index.

Research · Coming soon

Agent Commerce Evals.

Finding a plausible product is easy. Finding the right variant, in stock, at the actual delivered price is the test.

We’re designing tests of how well agents find products across merchant platforms. Below is the proposed method. Results and agent comparisons are coming soon.

What we plan to measure

01

Exact match

Does the product meet the request: model, size, color, budget and delivery destination? Check the exact variant, not just a similar product.

02

Price & delivered cost

Does the quoted price match the merchant, including shipping, tax and currency? Mark any total that cannot be verified as unknown.

03

Stock & freshness

Can that exact variant be bought when checked? Record the source and check time, then verify again at handoff.

04

Time & cost to answer

How long does a correct, verified answer take? Report response times, tool calls and token costs alongside accuracy.

05

Useful handoff

Can the agent take the right item to a supported purchase step? Report a successful handoff separately from a completed purchase.

Proposed method

Same task.
Same conditions.

Give each agent the same requests, delivery destination, currency and budget. Test across merchant platforms, including unclear requests and out-of-stock variants.

Compare browsing with catalog tools using the same model where possible. Record model versions, prompts, available tools and time limits. Repeat each task to distinguish consistent performance from a lucky answer.

Check answers against accessible merchant sources at the time of each run. Publish the evidence for each task, including failures and unknowns. Clearly label runs that use Index.

Tested agents, catalog coverage and a release date will be shared with the results. No scores or rankings have been published yet.

Building a commerce agent?

We’d like to hear which real shopping tasks it struggles with.

Why we’re building Index