Everything Compass Does
Task-based AI navigation audits that show you exactly what's blocking discovery — and what to fix.
💡 The demos below aren't sales mockups — they're real UI components lifted from Compass with sample data.
Core capabilities
Task-based auditing
Define what you want AI to accomplish on your site — "find pricing", "locate returns policy", "reach support" — and Compass tests whether it can.
- 140+ pre-built tasks across industries
- Universal, SaaS, E-commerce, Content, Financial, Healthcare, Enterprise categories
- Customise existing tasks for specific verticals, topics, or locales
- Create completely bespoke tasks for your unique needs
Select Tasks
Choose what you want the AI agent to find on your site
Hybrid navigation engine
Not a simple crawler. Compass uses ML to pre-filter link candidates and an LLM to make authentic navigation decisions — the same way real AI agents work.
- XGBoost pre-filtering for efficient link selection
- Claude-powered navigation decisions
- Two strategy modes: homepage navigation vs. search-first
- Real-time visibility into agent reasoning
Running Audit
acme-saas.com
Failure classification
When something fails, Compass tells you exactly why — not just "couldn't find it" but the specific failure mode so you know what to fix.
- Blocked: robots.txt, authentication, paywalls
- Navigation failures: buried links, confusing menus
- Depth failures: too many steps to reach content
- Content gaps: genuinely missing information
403 Forbidden - Bot detectedPage blocked by bot detection. Server returned 403 Forbidden when accessed without browser fingerprint.
Step-by-step navigation traces
See exactly what the agent saw, which links it considered, why it chose each path, and where it got stuck.
- Full navigation path visualisation
- ML scores and SHAP feature importance
- LLM reasoning for each decision
- Link-by-link candidate ranking
""Pricing" link is the most direct match for finding pricing information. Located in primary navigation with high text similarity score (0.42). Choosing over "Plans & Pricing" due to cleaner URL path."
"Task complete. Current page displays pricing tiers with monthly/annual pricing ($29/mo Starter, $79/mo Pro, $199/mo Enterprise). All information needed is present on this page."
"Navigation successful. Found pricing information in 2 steps (Homepage → Pricing)."
Actionable recommendations
Every audit produces specific, prioritised fixes with estimated impact scores — not generic advice but targeted improvements.
- Potential score calculation shows improvement ceiling
- Prioritised by impact on AI readiness
- Specific page and element references
- Implementation guidance for each fix
Real data foundation
Navigation traces used to train our models, across 250 sites and 494,197 links analysed.
Industry task packs
Pre-built tasks covering Universal, SaaS, E-commerce, Content, Financial, Healthcare, and Enterprise sites.
Project management
Organise audits by project. Track multiple domains, compare results over time, manage client work.
Audit history
Every audit saved with complete results. Compare changes, track improvements, demonstrate progress.
Coming soon
We're actively building. Here's what's on the roadmap.
Multi-tenant accounts
Team workspaces with role-based access and audit assignment.
Scheduled audits
Automated recurring audits to catch regressions before they impact visibility.
Why we test full navigation paths
Most AI tools can't actually browse — they land from search and that's it. Compass uses browser automation to test complete navigation paths, giving you an optimistic ceiling on your AI accessibility. If your site fails in Compass, it'll definitely fail in the real world.
How it works under the hood
Compass uses a hybrid ML+LLM architecture designed to simulate how real AI agents navigate websites.
The navigation pipeline
- Page analysis: When Compass lands on a page, it extracts all navigable elements — links, buttons, form fields, interactive components.
- ML pre-filtering: An XGBoost model trained on 494,197 links scores each candidate based on text, position, and context. This narrows thousands of links to the most promising candidates.
- LLM decision: Claude evaluates the shortlist against the task goal and selects the best navigation action, providing reasoning for each choice.
- Result classification: After each step, ML models classify the outcome — success, partial progress, or failure with specific category.
Why hybrid?
Pure LLM navigation is expensive and slow. Pure ML can't handle the semantic reasoning needed for complex navigation decisions. The hybrid approach gives us the best of both: efficient filtering with intelligent decision-making.
Training data
Our models are trained on 3,348 real navigation traces across 250 websites. Not synthetic benchmarks — actual agent behaviour from production systems.