Agent · Product Validation

Find out which product data is wrong — before your customers do.

Validation re-checks the data you already have against fresh sources. Every attribute comes back with a confidence score, the reasoning behind it, and the page it was read from. Every product comes back with an overall quality score.

Validation VLD-2291

Cordless impact driver, 18V brushless

71
Voltage18 Vmanufacturer datasheet98
Torque1,500 in-lbThree current sources state 1,825 in-lb. The stored figure matches a superseded model revision.dewalt.com44
Chuck1/4 in hex quick-releaseacmetools.com93
Weight2.8 lbSources disagree on whether the figure includes the battery.homedepot.com62

Overall quality 71 · 1 attribute contradicted

The situation

Enrichment has a finish line. Accuracy does not.

Filling a catalog is a project. It gets a budget, a deadline, and someone who owns it. Keeping it true is nobody’s project.

So specifications drift. A manufacturer revises a model and keeps the part number. A supplier updates a datasheet nobody reads. Someone pastes the wrong dimensions into row 4,812 and the mistake propagates into a feed, a marketplace listing, and a customer’s expectations.

From the outside the catalog looks finished. It is complete and quietly wrong, and the first person to notice is usually the one filing the return.

The score

What 0 to 100 means

The number is only useful if you know what to do with it. Three bands, three responses.

85 – 100
Multiple current sources agree with the value on record. Safe to publish without review.
65 – 84
Broadly supported, but with a caveat — one source disagrees, the wording is ambiguous, or the figure depends on a measurement condition. Worth a glance.
0 – 64
The value on record is contradicted, unsupported, or impossible to establish. These are the rows to work through.

Attribute scores roll up into an overall quality score for the product, which is what makes a catalog-wide reading possible — one number for a range, a brand, or a supplier, tracked across runs.

Why it is trustworthy

A score with its reasoning attached

A bare number invites an argument. A number with the contradicting source attached ends one.

The comparison

What you have on record, next to what current sources say. Stated plainly, in the same row.

The reasoning

Written out: how many sources agreed, what they disagreed about, and why the score landed where it did.

The citation

The page it read. Open it and settle the question yourself in ten seconds.

When to run it

Four moments that justify a sweep

After a supplier feed lands

New data arrives with the same confidence as old data and none of the scrutiny. Validate the delta before it reaches the storefront.

Before a marketplace submission

Hold back the low-confidence rows yourself rather than letting the channel reject them for you and cost you the listing.

After a migration or acquisition

You inherited a catalog. Validation tells you what you actually inherited, in a number you can put in a status report.

On a cadence for the top of the range

The SKUs that carry the margin deserve to be re-checked on a schedule. Trigger through the API and let the webhook tell you what moved.
Process

How a validation job runs

  1. STEP 1

    Point it at existing data

    Select products you already have. Validation needs no new input — the data on record is the thing under test.

  2. STEP 2

    It re-researches from scratch

    Each attribute is looked up again from current sources, independently of whatever is stored.

  3. STEP 3

    Compare, score, explain

    Stored value against found value, scored 0–100, with the reasoning written out and the source linked.

  4. STEP 4

    Work the exceptions

    Filter by score, re-enrich the failures, or export the exception list for whoever owns that part of the catalog.

At catalog scale

Built to run against everything.

Validation is only useful if it is cheap enough to run on the whole catalog and repeat it next quarter.

Bulk selection
Select from the products table, or target a whole catalog or category in one job.
Parallel processing
Products are checked concurrently with live progress. No per-job SKU cap.
Webhooks
A completion event fires when the job finishes, so a scheduled sweep can drive your own workflow.
Job history
Every run is stored with its inputs, model tier, cost, and results.
API
Create and read validation jobs programmatically, the same as any other agent.
Validation

What people ask once they see the scores

How is validation different from enrichment?
Enrichment fills empty fields. Validation tests full ones. Enrichment answers "what is this product?"; validation answers "is what we already wrote down still true?" Most teams run enrichment once on a range and validation repeatedly.
What does the score actually measure?
Agreement between the value on record and what current sources say, weighted by how well-supported those sources are. It is a measure of evidence, not of importance — a low score on a field nobody reads is still a low score.
Does it change our data?
No. Validation reports; it does not overwrite. Acting on a finding — re-enriching, editing, or ignoring it — is a separate decision you make.
Can we validate only some attributes?
Yes. Choose which attributes a job checks. Running against the six fields that affect feed acceptance is cheaper and more useful than running against all forty.
What does it cost to validate a whole catalog?
The same credit model as enrichment, and the estimate is shown before the job starts. Budget tier is often the right choice for a first sweep: it is looking for the obviously wrong rows, and it finds them cheaply.
Can we track quality over time?
Every run is stored with its scores, so the overall quality score becomes a trend rather than a one-off reading.

Point it at a hundred products you believe are fine.

This is the demo that tends to change the conversation. Pick a range you are confident in and see what the scores say.

5 credits on signup · no card required