Check your company's visibility in AI
For free, no credit card required.
Most people use AI to write, explain, or summarize something. Jev from TypeSafe AI serves a different purpose: making specific decisions based on provided information. For teams focused on visibility in AI, this opens an interesting area of applications - from organizing customer questions to analyzing responses and evaluating content.
What is Jev?
Jev is the AI model from TypeSafe, designed to return structured decisions that software can directly utilize.
You provide it with information for evaluation, specify questions, and acceptable types of answers. The model returns a choice, rating, or probability instead of freely written text. This is how its operation is described by TypeSafe documentation.
Imagine analyzing a hundred AI responses regarding your industry. You need to determine whether the brand was merely mentioned or actually recommended. You can ask the generative model to describe each response, but later you still need to translate these descriptions into comparable data.
The Jev approach involves defining categories in advance. The result can then go into a table, report, or the next stage of analysis.
Why does TypeSafe call it the System One model?
TypeSafe introduced Jev on September 15, 2026, as the first public model from the company's developing System One category. The name refers to the distinction between fast, intuitive judgment and slower, more elaborate reasoning.
According to the manufacturer, Jev was designed for automation, using a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. The company also describes parallel scoring instead of sequential text generation.TypeSafe announcement
In practice, it involves well-defined questions: what category does this case belong to, how to evaluate it according to the given criteria, and whether the provided information supports a specific statement.
How does Jev work? data, questions, and results
Working with the model starts with providing context, referred to in the documentation as state. This could be, for example, an AI assistant's response, a snippet of an article, or a customer's question.
You add questions to this context. Jev supports three basic types:
| Type | Purpose | Example use in content work |
|---|---|---|
| Choice | Choosing one of the specified options | Assigning the question to the topic: price, implementation, comparison, or support |
| Score | Rating according to described, ordered levels | Assessment of how well the fragment answers the question |
| New | Probability of the truth of the statement | Check if the provided text contains a price condition |
Several questions can be combined in one call. Each is evaluated independently against the same context. Introduction to Jev
The most important work therefore takes place before the model is launched: we need to determine exactly what we want to measure.
What does Jev have to do with GEO and AEO?
GEO relates to working on brand visibility and its content in generative AI responses. AEO focuses on preparing information to effectively answer user questions and be utilized by answering systems.
Both areas involve a lot of repetitive analysis:
- What questions do potential customers ask?
- Does the AI response recommend the company, or just mention its name?
- Does the article actually solve the problem stated in the title?
- Which sections require review by an editor?
Jev can be a component of a tool performing such tasks. However, its result is not a measure of the chance of being cited in ChatGPT or Google.
The following applications are editorial proposals from Semly based on the model's operation. They do not represent the results of our tests or a finished integration of Jev with Semly.
1. Organizing customer questions
The list of queries from search engines, sales conversations, and customer service quickly becomes difficult to browse. Some questions address the same issue, even though they are phrased differently.
Jev can be used to assign questions to previously defined categories.
For a company running a website on WordPress, this division could look as follows:
| Question | Proposed category |
|---|---|
| How much does monthly website maintenance cost? | Price |
| What does the subscription include? | Scope of service |
| How to transfer the site to a new agency? | Starting cooperation |
| Will a subscription or hourly billing be better? | Comparison of options |
This is an illustration of the expected classification method, not the result of the model invocation.
An organized list helps to notice which needs are already met by the site and which require new material. However, categories need to be checked against actual questions - especially those that fit into more than one group.
2. Distinguishing mention from recommendation in AI responses
The mere appearance of the company name says little about how it is presented.
Compare two sentences:
Company X operates in this market alongside several other providers.
For the described application, it is worth considering Company X due to the scope of service.
The first is a mention. The second contains a recommendation and justification.
In the proposed analytical process, Jev could classify the recorded responses according to clearly defined categories. It is also worth checking the presence of the brand, the recommendation, and the link to the source provided by the assistant. These are different pieces of information.
The model must receive material for analysis. If you are evaluating citations, also provide a list of sources. Based on text without links, it is impossible to reliably determine which addresses the assistant indicated.
3. Assessment of whether the content answers the question
An article may have an appropriate title and still not provide a useful answer.
Let's assume you are analyzing the guide "How much does it cost to maintain a WordPress site?" The general statement "the price depends on many factors" does not explain to the reader what makes up the cost.
Three descriptive levels can be prepared for the assessment:
- the text does not answer the question
- the text indicates cost components but does not explain the assumptions
- the text describes cost components and the conditions needed for their interpretation
This is an example task for the Score type. Documentation recommends describing specific situations and assessing one dimension at a time. Therefore, it is worth analyzing the completeness of responses, compliance with the offer, and the presence of examples separately.Score Documentation
The assessment obtained in this way can help organize editorial work. It should not be called a 'Google ranking result' or 'likelihood of citation by AI'.
4. Selecting materials that require checking
With a larger number of publications, it is helpful to preliminarily indicate sections that require attention.
You can design checking questions to see if the text contains:
- a specific promise of results
- a price or delivery time
- comparison with competitors
- a claim about customer results
Such analysis helps direct the material to the appropriate person. It does not replace fact-checking.
If we want to assess the article's compliance with the offer, we also need the current offer. The model should not guess whether the company actually provides the service within 24 hours.
Is Jev mistaken?
In TypeSafe communication, the phrase "zero hallucinations" appears. It is important to understand its scope: the manufacturer relates it to the compliance of results with specific types and response patterns. A correct format does not automatically mean a correct judgment.Manufacturer's explanation
The model can choose one of the allowed categories, yet still assign the case to the wrong one.
For Choice and Score, Jev returns a probability distribution and a confidence indicator. Documentation explains that this indicator results from the distribution of responses. One should not interpret a value of 0.9 as an automatic guarantee of 90% accuracy. Noul does not have a separate field for confidence. Response confidence documentation
In GEO applications, this means the need to compare results with manual evaluation of samples before using them in reports.
How to start using Jev?
TypeSafe provides a Playground where you can log in, enter data, and define questions. An API is used to connect the model to your own tool; the manufacturer also describes using the SDK.Getting started guide
The first test should be limited to one task. For example:
- Gather 50 recorded AI responses related to one industry
- Manually mark those that contain a recommendation for the specified brand
- Save the criteria that distinguish a recommendation from a regular mention
- Conduct classification using Jev
- Compare results and analyze discrepancies
- After refining the criteria, conduct another test on new examples
The number 50 is a suggestion for a small pilot, not a validation standard. Such a test allows assessing the idea's usefulness before building a larger integration.
A WordPress site owner does not need to implement Jev right away. It is primarily a tool for those building processes and software. It can work in the backend of a content analysis system, but it is not a ready solution for running a blog.
What about the promises of speed and low cost?
TypeSafe publishes comparisons indicating a significant advantage of Jev in selected automation tasks. However, the company notes that the most impressive results may correspond to the upper range of benefits achievable in practice.Description of tests and limitations
For the marketing team, a sensible test is their own process: the cost of analyzing a specific number of responses, the time taken, and the number of cases requiring correction.
Quick classification is valuable when it is sufficiently accurate for the task we want to assign it.
Frequently Asked Questions about Jev and GEO
Will Jev write a blog article?
Jev is intended for structured decisions, not for writing free text. In the editorial process, it can serve as an evaluating or classifying tool for materials.
Can you position a page in Jev?
The product described by TypeSafe is not a search engine presenting users with page rankings. A sensible application in GEO is to support analysis, not to compete for position 'in Jev'.
Does a high article rating in Jev mean that ChatGPT will quote it?
No. The rating pertains to defined criteria and provided data. It is not a forecast of decisions made by another AI system.
Are the described applications features of Semly?
The presented scenarios are proposals for using Jev. The article does not constitute an announcement or confirmation of the model's integration with Semly.
From the new model to a specific task
The most interesting application of Jev in GEO and AEO starts with a well-posed question. Can we consistently distinguish a mention from a recommendation? Identify texts that do not answer the question? Organize hundreds of customer inquiries?
If so, you can check if Jev will streamline this work. If the criteria are unclear, it is worth refining the assessment method first.
And if you want to start with your own company, check how it appears in AI responses.
Generate a free visibility report for your company in AI →
Information status: September 24, 2026. The operation of Jev is based on TypeSafe materials. Examples of applications in GEO and AEO are editorial proposals, not model test results.
Share:
