Autopilot Operating Contract

Autonomous website changes raise hallucination, brand-damage, permission, and rollback concerns. Output quality alone doesn't clear the bar. This contract states exactly what Autopilot can and can't do, how we prevent fabrication, and how you stay in control.

1. The Promise

Autopilot is built to close AEO gaps, not generate content. Here's what it does:

Detects gaps — Where ChatGPT, Perplexity, and Gemini don't cite you for your money intents
Generates grounded fixes — Based on YOUR existing content, never fabricated claims
Proposes changes for approval — Every intervention requires your explicit approval before publishing
Re-probes weekly — Measures whether citations actually moved, iterates on what works

This is not a content generator. This is a gap-closing agent that operates under strict boundaries.

2. Grounded-Source Requirements

How we prevent hallucination:

Tier 1: Deterministic (zero LLM involvement)

robots.txt updates — Allow AI crawler access (GPTBot, PerplexityBot, etc.)
llms.txt creation — Structured crawl guidance for AI assistants
Schema type selection — JSON-LD blocks for Organization, WebSite, FAQ, Product

These interventions are rule-based. No language model can fabricate a robots.txt directive.

Tier 2: Grounded (LLM-generated from your content, fact-checked)

Answer capsules — Direct-answer paragraphs generated from your actual page content
FAQ schema — Questions/answers extracted from your existing FAQ or documentation
Meta descriptions — Optimized for AI extraction, based on your page's actual content

Every claim in a Tier 2 artifact must be verifiable in your existing content. Before any Tier 2 intervention ships to design partners, it passes a 17-domain validation test (documented in /lab).

Tier 3: Prohibited (we never do this)

Inventing product features you don't offer
Fabricating pricing, plans, or availability
Making claims you don't make on your site
Writing blog posts, landing pages, or marketing copy

If an intent asks for something you don't offer, we skip it. Autopilot optimizes for truth, not coverage.

Evidence requirement: Every Tier 2 intervention includes a "source evidence" section showing exactly which page and excerpt the generated text came from. You can verify the claim before approving.

3. Approval Levels

What requires your approval:

Every intervention — Autopilot proposes; you approve before anything publishes
Query set changes — If we want to optimize for different intents, you approve the new set
Major site changes — Anything beyond schema/capsules/robots requires explicit sign-off

What doesn't require approval:

Re-probes — Weekly citation checks are read-only, no permission needed
Evidence collection — Fetching your pages to analyze them (read-only crawl)
Internal analysis — Gap detection, scoring, prioritization

Default is: we ask first. If you didn't explicitly approve it, we didn't publish it.

4. Change Boundaries

✓ What Autopilot can change (with approval):

Add JSON-LD schema blocks (FAQ, Product, Organization, WebSite)
Add/update answer capsules on existing pages
Update robots.txt to allow AI crawlers
Create/update llms.txt
Add meta descriptions optimized for AI extraction

✗ What Autopilot NEVER touches:

Product pages, pricing, features, claims (no fabrication)
Existing content (only additions, never deletions)
Third-party integrations, analytics, ads
Authentication, payments, user data
Navigation, site structure, URLs
Brand voice, tone, messaging (uses YOUR content as source)

Autopilot is metadata-only. It can't break your site because it never touches the parts that make your site work.

5. Audit Logs

Every action is logged. Design partners get read access to their complete audit trail:

Example audit log entry

{
  "intervention_id": "int_2026_07_20_001",
  "proposed_at": "2026-07-20T14:23:00Z",
  "approved_at": "2026-07-20T15:45:00Z",
  "approved_by": "[email protected]",
  "type": "schema_faq",
  "target_url": "https://yoursite.com/pricing",
  "change_summary": "Add FAQ schema for 3 pricing questions",
  "source_evidence": "Extracted from /pricing FAQ section",
  "commit": "a3f2c8d",
  "next_probe": "2026-07-27",
  "probe_result": {
    "citations_gained": 2,
    "citations_lost": 0,
    "verdict": "win"
  }
}

What was proposed (full diff)
What you approved (timestamp, user)
What was published (committed artifact)
What the next probe observed (citations gained/lost)

Access: Design partners get read access to their audit log. Exportable as JSON (your data, your records). No silent edits, ever.

6. Rollback

Every intervention is discrete and reversible:

Every intervention includes exact rollback instructions

Remove this schema block, revert this capsule, restore old robots.txt — spelled out in the approval request

Your CMS + version control = ultimate authority

Autopilot never hides what it did. You can undo any change through your normal publishing workflow.

Automatic rollback triggers (opt-in)

If re-probe shows citation DROP after intervention → flag for review, optionally auto-revert

If an intervention doesn't work, you can undo it. If it actively hurts, we can auto-revert based on probe data (you configure the trigger threshold).

7. Failure Handling

What if the agent is wrong?

Tier 1 interventions — Deterministic, can't fabricate (robots.txt is robots.txt)
Tier 2 interventions — Fact-checked before proposal; if you spot a fabrication, reject it and we fix the verification pipeline
Tier 3 interventions — Don't exist. We ship deterministic first, then add LLM-generated content ONLY after proving grounding works

What if a change breaks something?

Autopilot changes are additions/metadata, never core functionality
Schema errors → search engines ignore, your site still works
Bad robots.txt → crawlers fall back to default allow
Bad capsule → shows on your page but doesn't break layout

Liability

You approve every change → you own the outcome
We provide rollback instructions → you can undo
This is a design partner experiment → we iterate together on what's safe

If something goes wrong, we want to know. Design partners have a direct line to report issues — no support queue, straight to the builder.

8. Prohibited Actions (The Never List)

Autopilot will NEVER:

⛔

Publish without your approval

⛔

Invent product features, pricing, or claims

⛔

Delete or modify existing content (additions only)

⛔

Touch payment/auth/user-data systems

⛔

Change site functionality (only metadata/schema/capsules)

⛔

Optimize for misleading queries (if an intent asks for something you don't offer, we skip it)

9. Design Partner Agreement

What you're agreeing to:

Weekly re-probes of your site (read-only)
Proposed interventions (you approve before publish)
Shared learnings (anonymized case study if it works)
Honest failure reporting (if it doesn't work, we say so publicly on /lab)

What you get:

Free $79 deep audits (weekly, as long as the experiment runs)
First access to Autopilot (before paid launch)
Direct input on what gets built
Your audit log + rollback instructions
A direct line to the builder (no support queue)

Eligibility

Live website with real traffic (not a staging site)
CMS access (you can publish the changes we propose)
Willing to share anonymized results (citations before/after, not your brand unless you want to)
Honest feedback (if something doesn't work, tell us — that's the point)

This is a small cohort, chosen for fit. We're looking for sites where AEO gaps are a real constraint, not where they're nice-to-have.

Apply for design partner program

Read the public ledger to see proof we follow our own rules (we run Autopilot on our own site, in public).