AXO — Agent Experience Optimization

AXO measures something new: not how a human experiences your site, and not whether crawlers can read it, but how an autonomous AI agent fares when it tries to actually use your site — find what it offers, call its capabilities, complete a task. Enter a URL, run a scan, and watch the live scan execute a generated task against your page.

The score

You get an overall score built from weighted categories:

• Agent Discovery — the standard protocols agents use to find what you offer: llms.txt, robots directives, schema markup, API discoverability.
• Agent Capabilities — the tools your site exposes to agents, including WebMCP tools detected on the page.
• Agent Experience — the real UX a browser agent encounters: which steps of the generated task it could complete, where it got stuck.
• Commerce (optional section) — checkout-and-inventory readiness checks, reported separately and not counted in your score.

Blockers and recommendations

Below the categories:

• Blockers — the specific failures that stop agents outright.
• Recommendations — ranked fixes, including ready-to-paste WebMCP code where a missing tool definition is the fix, and a "Copy all instructions" button that bundles everything for your developers (or your coding agent).

Quota

AXO scans have a small monthly allowance per project (plan-dependent; the default tier includes a few scans per calendar month). The History panel keeps past scans with their overall scores so you can verify a fix actually moved the number.

Why this matters now

A growing share of "visits" are agents acting for a user — comparing products, checking availability, filling carts. A site that's merely crawlable is readable; a site with good AXO is operable. The gap shows up in checks like: can an agent find your API? Can it read structured prices? Does your bot protection challenge legitimate agents into failure?

Suggested loop

1. Scan your most transactional URL (not the homepage).
2. Fix Blockers first — one blocker usually costs more than ten warnings.
3. Ship the WebMCP recommendations if you expose any user-facing tools.
4. Re-scan after each change; History shows whether the score moved.