Chris Green splits an AI SEO audit across three layers

A technical audit now looks like three layers rather than one chat window. Writing in Search Engine Journal, Chris Green, Technical Director at Torque Partnership, describes the Chrome extension he is building to assist SEO work, and the split he describes is useful for any audit where facts can be checked and the output is dense.

Green names three roles and keeps them separate. Deterministic checks verify objective facts. A local language model, running on the user's own machine, turns the output into plain prose. The human makes every judgment call.

The three layers, with what each is for

The deterministic layer is the one most SEO software already does: HTTP status codes, canonical tags, robots.txt rules, differences between the rendered DOM and the server-side DOM. Green treats this as the only source of facts. If the extension says a canonical points at the wrong page, it is because the program read the canonical, not because a model summarised a page.

The local model layer is Gemini Nano, running on-device inside Chrome. Green uses it to turn technical output into plain English: a short explanation of what the finding is, or a description a developer could paste into a Jira ticket. He tested the model on tasks like comparing the rendered DOM to the server-side DOM and reports that it is useful for summarising but not reliable for complex technical judgments. That is why the model never gets to decide anything.

The human layer makes the judgment call. In Green's description, the extension does not autonomously audit a site. It gives a person cleaner data and explanations faster, and the person decides what to do with them.

What the piece does and does not claim

There are no benchmark numbers in the article. Green does not quote a time saving, a conversion rate, or a dollar figure, and the piece is written as a method rather than a result. The practical point is a split: facts from code, prose from a model, decisions from a person.

The reason to lean on an on-device model, in Green's framing, is that it stays on the user's machine. A site owner is not sending their own site data to a provider when the audit runs, and the extension can work without a per-user API key or a usage cap. On-device inference also means a quick turnaround on each check, which is the part that keeps the human engaged.

The shape of this is more general than SEO. Any audit workflow with dense output and expensive guesses, security review, infrastructure checks, code review, benefits from the same decomposition: let a program establish what is true, let a model turn the output into prose, and keep the person as the one who decides.

Source

Primary: Using AI To Assist With SEO Work, Not Replace The Worker, Chris Green, Search Engine Journal.