A new AI model dropped on September 15, 2026, and most SEO and GEO practitioners haven’t noticed yet. It isn’t a better chatbot or a smarter writing assistant. Jev, from TypeSafe AI, doesn’t generate text at all. It makes decisions — and it does it 193 times faster and 445 times cheaper than the frontier models GEO teams are currently using for the same job. For anyone running AI visibility programs at scale, that cost and speed difference isn’t a minor optimization. It changes what’s economically feasible.
What Jev Actually Is
TypeSafe AI calls Jev a “System One model” — a deliberate reference to Kahneman’s fast, intuitive thinking. Where large language models generate answers token by token, Jev returns one of three output types: a yes/no answer with a calibrated confidence score, a category label picked from a predefined list, or a numeric score on a defined scale. That’s the complete output set. It cannot write a paragraph or explain its reasoning. What it can do is run through 20,000 keyword classification decisions for $0.07, or scan 12,000 AI responses for brand citations for $0.34, at latencies between 70 and 500 milliseconds.
The model was built by Diogo Almeida, a co-inventor of RLHF (reinforcement learning from human feedback) at OpenAI, alongside Erik Gafni and Sasha Sheng. TypeSafe raised $40 million in its seed round and trained Jev using a method they call RLCD — reinforcement learning for calibrated decisions — which produces confidence scores that are mathematically calibrated rather than overconfident in the way typical LLM outputs tend to be. The zero-hallucination claim holds technically: Jev returns typed values, not open-ended text, so it cannot produce factual fabrications by construction.
Why GEO Teams Should Pay Attention Now
The operational problem in GEO is volume. A credible AI visibility monitoring program needs to check whether your brand is cited across ChatGPT, Claude, Gemini, and Perplexity — across hundreds of queries, at a frequency that detects changes. Running that against a frontier LLM costs real money and takes real time. Most teams either run too few queries to detect shifts or batch checks infrequently enough that they’re always reacting to stale data.
Jev prices input at $0.042 per million tokens with output charged at zero. At that rate, the economics of GEO monitoring change substantially. The Ryze AI team published benchmarks showing a nightly citation check across 4,000 answers from four AI engines runs to approximately $0.42. Running that same check weekly with GPT-5 class models costs roughly 80x more. The difference is between running a monitoring program daily and running it when budget permits.
The Eight GEO Jobs Jev Can Run Nightly
The use cases that fit Jev’s architecture share a pattern: there’s a bounded set of possible answers, the input is structured text, and you need to process hundreds or thousands of rows. GEO generates exactly this kind of work at every stage.
Citation monitoring at scale. For each AI response in your monitoring set, Jev answers: does this response cite your brand? A yes/no output with a confidence score. At $0.42 per 4,000 answers, you can monitor daily rather than weekly. The citation rate data you generate over weeks and months becomes the leading indicator that currently requires expensive manual audits to surface.
Competitor share-of-answer tracking. The more strategically useful companion to citation monitoring: when you’re absent, who gets cited instead? Jev classifies each answer for competitor mentions using the same classification format. The output is a share-of-answer table — per query category, per competitor — that tells you where the replacement visibility is going. This is information that currently gets assembled manually, quarterly, if at all.
Page answerability scoring. One of the structural determinants of whether a page gets cited in AI responses is whether it answers the relevant question in the first 40 words. Jev can score each page in your content index on a 0–100 answerability scale against a list of target queries. Pages in the bottom quartile go to a revision queue; pages above threshold stay. The rankspot.ai team benchmarks this at $0.21 for 10,000 pages.
Title-to-question alignment. A page can answer a buyer’s question well but use a title that doesn’t match how the question is phrased in AI prompts. Jev scores title-question alignment as a classification: strong match, weak match, or mismatch. Weak and mismatch pages get flagged for title revision independent of content quality. This diagnostic currently requires manual review or expensive embedding comparisons; at Jev’s cost, it runs across your full catalog for under $0.10.
Community thread triage. Reddit, Quora, and industry forums are significant AI citation sources. GEO teams that want to identify threads where competitors are being recommended exclusively — threads worth engaging with — currently scan manually or rely on keyword alerts. Jev classifies each thread: competitor-only recommendation, brand-inclusive, or neutral. The high-priority threads surface automatically; the rest drop.
Source type fingerprinting. When AI engines cite a category of content — vendor comparisons, Reddit threads, analyst reports, product documentation — knowing the dominant source type tells you where to invest. Jev classifies each cited URL by content format. Run this monthly across your share-of-answer data and you get a source-type map that shows whether the AI engines in your category are primarily citing peer review sites, media, or owned content. The content investment implications are direct.
Mention sentiment classification. Not all citations are positive. Jev classifies each brand mention as positive, neutral, or negative and flags negative citations with source attribution. This distinguishes visibility gaps (you’re absent) from positioning problems (you’re present but in the wrong framing). Both require action, but the action is different.
Content gap identification. For each buyer question in your query set, Jev determines whether an existing page adequately covers it: covers well, covers partially, or no coverage. The partial and no-coverage buckets become your content calendar. At $0.21 for 10,000 questions, this runs across a comprehensive query set rather than the 50-100 questions a strategist can reasonably assess manually.
The Untapped Angle: Reasoning-Mode Readiness Scoring
Extended-thinking AI changes which content gets cited (discussed in detail in yesterday’s piece on reasoning lift). High-reasoning AI systems select for evidence density, sub-question coverage, and deep problem-definition framing. These are classifiable qualities — Jev can score a page for reasoning-mode readiness without reading the content in full, using a bounded rubric: does the page contain quantified claims? Does it define a problem before offering a solution? Does it address multiple related sub-questions rather than a single narrow query?
A Jev-based reasoning-mode readiness score across your content index would give you a prioritized optimization list — pages that are one structural change away from better reasoning-mode citation performance versus pages that need substantive rewriting. No one appears to have productized this yet; the data pipeline that would produce it is straightforward to build with Jev’s classification outputs feeding into your CMS or content calendar system.
What Jev Cannot Do
The limitations matter. Jev returns no explanations — you get a probability and a label, not a rationale. It requires predetermined answer sets: you cannot ask an open question and receive an unexpected answer. Its published accuracy is 67.8% on vendor benchmarks, which TypeSafe characterizes as “triage-grade” — sufficient to route work into human review queues, not sufficient to replace human judgment on individual decisions. It is on a waitlist as of September 2026, not generally available.
The right mental model is a very fast, very cheap first-pass filter. Jev narrows 10,000 rows down to the 400 that need attention. The attention itself — the writing, the editorial judgment, the strategic call — still runs through an LLM or a human. The cost savings come from not running every row through the expensive step.
Getting Started
TypeSafe’s early access waitlist is at typesafe.ai. The documentation describes three API primitives: noul (yes/no with confidence), choice (category classification), and score (numeric on a scale). The recommended workflow for SEO and GEO applications is: collect inputs via Search Console, Screaming Frog, or your AI monitoring pipeline; pre-filter to remove clear negatives; submit the remainder to Jev; apply confidence thresholds to route outputs between automatic action, human review queue, and drop; feed what passes to an LLM for any writing work required.
For GEO teams running citation monitoring, the day-one application is replacing per-response LLM classification calls with Jev classification calls for the yes/no brand citation question. The accuracy tradeoff is real — you will miss some citations and falsely flag some non-citations — but at a 67.8% accuracy on a 70ms/500ms latency at essentially zero cost, the check becomes daily rather than weekly and the false positives go to human review. That is a better program than a perfectly accurate check you run when budget allows.
To see where your brand currently sits in AI citation coverage — and identify which content gaps are creating the most visibility loss — run a free audit at ai-visibility.llmagnet.com.