Search and seller-storefront listings cost 1 point per returned record. A lookup that finds nothing is not billed.

Can I query marketplaces other than amazon.com?

Yes. Each source takes an optional domain parameter that defaults to amazon.com and accepts any of the 19 supported Amazon marketplaces. The same ASIN can come back with a different price, availability, offer set and seller depending on the domain.

Is the data live or cached?

The source pages describe each lookup as read at request time. Price, stock and the winning seller can change within minutes, so treat every response as a snapshot with a timestamp you add yourself, and re-run the lookup when freshness matters.

Can an AI agent call these Amazon sources through MCP?

Yes. These sources are available over REST and as MCP tools named run_amazon_product, run_amazon_product_offers, run_amazon_seller and run_amazon_seller_products.

Why are some fields null in the response?

Amazon does not publish every field for every listing. In the captured sample, brand, byline and the seller id are null, and the Amazon Resale seller profile returns null for every rating window. The sources return what the page publishes and do not fill gaps with guesses.

How is this different from Amazon's own official APIs?

This article does not compare feature lists, because we cannot verify current terms for Amazon's programs. Check Amazon's own documentation for eligibility and access rules. What these sources offer is pay-per-lookup access with no seller account required, one key shared with every other NeuralVerge source.