What is an autonomous SEO agent?

From a report to a completed task

An autonomous SEO agent is software that can inspect a website, investigate supported search problems, prepare changes and check the result without needing a separate instruction for every step. Its useful output is a traceable task: what was observed, what changed and what was verified.

Consider a homepage redesign that leaves the main collection button pointing to a deleted URL. A report can list a broken link. An agent can follow the destination, find the working collection, inspect the source, propose a small patch and run the project’s checks. After the patch is approved and deployed, it can revisit the served page.

That last step matters. A source edit can be correct while a deployment still serves the old version. The task should remain open until the website evidence catches up with the code.

What happens automatically, and what stays with you?

In Your Optimiser’s hosted workflow, a monitored repository update can trigger the investigation. The agent focuses on relevant files and pages, evaluates evidence, stages a permitted repair, runs configured checks and opens a pull request when justified. Your team reviews and merges it. Deployment checks then establish whether the repair reached the site.

The distinction is deliberate: autonomous investigation and preparation can coexist with human publication review. Local installations have their own action policy. Enabling a model does not grant it permission to rewrite every file, publish unsupported content or change intentional private pages.

• You choose the repository, monitored branch and permitted directories.
• You supply business facts that cannot safely be inferred from code.
• The application enforces publication permissions and verification requirements.
• Missing evidence stays a dependency rather than becoming a fabricated fix.

What we observed in a recorded repair

Our 52-second demo condenses a real local run on a fictional lighting shop. The fixture had four confirmed defects: a missing description, a broken collection link, conflicting canonical declarations and incorrect product offer data. The agent read two source files, proposed patches and ran the fixture’s checks in an isolated environment.

The worker rechecked the served pages and recorded four resolved defects with no regressions detected in that audit. The private preview and crawler preferences stayed unchanged. A repeat audit created no new commit. A subsequent docs-only merge produced no inference request because it affected no configured website route.

This is evidence of one repair workflow, not a customer case study or an estimate of ranking impact. The demo used local Git. It did not publish a live GitHub pull request, and it did not measure organic traffic or AI citations.

Source: Inspect the recorded repair evidence ↗

A healthy page is a successful outcome

An agent that edits something on every run is creating work, not necessarily value. A new background colour may leave the title, links and public content intact. A sensible result is to record the relevant checks and stop. A shared template change, on the other hand, can justify looking beyond the edited file because many pages inherit its output.

This is why change scope and periodic checks belong together. Events make the workflow timely. Reconciliation helps catch missed events and website changes outside the repository. Neither should turn into repeated model calls for an unchanged, already verified observation.

Evaluate the work, not the activity

Ask for the evidence behind each finding, the actual diff, the checks that ran and the post-deployment observation. A confident explanation is not a test result. A screenshot can reveal an overlapping menu, but it cannot establish whether a search engine indexed the page.

For a trial, include a known defect, a healthy update, an intentionally private page and a failure in the project’s own checks. The agent should repair the defect when policy permits, leave the healthy page alone, preserve privacy and refuse to label a failed check as verified.

Use verified fixes, unnecessary changes, regressions, elapsed time and operating cost to assess the workflow. Track search impressions, clicks, enquiries and AI-answer citations separately. Those outcomes need observations over time and cannot be guaranteed by an autonomous agent.