Jev for SEO: intent labels, audits, and page quality

Closed questions over SEO data

TypeSafe’s API takes state and a typed question. Choice picks from options you list. Score rates a rubric. Noul returns a probability for a statement. The models page lists jev-1.13.0 and jev-1.14.0 at $0.042 per million input tokens, with output free.

TypeSafe’s jaggedness page for jev-1.13 says Jev is not a calculator, reads dates as text, and is not trained to generate text. Accuracy falls when the state fills with material the question does not need. The write-ups below keep crawls, joins, regex splits, and date windows in code, and send Jev a short passage.

Sources:[1] TypeSafe AI[34] TypeSafe AI[48] TypeSafe AI

Tenpoint Labs: five jobs on 3,566 answers and about 30,000 queries

Mikal Khoso of Tenpoint Labs, writing on 9 October 2026, ran jev-1.13.0 on five SEO jobs between 3 and 9 October. The tests use 3,566 AI answers and about 30,000 search queries and keywords. tenpointlabs.com supplies 2,944 of the queries: Search Console web search from 9 June 2025 to 9 October 2026. Search Console omits anonymised queries. No GA4 data is used. Client and prospect accounts are described by industry and never named. Tenpoint Labs is a TypeSafe customer.

Brand mentions: 140 of 140 sampled named verdicts were correct. About 6% of real mentions were lost because a search step found a short form of a name and the question asked about the stored long form. Search intent: about 86% agreement with a reader on 14,902 queries across five sites Jev had not seen. A paid classifier scored about 48% on an earlier set of 1,326 keywords. Jev agreed with that classifier on 54.9% of keywords. On 120 sampled disagreements, a blind reader sided with Jev on 86 and with the classifier on 17.

Query fan-out: two readers labelling the same 30 queries disagreed on 24 of them. The first question asked whether an AI had written the query. Only 5 of 7,157 queries reached probability 0.8. A branded versus non-branded split on Tenpoint’s own 2,944 queries agreed with a regex on 99.4%; Khoso keeps the regex. Keyword relevance with 16 labels on 2,804 keywords from one site was roughly level with a larger model. The post calls that unproven beyond one site.

In production they count 0.8 or higher as a yes and 0.2 or lower as a no, and send the band in between to a larger model or a person. A search query with one question used about 420 input tokens. An AI answer with its questions used about 1,040. Median response time was near 400 ms. Khoso labelled 120 rows himself after a larger model labelled a sample. He agreed with that reader on 26 of 30 brand mentions, 21 of 28 intent rows, 14 of 24 keyword-relevance rows, and 6 of 30 fan-out rows.

Sources:[180] Mikal Khoso[34] TypeSafe AI

Nico Gorrono: three tests on two live sites

Nico Gorrono, writing 24 September 2026, ran three jobs on two live sites. On airankingskool.com he had 164 pages and 3,863 paragraphs. A local embedding model shortlisted four pages per paragraph. Jev judged 3,182 paragraphs for $0.12. Rules required at least 70% confidence, at most three new links per page, and at most eight per target. The first run proposed 147 links. Gorrono approved 109 and rejected 38. Rejections included hub pages taking a competing post’s keyword, anchors inside quoted prompts, and related targets that would mislead. On two different days Jev made the same pick 96.7% of the time. The Claude Code skill README at commit 5f76ff2 reports the same 3,182 decisions on that 164-page site, 112 links proposed, and $0.75 including Claude for anchors.

On propiedadesaysen.com he pulled 90 days of Search Console query and page data through Windsor.ai. Python found 64 searches where at least two pages each got impressions. Jev read the query plus a short snapshot of each page. At 70% confidence the split was 26 wrong page winning, 19 not sure, 18 right page already winning, and 1 splitting signals. The query "patagonia real estate" was searched 255 times. Google showed the English homepage 194 times at average position 12.5. Jev gave the dedicated Patagonia real estate page an 81% probability of being the owner, a page Google had shown 4 times at position 52. The homepage took 14 "for sale" searches the for-sale page should own: 633 impressions. Only 3 of 64 fix suggestions (leave, strengthen, differentiate, or merge) reached 70% confidence. The Jev bill for all 64 searches was $0.003.

For ChatGPT visibility he pulled 40 answers, 20 in Spanish from Chile and 20 in English from the US, through DataForSEO at $0.004 per answer. Jev found the site mentioned in 11 and recommended in 2. The mentioned-or-not call matched a plain text search on 40 of 40. Eight answers had no mention where Gorrono judged one was plausible. Forty answers cost about 16 cents to fetch and a fifth of a cent for Jev.

Sources:[181] Nico Gorrono[182] Nicolas Gorrono

jev-seo: a crawl, 52 rules, and 13 questions per page

Agrici Daniel’s jev-seo, at commit 55a184a, crawls from a homepage, checks 52 rules tied to Google Search Central, measures Core Web Vitals through PageSpeed Insights, and asks Jev 13 questions per page and 5 per site. Standard mode costs about a cent in Jev per site. An optional --full mode adds rankings, keywords, and backlinks from DataForSEO for about $0.30. The default page cap is 60, and the report says so when a site is sampled. Scores rank work. They do not predict rankings or traffic. Without a TypeSafe key the audit still runs and marks the Jev sections as not assessed.

evaluation.md is dated 2026-09-22. Fifty-nine pages judged twice: page type 58/59 identical, intent 59/59, suggested action 54/59. Score answers moved by 0.001 to 0.023 on average. When both runs were decisive on page type they agreed 44/44. Against a blind second model on 30 pages and 40 keywords, helpfulness matched on 29/30 at a 0.45 threshold and specificity on 28/30. After wording changes, decisive helpfulness and specificity were 28/29, keyword relevance 27/30, opens with the point 23/23, and clear next step 16/16. The second judge is another model. The file says there is no human-labelled answer key yet. GitHubDaily posted the tool on 30 September 2026 and rounded the two-run score gap to 0.03 on those 59 pages.

Sources:[183] Agrici Daniel[184] Agrici Daniel[185] GitHubDaily

SERP Lens: page quality for under a penny a page

Sam Underwood, writing in October 2026 and posting on 20 October, put Jev in SERP Lens to score page quality for under a penny a page. The rubric draws on Google’s helpful content guide and quality rater guidelines. Rows include Depth, Production quality, Padding, Gets to the point, Written for people, Supported claims, and Trust signals. Capture strips sidebars, headers, footers, and cookie banners so Jev sees main content. Questions vary by page type: an article is asked how quickly the main content delivers the title’s promise, a product page how quickly a buyer sees what the product is, its price, and its availability.

The first attempt used many calls per page, scored Wikipedia articles as low quality, and could differ across two runs on the same page. Overlapping rubric levels hid about half of the quality results. After one request per page, clearer levels, and a tighter capture, agreement with reviewers within one level moved from 64% to 82%. Reviewers were Underwood plus Astra, Fable, and Opus 5.5. He writes that Jev agreed with that group as much as they agreed with each other. The post does not print the held-out page count. SERP Lens also uses Jev to check whether a page answers Google Search Console queries, People Also Ask questions, and AI Overview text, and to label SERP formats and site types. Underwood builds SERP Lens.

Sources:[186] Sam Underwood[187] Sam Underwood

What stays in code

Across these write-ups, crawls, Search Console joins, regex brand splits, word counts, and date windows stay in code. Jev judges meaning on a shortlist. A person still approves internal links, merges, and redirects. TypeSafe’s coding-agents page says Jev is not a drop-in replacement for the model behind Claude Code, Cursor, or similar tools. Keep that model for writing and use Jev for typed decisions.

Sources:[162] TypeSafe AI[48] TypeSafe AI[181] Nico Gorrono[180] Mikal Khoso

Questions about Jev for SEO