← back to all posts

Chris Green Built a Chrome Extension for This. In Orakul It's Just Text

Chris Green Built a Chrome Extension for This. In Orakul It's Just Text

On September 30, Search Engine Journal published Chris Green's "Using Local (AI) Compute To Reduce Reliance On Frontier Models." The core claim: "Not every SEO task needs a frontier model." I agree with it enough to have built a product on that principle. Below: what Green actually proposes, where my stack doesn't cover his design, and how the same architecture gets assembled in text instead of code.

What Green proposes

He describes three tiers of compute. Deterministic code handles exact operations. Light interpretation goes to a small model running locally — in his case Gemini Nano, built into Chrome. And only where real judgment is required does a frontier model get called. In his words:

"Exact computation happens in code, lightweight intelligence happens locally, and expensive intelligence is called only when it is truly needed."

He lists the routine work that needs no frontier model outright: extracting and deduplicating URLs from an XML sitemap, fetching pages and comparing HTML, checking response codes, matching elements, identifying canonical relationships, detecting destination changes, comparing raw HTML with the rendered DOM, and analyzing link attributes — anchors, destinations, broken links.

Look at that list and answer honestly: how many of those items are you paying a frontier model for right now?

Where his own test broke his design

The most valuable part of the article isn't the idea — it's that he published the failure. He built a Chrome extension that examines discrepancies between raw HTML and the rendered DOM, then pitted Nano against Gemini Flash and ChatGPT Luna on the same evidence. Nano proved unreliable for final judgment calls, while the larger models, in his words, handled the same evidence "considerably better."

His conclusion wasn't "local models don't work" but an architectural one: code computes, Nano presents and lightly interprets, a larger model steps in where judgment is needed. And he states a separate requirement for a model of any size:

"It has to respect the facts... and avoid inventing information."

What I don't have

Let me say this upfront, because what follows is an ad for my own product. Orakul does no local compute. Of Green's three tiers, two are implemented: tools that fetch the data, and a model that processes it. A small local model is still only on my roadmap.

So I'm not offering a finished implementation of his design. I'm offering the part of it that, in my view, delivers most of the effect — and I'd put the principle more bluntly than he does: the model is a processor of data, not a source of it.

What this looks like in Orakul

Green's case study is a Chrome extension he wrote in code. My closest equivalent is a technical audit agent: it takes a URL and returns a finished report. It works like this:

In steps one through four, the data is fetched by tools, not by the model. The model appears at step six and writes the text from what the tools collected. That's Green's split exactly, minus the local tier.

The difference is in the assembly. An extension has to be written and maintained. An agent in Orakul is a table: step order, instruction in plain text, selected tool. Zero code.

How to check that the model isn't making things up

Green's "respect the facts" requirement sounds good, but you need a way to verify it. I have measurements from a live store, which I wrote up in the contrast analysis piece. Three facts from it.

The agent states its own sample size. Across four page reports it wrote that it managed to open 4 competitors out of 8, 6 of 9, 6 of 9, and 3 of 9 — Pinterest, Scribd, and direct PDFs return no HTML. A model that invents data doesn't add that caveat; it has nothing to constrain.

Agent report: instead of a list of edits, a finding about intent mismatch and too small a sample
Where the data ran short, the agent refused to produce a checklist. Three pages out of nine isn't a sample, and it said so itself.

The agent refused the job. For "bass clef note naming worksheets pdf," the results were free PDFs while our page was a paid product. Instead of a generic list of improvements, it returned a finding about intent mismatch and called the sample insufficient. A refusal is what respecting the facts looks like in practice.

The numbers come from a connected source. The agent pulls the set of pages ranking 10–20 from Google Search Console, not from its own head.

Agent chat in Orakul: a Google Search Console table with pages and average positions
This table came from Search Console through a connected tool. The model formatted it; it didn't invent it.

And where no demand data existed, the report says "Wordstat: zero, Bing: insufficient data" instead of a plausible-looking number. For a report you later show a client, that matters more than polish.

What it costs in hours

The audit agent removes roughly three hours of manual work per report. At ten audits a month and a junior rate of $5 an hour, that's about $150 saved monthly. A modest figure, but one that's calculated rather than promised. And that's only the effect visible on a timesheet; the effect of not calling a frontier model to parse a sitemap shows up separately, on the API bill.

What to take away

If you want to build an agent like this in text, with no code, it lives in Orakul. And Chris's article is worth reading in full: it's a rare case of an author publishing both the design and the place where it didn't work for him.

Maxim Safianov
Maxim Safianov

I work at the intersection of technical SEO and software engineering: helping sites stay visible for classic search and AI answers alike, and building the Python tooling that automates it.

No comments yet — be the first to share your thoughts.

SEO

Mueller: "AI Systems Build on SEO." Why the Technical Base Decides GEO and AEO Too

On November 2, Search Engine Journal published an interview with John Mueller: technical SEO remains the foundation for AI search too, and a strategy built entirely on listicles will one day simply switch off. Here is what he said verbatim, what retellings add on his behalf, and why I agree — starting with a site that spent 89 days unreachable for crawlers for three server-side reasons.

05 Nov 20266 min
SEO

Did Your Rank Tracker Show a Collapse in Late September? Google Likely Served It a Fake SERP

Nozzle analyzed 27.5 million SERPs and found that from September 20 Google began serving rank trackers substituted results — no errors, no captcha, outwardly normal. At the September 29 peak, 27% of SERPs were poisoned, and the tracked site fell out of the top 10 in 58.6% of them versus 2.7% in the baseline. Here is the data, a three-minute test to tell a poisoned SERP from a real drop, and why I stopped tracking positions.

14 Oct 20266 min
GEO

Google's Top 10 Still Feeds AI Answers. Three Signals the Rules Are About to Change

The common view is that AI answers have already changed search. Under the hood it is still the same two-stage ranking pipeline. I checked this on data from anastasiya.studio: 74% of pages with impressions in AI Overviews and AI Mode rank in the top 10 and collect 92.6% of AI impressions. But Google has already published the theory that removes that pipeline, started paying for contributions to AI answers, and opened a Search Console report for them. Here are the three signals and what to do right now.

17 Sep 202610 min