Alt Benchmark

What can today’s AI agents actually do on real websites?

A quarterly, open-methodology study: several agent vendors attempt everyday discovery and qualification tasks (find a product, read a policy, get an estimate) on the public pages of top sites. Replayable evidence for every result.

Logged-out, read-onlyrobots.txt respectedSigned identity
Edition #1 · planned task setPublishes mid-Nov 2026
TaskTypeWhy it matters
Find a product / room meeting 3 constraintsDiscover≈76% of agent activity is discovery (HUMAN)
Answer: is cancellation / return free?QualifyPolicies agents cannot resolve stall the sale
Get a price or quote estimateQualifyNo personal data entered
Find accessibility informationQualifyLitigation + agent completion both depend on it

Schedule

#1 Holiday 2026: read-only discovery tasks

Retail and travel. Details

Travel booking funnels

Planned

Insurance quotes

Planned

Government services

Planned

Get it first

Be notified when edition #1 publishes

Journalists: press kit and accuracy table.