Alt Matrix

One grid. Every agent. Every kind of customer.

Rows are people: first-time mobile shoppers, older adults, keyboard users, non-native readers. Columns are AI agents from different vendors. Each cell says whether they finished your funnel, and if not, the exact step, with the recording.

checkout · northwind-travel.exampleSample data
Sample Alt Matrix: personas by agents, pass or fail with blocking step
Persona ↓ / Agent →Claude CUGemini CUBrowser UseOpenAI CUA
First-timer, mobilepatience 5 · iPhone
Impatient shopperpatience 2
Older adult, low dexteritymin target 44px
Non-native reader (B1)CEFR B1
Keyboard-onlyno pointer
Column fails → agent-class problemRow fails → persona problemSame step in ≥3 cells → product bug
Click any cell to see the blocking step and evidence.

Three patterns tell you what to fix first.

Same step, many cells

When one step fails in three or more cells across rows and columns, it is a product bug: the highest priority.

A failing column

One agent fails everywhere: usually bot management, a CAPTCHA or a control that agent cannot operate.

A failing row

One kind of person fails with every agent: small targets, complex copy, keyboard traps.

Evidence, not a number

  • Pass / fail against machine-checked success criteria
  • The blocking step with screenshot, DOM snapshot and accessible names
  • Full replay, console and network logs
  • Persona parameters and the agent’s model version and date
  • “Behavioural approximation” label on consumer-agent columns
The score stays private

No public 0-100 badge

Each job has a private Alt Matrix score (persona pass rate, agent pass rate, Confirmed findings) for tracking release over release. We never publish a score for your site: readiness scores are a commodity, and the evidence is what you act on.

What an Alt Matrix does not prove.

Personas are simulations, not people. Findings are evidence of behaviour we observed on a given date with a given model version. They are never a statement that a site is “compliant”, “certified” or “agent-ready”. Consumer-agent classes are behavioural approximations, not the named products. Every published precision figure states its sample size and confidence interval. Read the full limitations statement →

Watch a job fill, cell by cell.

Signed agents run in real browsers, two sessions per origin at most, with a kill switch one click away. Product preview on sample data.

A running job: twelve personas by five agents, cells resolving to pass or fail, with Pause and Kill job controls.

See where agents fail on your funnel.

Run one free cell on a property you own. No sales call, no card.