TypeSafe AI’s Jev can classify keyword search intent without generating an open-ended explanation. You define the possible answers, send the keyword and its context as state, and receive a typed decision with probabilities and a confidence value. That makes Jev useful for high-volume SEO workflows where consistent labels matter more than prose.
This guide shows how to model keyword intent as a Jev Choice question, interpret the result correctly, route uncertain keywords to human review, and avoid confusing model confidence with SEO truth.
What Is TypeSafe AI’s Jev?
Jev is TypeSafe AI’s first public System One Model. TypeSafe introduced it in September 2026 as a model built for fast, structured decisions inside software.
A conventional generative model returns text. Jev evaluates typed questions against the state you provide and returns structured values that application code can use directly. It does not write an article, produce a free-form rationale, or replace an SEO specialist.
TypeSafe currently documents three question types:
| Question type | Best used for | Returned value |
|---|---|---|
| Choice | Select one option from a fixed set | Choice, probabilities, confidence |
| Score | Place something on an ordered scale | Score, probabilities, confidence |
| Noul | Evaluate a yes-or-no statement | Probability from 0 to 1 |
Keyword intent is a natural Choice task because the answer must come from a controlled set of labels. See TypeSafe’s Jev introduction for the current model concepts and API documentation.
Why Use Jev for Keyword Search Intent Classification?
A keyword export can contain hundreds or thousands of queries. Reviewing every row manually is slow, while asking a text-generating model for an explanation can produce inconsistent labels and output formats.
Jev gives the workflow a fixed contract. For every keyword, the application can expect one of the allowed intent labels plus a probability distribution and confidence value. Those results can be filtered, counted, routed, and reviewed without parsing paragraphs.
This is useful for:
- separating informational queries from purchase-oriented terms;
- identifying ambiguous keywords that need live SERP review;
- grouping keywords before content planning;
- mapping clusters to guides, comparison pages, product pages, or existing URLs;
- keeping classification output consistent across a large export.
Jev does not determine what Google will rank. Search results can vary by country, language, device, freshness, and query context. Treat the model’s classification as a structured first pass, then validate important or uncertain terms against the live SERP.
If you need a broader introduction before automating the workflow, read Screpy’s guides to user search intent and matching content with search intent.
Define the Intent Options Before Calling Jev
Use labels that lead to clear SEO actions. A practical starting set is:
| Intent | Searcher’s likely goal | Typical page type |
|---|---|---|
| Informational | Learn, understand, or solve a problem | Guide, tutorial, definition |
| Navigational | Reach a known brand, product, login, or page | Homepage, product page, help page |
| Commercial investigation | Compare options before deciding | Comparison, alternatives, category page |
| Transactional | Complete an action now | Product, pricing, signup, booking page |
| Ambiguous or mixed | More than one goal is plausible | Manual SERP review |
The fallback option matters. A broad keyword such as “website audit” can refer to a guide, a tool, or a professional service. Forcing every uncertain query into one of four labels creates false precision.
Intent is also different from funnel stage. A keyword can be informational and still influence a purchase. Keep “What does the searcher want now?” separate from “How close is this person to conversion?”
Build a Jev Choice Question for Search Intent
TypeSafe’s documented Choice structure contains:
state: the keyword and relevant context;model: the Jev model requested;questions: one or more typed questions;instructions: the decision Jev should make;criteria: the allowed option keys and their definitions.
A search-intent request can follow this structure:
{
"state": {
"keyword": "best SEO audit tools",
"country": "United States",
"language": "English",
"brand_terms": ["Screpy"]
},
"model": "jev-1.13.0",
"questions": {
"primary_intent": {
"type": "choice",
"instructions": "Choose the primary search intent expressed by this keyword.",
"criteria": {
"informational": "The searcher primarily wants to learn or solve a problem.",
"navigational": "The searcher wants to reach a known brand, product, or page.",
"commercial_investigation": "The searcher is comparing or evaluating options before acting.",
"transactional": "The searcher wants to buy, subscribe, book, download, or start now.",
"ambiguous_or_mixed": "The wording supports multiple intents or lacks enough evidence."
}
}
}
}
The model name above matches the version shown in TypeSafe’s documentation when this guide was reviewed. Check the current Choice documentation before implementing the request.
Criteria should describe the meaning of each option rather than rely on keyword modifiers alone. “Best” often signals commercial investigation, but the full query and available context should drive the classification.
Understand the Jev Choice Response
A Choice answer contains the selected option, the probability assigned to every option, and a confidence value. An illustrative response shape looks like this:
{
"answers": {
"primary_intent": {
"type": "choice",
"choice": "commercial_investigation",
"confidence": 0.82,
"probabilities": {
"informational": 0.06,
"navigational": 0.01,
"commercial_investigation": 0.85,
"transactional": 0.05,
"ambiguous_or_mixed": 0.03
}
}
}
}
The numbers above illustrate the response format; actual probabilities depend on the state, criteria, and model response.
The choice is the option with the highest probability. The probabilities object shows how the model distributed probability across all allowed answers. The confidence value summarizes how concentrated that distribution is.
Jev’s confidence is returned from 0 to 1. If a spreadsheet displays 82%, the application has converted 0.82 into a percentage. Jev does not natively return an 82-out-of-100 SEO score.
Set Confidence Thresholds With Your Own Data
Confidence does not prove that a label is correct. It indicates how strongly the returned probability distribution favors one outcome.
TypeSafe recommends choosing thresholds according to the risk and performance of the specific workflow. There is no universal confidence threshold for keyword intent.
Start by creating a reviewed benchmark:
- Select a representative set of branded, generic, long-tail, local, comparison, and purchase-oriented keywords.
- Have an SEO reviewer assign the expected primary intent.
- Run the same Jev Choice question across the sample.
- Compare labels and confidence values with the reviewed decisions.
- Adjust criteria and routing thresholds before processing the full export.
A conservative initial routing policy might be:
confidence < 0.60 -> manual review
confidence 0.60–0.84 -> accept for low-risk grouping, review before page creation
confidence >= 0.85 -> accept unless the keyword is strategically important
ambiguous_or_mixed -> always inspect the live SERP
These values are workflow examples, not Jev or SEO benchmarks. Replace them with thresholds supported by your own validation data. TypeSafe’s confidence documentation explains why the appropriate cutoff depends on the action your software will take.
Add Page Type as a Separate Typed Decision
Jev does not produce a free-form recommendation such as “Create a comparison page because the word best implies evaluation.” If you need a recommended page type, model it as a second Choice question.
For example:
"recommended_page_type": {
"type": "choice",
"instructions": "Which page type is the best candidate for this keyword?",
"criteria": {
"guide": "An educational article, tutorial, or definition.",
"comparison": "A page comparing products, services, or alternatives.",
"category": "A collection or category of multiple relevant options.",
"product": "A page presenting one product or feature.",
"pricing_or_signup": "A page designed to start a purchase or signup.",
"existing_navigation": "An existing brand, login, support, or destination page.",
"serp_review_required": "The keyword alone does not support a reliable choice."
}
}
TypeSafe says questions in the same call are evaluated independently against the shared state. Keeping intent and page type separate prevents one vague question from combining several judgments.
Your application should create the remaining workflow fields:
- preserve the original keyword and row ID;
- store Jev’s intent and page-type answers;
- convert confidence to a display percentage if desired;
- calculate review status from your thresholds;
- attach a human note after SERP review.
If you want a written rationale, generate it separately or let the reviewer add it. Do not present free-form prose as a native Jev output.
Validate Jev’s Label Against the Live SERP
Keyword wording is only one source of evidence. Before creating or substantially changing a page, review the current results in the target market.
Check:
- the dominant organic page type;
- whether results favor guides, comparisons, products, tools, or local services;
- AI Overviews, featured snippets, shopping results, local packs, and videos;
- whether the query has strong brand or navigational signals;
- whether different intents share the first page.
If Jev selects transactional but the SERP is dominated by comparisons, inspect the conflict rather than automatically choosing either side. The wording may suggest immediate action while Google’s current results reflect evaluation. Record the final decision so repeated patterns can improve the criteria.
This step is especially important for broad terms, “free” modifiers, brand-plus-pricing queries, local services, and emerging product categories.
Turn Classified Keywords Into an SEO Workflow
Once the labels are validated, use them to decide what happens next.
- Map navigational terms to existing destinations.
- Group compatible informational terms into one complete guide.
- Route commercial clusters to comparison, alternatives, use-case, or category pages.
- Route transactional terms to pages with a clear action.
- Flag mixed-intent terms for SERP review.
- Check whether a relevant URL already exists before creating new content.
Do not create one page for every keyword variation. Screpy’s guide to finding long-tail keywords explains how to qualify terms by search intent and business fit, while AI-assisted keyword research provides the broader research context.
Google’s guidance on generative AI content focuses on usefulness and originality rather than whether AI participated in the workflow. Jev can organize decisions, but it cannot make thin or duplicated pages valuable.
Where Jev Helps—and Where It Does Not
Jev is well suited to bounded decisions with predefined answers. Keyword classification fits that pattern when the criteria are clear.
It is not the right tool for every SEO task:
- It does not write articles, title tags, or meta descriptions.
- It does not return a free-form explanation.
- It does not fetch or interpret a live Google SERP unless your application supplies that evidence as state.
- It does not know your commercial priorities unless you include relevant context.
- It does not guarantee that Google agrees with the classification.
- It does not remove the need for reviewed examples and threshold calibration.
The useful division of work is simple: Jev makes a typed first-pass decision, code routes the result, and an SEO reviewer handles ambiguous or high-impact cases.
Frequently Asked Questions
Is this the Jev model from TypeSafe AI?
Yes. This guide covers Jev, TypeSafe AI’s first public System One Model for typed decisions inside software. It is unrelated to other uses of the abbreviation JEV.
Can Jev write SEO content?
No. Jev returns typed decisions rather than generated prose. Use it for classification, scoring, filtering, or routing. Use a writing model or a human editor when the task requires text.
Can Jev classify keyword search intent?
Yes, search intent can be modeled as a Choice question with predefined criteria. The result should still be validated against reviewed examples and live search results.
Does Jev replace keyword research?
No. It can classify a prepared keyword set, but it does not replace query discovery, volume and difficulty analysis, competitor research, SERP review, or business prioritization.
Should low-confidence keywords be discarded?
No. Low confidence often identifies ambiguous wording or missing context. Route those keywords to review instead of deleting them.
A Practical Jev SEO Decision Pipeline
A reliable implementation follows this order:
- Collect and clean the keyword set.
- Preserve market, language, brand, and current-page context.
- Ask Jev separate Choice questions for intent and page type.
- Store the selected options, probabilities, and 0-to-1 confidence.
- Route uncertain and high-value terms to human review.
- Validate proposed page types against the live SERP.
- Map approved clusters to existing or planned URLs.
- Measure results and recalibrate the workflow with reviewed data.
This approach uses Jev for what it is designed to do: produce fast, structured decisions that software can act on. It keeps the final SEO judgment grounded in live search evidence, site strategy, and human review.