Responsible SEO Automation

Responsible SEO automation makes investigation faster without hiding decisions or expanding permissions. The default should be read-only analysis with reproducible evidence. State changes should be narrow, explicit, and confirmed.

Operating principles

Discover identifiers

Call list_projects, and call list_crawls when the task needs a crawl. For rankings or AI visibility, discover the relevant keyword, segment, prompt, and stored analysis instead. Never invent or reuse unverified identifiers from an older conversation.

Retrieve only what the task needs

Use structured filters and bounded pagination. A smaller, relevant dataset is easier to verify than a large export.

Separate fact from recommendation

Report the observed field and value before explaining impact. Label hypotheses as hypotheses.

Preserve human authority

Do not infer authorization to create, update, delete, start a crawl, add tracked keywords, or add monitored prompts from a request to analyze, diagnose, or recommend.

Require exact confirmation

If a tool accepts confirm=true, the confirmation must refer to that exact action and target.

Read-only automation template

This is a read-only recurring review.
Use the latest completed crawl. Do not start a new crawl.
Retrieve only the records needed for the requested checks.
If data is missing, report the gap instead of changing account state.

Change-planning template

Plan the requested change.
Show the project UID, current value, proposed value, expected effect, and risk.
Do not call a state-changing tool. Wait for confirmation.

Reporting provenance

Every durable report should include:

• Project UID.
• Crawl UID or comparison pair.
• Crawl timestamp.
• For ranking or AI visibility reports, the matching analysis date, device, location, model, keyword, or prompt instead of a crawl UID.
• Filters and operators.
• Result count and sampling limit.
• Known exclusions or missing data.
• Generation or review date.

Avoid these patterns

• Starting a crawl because the existing dataset looks old.
• Deleting a project after a broad cleanup request.
• Treating a quick-win label as sufficient evidence.
• Hiding pagination or sampling.
• Combining unrelated changes into one confirmation.
• Filling remaining plan capacity with keywords or prompts without a user request.
• Claiming a causal performance diagnosis from correlated metrics.

See Security and permissions for server-level guardrails.