Jev SEO: how we use Jev for SEO content review

Jev SEO is how we stop the same 4 defects reaching your editor twice. They are an opening that never answers the title, an undefined term, an unsourced claim, and an ending with competing asks. A script asks TypeSafe AI Jev 36 fixed questions about every draft. Code counts the rest. The editor decides.

What is Jev SEO?

Jev SEO is our name for using TypeSafe AI Jev, a decision model, to review SEO drafts before they are published. The script sends Jev the title, the meta description (the summary shown under a search result), the target keyword, headings, opening, body and a list of the page’s visual elements. Jev answers 36 fixed questions, such as whether the opening answers the title and whether claims name a source. It returns a probability, a score or a choice for each one.

Jev is not a chat model and it does not write. TypeSafe describes it as a System One model that picks a defined answer from supplied evidence. That is what a review question needs. The model itself is covered in What is Jev AI. This article covers the SEO use only.

SEO is search engine optimisation, the work of helping people find a page in search. AEO and GEO, answer engine and generative engine optimisation, ask whether an answer tool can quote the page. All three need the same things from your draft: an early answer, defined terms, sourced claims and one next step. Our SEO and AI search service holds client content to that standard. The review script holds ours to it.

Code counts, Jev judges, the editor decides

The review runs in two layers, because a language model cannot count reliably. The first layer is plain code. It measures lengths, counts headings and tables, finds banned characters and computes a reading score. The reading score is Flesch reading ease, a 0 to 100 scale where higher is easier to read. Every check is a fixed threshold, run before any model is called.

The second layer is Jev. It takes the questions code cannot answer. Does the opening name a problem you already have? Do the headings tell the story on their own? Is the next step the right size for you? Scores and choices carry a confidence value, and the script turns the answers into six lens scores.

The six lenses Jev SEO scores

Every question belongs to one lens, except one that only classifies search intent. A lens score is the average of its answers, from 0 to 100. The gate compares each score with a floor for the article type. The draft fails if the overall score is under its floor, or if more than three lenses miss theirs.

The readability floor is a target, not a block. Most of our guides have scored between 50 and 60 on it since the gate was introduced, and none has reached 65. The miss is reported, not hidden.

What Jev SEO catches, and what it misses

Jev catches on the first read what your editor finds on the third. In our drafts the common flags are an opening that defines the topic instead of answering the title, a term used before it is explained, and a claim with no source beside it. It also catches a quieter one: a pitch that leads with the technology instead of what you save.

Jev misses anything it cannot see. It does not crawl the live site, read Search Console (Google’s free report of search traffic) or know how Google will treat the page. It reviews the draft text and the list of visuals the script gives it, nothing more. It does not verify that a linked source says what you claim, so the editor still opens the link. It cannot count, which is why code does the counting.

Jev can also be wrong. A question about voice or proof is a judgement, and a confidence of 0.6 is a weak one. When a flag looks wrong to you, read the passage, decide, and move on.

The review workflow we run, step by step

Autonomous Technologies is a Vancouver-founded Shopify agency that publishes operator guides for ecommerce teams. Every guide goes through this loop before the founder sees it in the CMS, the publishing system. You can run it with your own questions.

1. Brief first. Name the reader, the question the title must answer, the evidence on hand and the one next step. Jev scores the draft against the title, so a vague title gives you vague flags.
2. Run the gate on the Markdown source. One command runs the mechanical checks, then the Jev call, and prints pass or fail with every miss listed. A run takes a minute or two; we have not timed it formally.
3. Fix hard fails before reading the lenses. A missing caption or a 62-character title is a two-minute fix. Clear those first.
4. Read each flag in context. The editor opens the passage, decides whether the flag is right, and changes the draft or leaves it. Keep a short note per run saying what changed and why.
5. Re-run, up to five times. Each run is logged with its scores. If the draft still fails after five, the problem is the brief, not the wording. Go back to step 1.
6. Render the page before publishing. A source file cannot show whether the tables, image and call to action render. Open the page at desktop and 390 pixels wide and check that the structured data, the machine-readable summary of the page, matches what a reader sees. Google’s structured data policies require that match, read 9 October 2026.

Publishing stays a separate decision. The script writes a log row. A person presses publish.

A worked example: this article’s first gate run

On 9 October 2026 the previous version of this article went through the gate as an operator guide. It failed. These are the real scores from that run.

Visual was the biggest gap. The old body had one table, one image and no callout, so the emphasis blocks counted as absent and the captions as labels. Structure missed because the headings named topics rather than making points. SEO missed because the opening defined the subject instead of answering the title.

The fix is the article you are reading: captioned tables, two callouts, headings that state a point, and an opening that answers within 55 words. Its own score is not quoted here, because quoting it changes the text being scored. Whether it changes search traffic is not measured yet. Search Console gets re-pulled after 20 October 2026.

Where Jev SEO breaks

The main hesitation you will have is cost and effort: another tool, another step before publishing. The step costs a minute or two per run, and the API call, the request the script sends to Jev, is metered per request. The effort is in the first week, writing the questions down. After that they do not change.

It breaks when the questions are vague. “Is this good?” produces noise. “Does the opening answer the title within its first hundred words?” produces a flag you can act on. Every question in our script names one passage and one criterion.

It breaks when you treat a score as a verdict. A draft can clear every floor and still be wrong about a fact, because Jev does not check facts against sources. Google’s people-first content guidance, read 9 October 2026, puts first-hand expertise at the centre of helpful content. No review script supplies that. It only makes your standard easier to apply the same way every time.

And it breaks on a thin brief. A draft with no evidence to cite fails the sourcing question every run. Go back and get the evidence.

Thresholds and gate scores above were read from our own scripts and logs on 9 October 2026. Re-check by 9 April 2027, or sooner if TypeSafe changes the Jev API.