apify-product-lookup

Apify product lookup

A model answers product questions from training data that was fixed months ago, or by

reading a page as prose with no product fields. This skill routes the question to a hosted

Actor that reads retailer pages on request and returns structured records: price, currency,

stock, rating, image URLs, and a link.

Written by a product marketing manager at Apify. It calls

E-commerce Scraping Tool, a paid

first-party Apify Actor, so weigh the framing accordingly. Everything below is what

testing against the live Actor returned, including where it does not work.

Setup

The MCP endpoint is https://mcp.apify.com. Appending

?tools=apify/e-commerce-scraping-tool narrows the server to this one Actor, which makes

tool selection more reliable when product data is the only job. Authentication is OAuth on

first use for interactive clients, or a bearer token for unattended ones. An Apify account

is required and the Actor is paid, billed per run started plus per product returned.

Pick the input before calling

| The question | Input to send |

|---|---|

| About a specific page the user gave you | detailsUrls: [{"url": "..."}] |

| "Find me an X under $Y" with no URL | keyword plus marketplaces |

| "Compare X across stores" | keyword plus several marketplaces |

Always send maxProductResults and additionalProperties: true. Without the second,

stock, rating, list price and identifiers are all missing, because they are nested there

rather than at the top level.

The call is two steps, sometimes three

1. Call apify--e-commerce-scraping-tool (two hyphens, not a slash). It returns run

metadata and a datasetId. It does not return products.

1. Check status. If it is not SUCCEEDED, call get-actor-run with the runId and

a waitSecs until it is. The Actor tool returns when its own wait window elapses, not

when the run finishes, so RUNNING is a normal answer and the dataset holds nothing at

that moment.

1. Call get-dataset-items with the datasetId, a limit, and fields.

Two failures live here. Stopping after step 1 returns a result that reads like success and

holds no product data. Skipping step 2 fetches an empty dataset and reports the product as

not found, intermittently, depending only on how fast the retailer answered.

Project with fields, and read them defensively

The full record is large: one Amazon product measured about 88 KB across 142 fields, most

of it marketing content and review text. Pass fields in dot notation, naming only what

the question needs.

Projected output comes back flat. The dots stay in the key names, so it is

item["offers.price"], not item["offers"]["price"]. Walking the nested path against a

projected response finds nothing, which is indistinguishable from the retailer not

reporting the field.

The four differences that bite hardest:

• offers.price is a number on some retailers and a string on others.
• Currency is offers.priceCurrency on some and offers.currency on others, and the value

may be a symbol or an ISO code.

• Stock and rating live under additionalProperties, as inStock and stars, not at the

top level. Top-level rating was null on a product whose stars was 4.2.

• brand frequently carries marketing text such as Visit the Sony Store rather than a

brand name.

Answer honestly

• Say when the data was read. The point of the call is that the answer is current, so

make that visible.

• Never claim a product is unavailable because stock was absent. Many retailers do not

report it. Absent means unknown.

• If every field comes back empty, suspect the URL before the retailer. An

unresolvable URL returns an item with no fields rather than an error.

• Quote the source URL so the user can check it.

Scope and limits

• Latency is high and unstable. Measured end to end over MCP across twenty questions:

14.7s fastest, 24.1s median, 95.8s slowest. Treat a runtime call as verification before

the user acts, not as something to put inside every turn.

• The marketplaces enum arrives truncated over MCP, because the server caps how many

characters of a long enum it forwards. Measured on 2026-08-27: of 249 values, 122 fit and

127 were dropped. A validation error naming marketplaces means the retailer cannot be

named in a keyword search on this connection, not that it is unsupported. Use

detailsUrls, which takes arbitrary URLs and is unaffected.

• Field coverage is a property of the retailer, not the tool. On one measured run,

Amazon returned 44 keys under additionalProperties and two other retailers returned 4,

with stock and rating absent on every record from one of them. There is no unified schema

to rely on.

• Cost. Cap with maxProductResults. Prefer one call with several URLs or marketplaces

over several calls, and never re-run to check a result you already have.

• additionalProperties can reach roughly 100 KB for one product. Never paste it into

a reply.

• A rating or stars of 0 means absent, not a zero-star product.