Chris Green argues for a three tier split, deterministic code, a browser local model and a frontier model, with Chrome's Gemini Nano doing the middle tier

Most of the AI work in an SEO pipeline should not be paying frontier-model prices for it, Chris Green argues in Search Engine Journal. Green, technical director and senior consultant at Torque Partnership, published the piece on 30 September and sets out a three-tier split: deterministic code for exact tasks, a local model such as Chrome's built-in Gemini Nano for light interpretation, and a frontier model reserved for genuine reasoning.

The piece opens with the pattern he says is common: a team defaults to a frontier model for every task, including the ones a small local model would answer just as well, and the ones a plain script would answer exactly. The cost, privacy and reliability case for splitting the work is the same case every time, Green writes: cost scales with calls, privacy is better when data stays local, and a local tier does not go down with a remote provider.

Tier one, code

Green's examples for the deterministic tier are URL extraction, HTML comparison and checking an HTTP response. These are tasks where a model's answer and a script's answer match, and the script's answer is faster, cheaper and does not vary. He names Exactly Matchy, a Chrome extension he builds that checks whether a page is part of a large-language-model retrieval pipeline, as the kind of tool that runs on the client rather than paying for a call per check.

Tier two, a browser-local model

The middle tier is where Chrome's Gemini Nano sits. Gemini Nano ships inside Chrome, runs on the device, and can be called from the browser without a network trip. Green says it is good enough for the extraction, classification and short-summary jobs that make up most of the AI calls in an SEO workflow. He compares it against Gemini Flash and ChatGPT Luna in the piece, and notes that it does not need to replace a frontier model to be useful: it just has to handle the common case so the frontier model only sees the uncommon one.

Tier three, a frontier model where it earns the price

Green's rule for the top tier is that it only runs when a task needs reasoning that the local model will not do well. The design question he poses to the reader: for each AI call your pipeline makes, is the answer exact (use code), is it light interpretation on a page the user is already on (use the local model), or does it need reasoning (pay for it)? He also argues the discipline encourages better system design, because asking the question forces a team to say what the AI call is for.

Green's figures and the Gemini Nano comparison are his own, from his own benchmarks. The piece is a design argument rather than a measurement study.

Source

• Primary source: Search Engine Journal: Using local AI compute to reduce reliance on frontier models, by Chris Green, 30 September 2026