JevSEO, Explained

JevSEO is an open source SEO scoring tool that audits web pages across search engines, answer engines, and generative Artificial Intelligence (AI) assistants. Software developer epergaboni created the project under the MIT license. The software runs semantic evaluations using the Jev decision model from TypeSafe AI.

In software engineering, Jev refers strictly to TypeSafe AI's scoring model. At Layer3Labs, we build and run automated content systems for business clients, where evaluating how automated systems extract information has become as necessary as standard search optimization.

JevSEO handles this challenge by evaluating 45 distinct page signals. It provides separate scores for traditional discovery, direct answer extraction, and citation potential.

The Core Premise and Community Origins of JevSEO

JevSEO is an open-source evaluation tool maintained by epergaboni that scores pages across search engines, answer engines, and generative AI assistants. The project is an independent community tool rather than a TypeSafe AI product. Full documentation lives at jevseo.epergaboni.com.

The software evaluates pages against three distinct computer audiences. Traditional search engines evaluate comprehensive topic coverage and user intent. Answer engines look for compact, self-contained statements. Generative AI assistants search for verifiable assertions with clear provenance.

These three goals frequently conflict. Writing an exhaustive guide satisfies traditional search crawlers. However, long discussions often bury the primary answer. Placing short summary blocks at the top helps answer extraction systems, but that layout can disrupt traditional reading patterns. JevSEO measures how well a single page balances these competing demands.

• Community project: open-source software created by epergaboni under the MIT license, completely independent of TypeSafe AI.
• Three-reader model: audits content specifically for search crawlers, answer parsers, and conversational models.
• Structural tension: identifies areas where optimizing for answer snippets harms traditional document depth.
• Independent reporting: scores each distribution channel separately rather than averaging them into a single metric.

The Three Scoring Pillars for Search, Answers, and Citations

JevSEO separates content evaluation into Search Engine Optimization (SEO), Answer Engine Optimization (AEO), and Generative Engine Optimization (GEO) to measure how well a page satisfies three conflicting audience requirements. Each pillar produces an independent numeric score alongside a prioritized list of remediation tasks. This separation prevents strong metadata from disguising unextractable body copy.

The SEO pillar reviews whether the content satisfies search intent, covers expected subtopics, and provides an original angle. It measures genuine information gain. This score rewards thorough topical depth.

The AEO pillar tests whether an automated parser can extract an isolated, self-contained answer running 40 to 60 words. The text must answer the user prompt immediately. It cannot rely on context from prior headers or following paragraphs. The GEO pillar measures whether assertions include verifiable attribution, clear sourcing, and unambiguous facts. Conversational assistants require quotable evidence to cite a source safely.

• Search Engine Optimization (SEO): measures user intent coverage, thematic breadth, and unique information gain.
• Answer Engine Optimization (AEO): inspects body copy for self-contained answers running 40 to 60 words.
• Generative Engine Optimization (GEO): verifies explicit source provenance, factual assertions, and quotable data points.
• Isolated remediation: delivers dedicated fix lists for each audience instead of generic optimization tips.

The Architecture Split Between Code Checks and Model Judgments

JevSEO splits its 45 diagnostic signals into 20 deterministic code checks and 25 semantic model queries to separate objective technical hygiene from nuanced editorial judgment. The 20 deterministic checks run entirely in local software without consuming network calls or Application Programming Interface (API) tokens. These measurements evaluate objective document characteristics like title length, missing meta tags, heading hierarchy, and indexability flags.

The deterministic half runs in code and never hallucinates. It returns binary pass or fail indicators without model variance.

The remaining 25 signals evaluate nuanced editorial qualities. These qualities include tone, answer placement, clarity, and topical depth.

JevSEO sends these 25 semantic questions to Jev in a single API request evaluated against written rubrics. Jev was trained using Reinforcement Learning for Calibrated Decisions (RLCD) to produce structured outputs rather than conversational prose. All questions evaluate in parallel against the supplied document state. Adding checks does not meaningfully inflate response latency.

• 20 deterministic signals: checks header hierarchy, title length, meta tags, and indexing rules via standard code.
• 25 semantic signals: evaluates clarity, intent fulfillment, and answer placement using written rubrics.
• Zero token waste: computes objective page measurements locally without querying external models.
• Parallel scoring: processes all 25 rubric evaluations concurrently against a single document context.

Full Site Crawling and the Six Content Verdicts

JevSEO crawls multi-page websites to assign one of six categorical actions to every Uniform Resource Locator (URL) while preserving the complete probability distribution behind each recommendation. The crawler reads robots.txt instructions and prioritizes Extensible Markup Language (XML) sitemap entries. It restricts requests to a single hostname origin.

The crawl assigns one of six actions to each page: leave, improve, rewrite, merge, split, or prune. A leave verdict signals strong performance across all three pillars. A prune verdict flags thin or obsolete content.

JevSEO preserves the complete probability distribution across all six potential actions. Operators can review the margin between the top recommendation and the runner-up classification. This visibility reveals close decisions.

Each verdict includes an effort rating, an opportunity ranking, and a calibrated confidence score. Uncertain findings receive lower confidence ratings to flag them for manual editorial review.

• Controlled crawling: obeys robots.txt directives, discovers URLs via sitemaps, and remains within one origin.
• Six action verdicts: classifies inspected pages into leave, improve, rewrite, merge, split, or prune categories.
• Distribution visibility: displays probability scores for secondary actions when a page sits on the boundary between two verdicts.
• Effort scoring: pairs every verdict with an opportunity rating and confidence metric to prioritize developer backlogs.

Execution Speed, Token Costs, and Local Installation

JevSEO audits individual pages in roughly 2.3 seconds at a cost of approximately $0.0001 per page by running semantic questions in parallel against TypeSafe AI's Jev model. The project bases this cost figure on an average consumption of 3,969 input tokens per page. At TypeSafe AI's published rate of $0.042 per million input tokens, 3,969 tokens costs roughly $0.00017.

Output tokens on Jev are free. A crawl across ten pages costs approximately $0.002. This cost advantage grows at higher volumes.

In a workload of 100,000 requests of 1,000 input tokens each with 50 output tokens, Jev costs $4.20 because outputs carry no charge. For that same workload, OpenAI GPT-5.6 Luna costs $26.00 based on $0.20 per million input tokens and $1.20 per million output tokens. Anthropic Claude Haiku 4.5 costs $125.00 for those requests, charging $1.00 per million input tokens and $5.00 per million output tokens.

However, general Large Language Models (LLMs) can write prose and generate code, which Jev cannot do.

TypeSafe AI released Jev in early access on September 15, 2026. A report in Forbes noted rapid developer interest, and Cloudflare published documentation for hosting the model.

Running JevSEO requires Node 22 or higher, pnpm, and a TypeSafe AI credential. Install dependencies using pnpm install and start the local server with pnpm dev. Paste your credentials in the browser interface at localhost:3000/settings. All credentials and crawl records stay on your machine.

• Sub-cent pricing: processes standard pages for fractions of a cent under TypeSafe AI's published rates.
• Rapid turnaround: returns complete multi-pillar audits in approximately 2.3 seconds per URL.
• Comparative benchmark costs: 100,000 queries cost $4.20 on Jev, $26.00 on Luna, and $125.00 on Haiku.
• Local credential security: runs entirely on local infrastructure with API keys pasted at localhost:3000/settings.

Documented Limitations and the Model Accuracy Tradeoff

JevSEO enforces five strict operational constraints, including no content rewriting, no JavaScript rendering, a 28,000-character truncation limit, unvalidated scoring weights, and reliance on Jev's 67.8% benchmark accuracy. The software does not rewrite content, produce replacement sentences, or draft meta tags. It only evaluates existing text against structured rubrics.

The crawler does not execute client-side JavaScript. Pages built with single-page application frameworks return empty Hypertext Markup Language (HTML), creating false warnings. The tool does not predict actual search engine ranking positions. It evaluates rankability, extractability, and quotability instead.

Long articles encounter a strict length cap. The software truncates any submitted document beyond 28,000 characters, discarding trailing text. The scoring weights across the 45 signals are editable defaults rather than validated values.

Semantic checks inherit the accuracy profile of the underlying Jev model. On TypeSafe AI's internal dashboard, Jev aggregates 67.8% accuracy compared to 74.1% for an unidentified best comparator. Task results range from 61.7% to 76.0%.

No independent third-party benchmark exists as of September 22, 2026. TypeSafe AI provides no documented free tier or free credits.

• Diagnostic only: identifies structural flaws without drafting replacement copy or generating metadata.
• No script execution: skips client-rendered JavaScript, leaving dynamic single-page web applications unparsed.
• 28,000 character limit: truncates long-form guides and documentation pages past the 28,000 character boundary.
• Accuracy tradeoff: semantic questions inherit Jev's 67.8% internal benchmark against an unidentified 74.1% comparator.
• Configurable weights: scoring coefficients reflect project defaults rather than scientifically verified ranking formulas.

Who Should Run JevSEO and When to Pick Other Tools

JevSEO serves editorial teams managing static or server-rendered websites, but it cannot support dynamic single-page web applications or teams seeking automated article rewrites. If you manage static documentation, articles, or knowledge bases, the tool provides rapid diagnostic scoring. It exposes where answers are buried or provenance is missing.

The software does not fit teams needing automated copy generation. It cannot audit applications that require client-side JavaScript rendering. Organizations facing dynamic script requirements need a hosted crawler capable of rendering JavaScript in headless browsers.

Teams seeking automated content remediation need a separate text-generation model. JevSEO replaces neither. Our recommendation would change if TypeSafe AI fails to document improved accuracy on independent benchmarks, or if commercial crawlers integrate extraction scoring natively at comparable speeds. To test this open source SEO scoring tool today, clone the repository, configure an API key, and audit a ten-page directory to inspect the verdict distribution.

• Recommended for static publishing: ideal for static site generators, documentation directories, and server-rendered editorial hubs.
• Fast triage: identifies pages failing basic answer extraction criteria in approximately 2.3 seconds per URL.
• Not for dynamic apps: unusable on client-rendered React, Vue, or Angular applications without pre-rendering.
• Strict tool boundaries: replaces neither enterprise JavaScript crawlers nor text-generating language models.