Quick Answer: Yes, Jev AI is good for SEO and AEO judgment gates once the rules are written down, scoring 96% on our 92-case blind benchmark. Every answer at 0.85 confidence or above was correct; bare questions scored 89%. It cannot write, browse or abstain, so it sits inside a pipeline, not in place of the agent running it.
Most of what an SEO or AEO team calls analysis is a stack of yes/no calls. Does this topic cannibalize an existing page, does this paragraph open with the answer, does this sentence hand a win to a competitor.
We make those calls today with general-purpose LLMs, and they are slow, expensive and inconsistent between drafts. Jev AI, the model TypeSafe AI released on 15 September 2026, does only that kind of call and returns a calibrated probability. We spent a day benchmarking whether the probability is worth trusting for SEO work.
What is Jev AI and what can it not do
Jev AI is a judgment model: it takes a state plus typed questions and returns probabilities, never text. TypeSafe AI calls the product System One, and the Jev API exposes it as a single endpoint. A Noul is a yes/no returning P(true), a Choice picks from up to 255 options with a probability each, and a Score rates on ordered levels.
That constraint decides how you use it. Jev cannot summarise a page, extract a brand name, explain its reasoning or make a tool call; one request is one state in and one set of numbers out. It is not fine-tunable, so what you know about a client has to travel inside the question every time.
TypeSafe lists input at $0.042 per million tokens with output free, and quotes 70 to 500 ms per call. Our median from India was closer to 880 ms. A 92-case benchmark was still a sub-minute, half-cent exercise.
The pattern that makes it useful
Jev sits at the one step in a pipeline where a judgment is needed, and nowhere else. An agent gathers data and assembles a state, Jev judges every item, the agent applies a threshold, and a text model explains only what a human must read.
For a cannibalization check the orchestrator still pulls GSC queries, searches the corpus and fetches candidate pages. Jev then answers "would a searcher of the new topic's keywords be satisfied by this existing page" for every pair, in parallel, in under a second. That is the step most likely to be sampled rather than run in full when twenty pages will not fit in context.
How did Jev score on the SEO and AEO benchmark
On a 92-case blind benchmark across 11 suites, the Jev model scored 89% with bare questions and 96% with principle-based rubrics. The suites mirror the calls an account manager makes weekly, from cannibalization and search intent to the answer-first and concession checks in our AEO audit checklist.
| Suite | Cases | Bare question | With rubric |
|---|---|---|---|
| Cannibalization | 14 | 100% | 93% |
| H2 restates H1 | 8 | 75% | 100% |
| Search intent | 12 | 67% | 83% |
| Answer-first paragraph | 8 | 100% | 100% |
| Competitor concession | 8 | 100% | 100% |
| SEBI compliance | 8 | 88% | 100% |
| Needs primary source | 6 | 100% | 100% |
| Citation domain type | 8 | 88% | 100% |
| Brand sentiment | 8 | 75% | 88% |
| Competitor detection | 8 | 100% | 100% |
| Rerank chunks | 4 | 100% | 100% |
| Total | 92 | 89% | 96% |
A separate agent that never saw the rubrics wrote the cases, about a third of them deliberately hard: adjacent topics, mixed-sentiment sentences, a regulator that looks like a competitor. Both passes together used about 118,000 input tokens, roughly $0.005.
Where the rubric earned the seven points
Every case the rubric fixed was a missing rule, not missing intelligence. The bare question called "Understanding Company Registration" a distinct H2 under the H1 "Company Registration in India". The rubric states that overview and "understanding X" headings restate the H1, and the answer flipped.
The bare question read "protect your capital" as neutral finance copy. Once the rubric quoted the ad-norm rule on implied safety, it became a breach at 0.93. A multi-brand supplement retailer typed as a brand's own store became an aggregator the moment the rubric said multi-brand retailers are aggregators.
Not one fix came from asking the question harder. This is the same discipline as structuring content for AI: the rule has to be on the page, or in the question, in words.
Is the confidence number actually trustworthy
Yes, and this was the more important result. With rubrics, 71 of 92 answers came back at 0.85 confidence or above, and all 71 were correct. Bare questions produced 64 answers in that band, also all correct.
Everything Jev returned below 0.7 was right between 50% and 80% of the time, which is coin-flip territory, and it said so by reporting the low number. The pattern held across three runs on different case sets, so it is not an artefact of one batch.
That is what turns a judge into a gate. Auto-accept above 0.85, human review from 0.5 to 0.85, reject below, and an account manager clears 77% of judgments with zero errors while reading the other 23%.
The four misses, and whose fault they were
Two were taxonomy disagreements between humans. Jev called the query "company registration" transactional at 0.87 because our rubric says bare service nouns are transactional, while the test author had labelled it commercial. "ELSS vs PPF" went commercial at 0.86 where the author said informational; both are named investable instruments, so the rubric needs a sharper line.
One was a genuine regression: a borewell-water purifier topic against an existing RO/UV page came back 0.66 cannibalize, and the bare question had it right. One was a mixed sentence, "the most widely used, though users report slow turnaround", which landed at 0.50 neutral. Every miss sat below 0.87 and three of four below 0.7, so a threshold catches all of them.
How we kept the benchmark honest
A judgment model will score 100% on any test whose answers appear in its rubric, so the protocol matters more than the score. We built the evaluation so that the rubrics and the test cases could not have seen each other, and the 96% above is the result of that protocol only.
Rubric rules are stated as principles, never as lists of cases. Worked examples are drawn from unrelated domains such as HR software and hotels, so nothing in a rubric resembles a legal, finance or supplement page. A separate agent with no sight of the rubrics wrote all 92 cases in our client verticals.
That separation is what makes the number usable for planning. It also sets the bar for any vendor benchmark you read: if the same author wrote the questions and the rubric, treat the score as rubric-following, not generalisation.
Which Jev use cases should not run alone
Three failure modes showed up consistently, and none is fixed by a better rubric. The first is "none of the above". Asked which of four chunks answers "how long does trademark registration take", Jev picked one at 0.77 even though none stated a duration.
A separate presence Noul per chunk, "does this state a duration?", returned 0.01 to 0.03 for all four. So the gate works, but only if every Choice question is paired with a presence check.
The second is that a rubric gap produces a confident wrong answer, not a hedged one. The retailer typed as a brand store came back at 0.90 before the aggregator rule existed. Jev signals uncertainty about ambiguous inputs, not about rules you forgot to write, so each rubric needs its negative cases enumerated and an owner per client.
The third is local intent. A Bangalore-specific page against a national page came back 0.67 cannibalize, and adding "SERP shows a local pack" to the state only moved it to 0.58. Geo-variant intent stays a code rule, not a judgment.
Which SEO gates are worth wiring Jev into first
The gates that are hardest to hold by hand are the ones Jev already clears at 95% or better on unseen cases. Concession detection caught the parity pattern ("both X and the client handle tables well"), the subtlest way a comparison page hands a rival a win. The H2-restates-H1 check caught the overview-heading pattern that slips past most editors because it reads naturally.
Cannibalization is the strongest suite overall, 34 of 36 across all three runs, even with only titles, H2s and GSC queries in the state. Search intent and mixed sentiment are the weakest and should be tiebreakers. In order: concession, client hard rules, H2=H1, answer-first, cannibalization, then everything else behind a review band.
That order matters more for AEO than for classic SEO, because the failures it catches are the ones that stop a page being quoted. The AEO audit versus SEO audit comparison covers why, and the GEO metrics page shows what a cleared gate should move.
Why we ran the experiment
Unveilr delivers AEO as a managed service on proprietary AI agents, built by a team of IIT graduates and backed by AJVC, with engagements from ₹65,000 a month. Every rupee of margin in that model depends on judgment steps being cheap, consistent and exhaustive rather than sampled.
A judge that costs half a cent per hundred decisions and flags which ones to double-check changes what "reviewed" can mean. The method is the same one our answer engine optimization guide describes; the gates just run on every chunk now instead of a sample.