How to Classify Cited URLs in Your Industry with Jev

Classify cited URLs with Jev only after you define the taxonomy. Write one label list and set of decision rules for each independent question, test those rules on a human-labelled sample, and then encode them as Jev choice, score or noul questions. Jev returns typed labels, option probabilities and confidence, which makes it suitable for high-volume classification with an explicit review queue.

The useful unit is a page record with enough evidence to classify it: URL, domain, title, H1, page text or excerpt, structured-data type, publication details, prompt segment and supported claim. A URL string alone can identify some domains, but it cannot reliably distinguish an editorial comparison from a vendor’s own comparison page.

Why should the taxonomy come before Jev?

Jev chooses from the options you define. It cannot repair overlapping labels or decide which business question you meant to ask. If editorial, listicle, competitor and review platform sit in one list, the classes mix ownership, publisher model and page format. A competitor’s listicle could fit three options.

Use separate axes:

1. Ownership: who controls the property?
2. Publisher type: what kind of organization or surface published it?
3. Page format: what kind of page is it?
4. Primary intent: what job does the page perform for its reader?
5. Actionability: what can the brand realistically do about this citation?
6. Industry role: which vertical-specific role does the page play?

Each axis answers one question and returns one label. Multi-label output across axes comes from asking several typed questions in the same Jev request, not from forcing every combination into one giant list.

The first five axes transfer across industries. The sixth is where the taxonomy becomes useful for a specific market. Tourism needs to separate official properties, online travel agencies, review platforms, destination marketing organizations, transport operators and government advisories. A generic company label loses those differences.

What evidence should each page record contain?

Send the smallest state that supports the decisions. Jev charges by input tokens, and extra boilerplate can introduce noise.

{
  "observed_url": "https://example.com/best-payroll-tools",
  "canonical_url": "https://example.com/best-payroll-tools",
  "domain": "example.com",
  "title": "Best Payroll Tools for Distributed Teams",
  "h1": "The 12 Best Payroll Platforms for 2026",
  "excerpt": "We independently reviewed...",
  "schema_types": ["Article", "ItemList"],
  "author": "Jane Smith",
  "organization": "Example Media",
  "prompt_segment": "commercial_comparison",
  "supported_claim": "Product A supports contractor payments in 120 countries",
  "known_brand_domains": ["client.com"],
  "known_competitor_domains": ["competitor.com"]
}

Keep known brand and competitor domains outside probabilistic classification where possible. Exact domain rules are cheaper and more reliable for ownership. Use Jev for the ambiguous remainder and for content-dependent axes.

Which ownership labels should you define?

Ownership determines whether the page is a first-party asset, competitor evidence or an independent source.

Apply an exact domain map first. If a parent company owns several product domains, document the relationship in the map. Do not infer ownership from whether the brand appears in the article.

Which publisher-type labels should you define?

Publisher type describes the surface, independent of the page’s format.

A company blog is company even when the article has an editorial tone. A review article on a newspaper site remains editorial_media; its page format can be comparison_roundup.

Which page-format labels should you define?

Page format captures the document’s structure and expected evidence type.

Choose the primary format. Store secondary traits such as has_table, has_original_data and has_step_list as separate yes/no fields. Otherwise, the label list expands until the classes become impossible to apply consistently.

Which intent labels should you define?

Intent describes the page’s primary reader job, not the prompt that produced the citation.

Mixed intent should be rare. Define it as a genuine tie, not a fallback for uncertainty; unclear handles insufficient evidence.

Which actionability labels should you define?

Actionability converts classification into a work queue. It depends on the focal brand, so supply brand context in the state.

Actionability is a recommendation class, not an objective property of the URL. A government page may be reference_alignment for one project and monitor_only for another.

Which tourism-specific source roles should you define?

Add tourism_source_role when the cited-URL set covers hotels, destinations or travel planning. Classify the page’s role in the traveller’s decision, not the legal form of its owner.

Domain rules can suggest a role, but they cannot finish the classification. Tripadvisor hosts review pages and traveller forums. A page-level classifier should label the destination page review_platform and a forum topic traveller_community even though both URLs share tripadvisor.com.

How does the taxonomy work on a tourism validation set?

The table below uses public Istanbul and travel pages extracted on 11 October 2026. It tests the taxonomy; it is not a claim that an answer engine cited these pages. A real benchmark must start from URLs observed in saved LLM answers.

Two rows should trigger manual review during taxonomy development. GoTürkiye’s ownership and operating structure must come from a maintained organization map rather than page tone. Lonely Planet may be reachable for editorial work in one campaign and unrealistic in another, so actionability needs client and market context.

Tavily could extract seven of nine selected pages. The St. Regis Istanbul and Turkish Airlines Touristanbul pages failed in the basic extraction attempt. Keep these failures in the dataset, label their content evidence as missing and route them to refetch or manual review. Dropping blocked pages would make the classifier test easier while hiding a common production condition.

The tourism role can be encoded as another choice question:

{
  "tourism_source_role": {
    "type": "choice",
    "instructions": "Classify the role this page plays in a traveller's decision. Classify the page, not the domain. Use other_or_unclear when content evidence is missing.",
    "criteria": {
      "official_property": "First-party page for one hotel, resort, attraction, or tour product",
      "ota_booking_platform": "Intermediary comparing or booking inventory from multiple suppliers",
      "review_platform": "Page where ratings and traveller reviews are central evidence",
      "destination_marketing_organization": "Official or delegated body promoting a destination",
      "travel_editorial": "Independent destination advice or travel journalism",
      "traveller_community": "User-led travel discussion, question, or experience thread",
      "government_advisory": "Public safety, entry, health, visa, or regulatory guidance",
      "transport_operator": "Airline, rail, ferry, airport, or local transport service page",
      "research_reference": "Statistical, academic, library, or standards evidence",
      "social_video_creator": "Social or video creator page used as the source",
      "other_or_unclear": "No defensible role from the supplied evidence"
    }
  }
}

Which tourism prompts should seed the real cited-URL set?

Use one destination and one accommodation segment for the pilot. An Istanbul hotel set can cover the main source roles without mixing unrelated countries.

Freeze the traveller origin, dates, currency and party type where they affect the answer. Run the same prompt set across the selected engines and at least three repetitions before treating the resulting URL set as the Jev input corpus.

Which auxiliary gates should you add?

Use Jev noul questions for calibrated yes/no gates and score for ordered assessments.

• needs_human_review: conflicting evidence, inaccessible content, ambiguous ownership or policy risk.
• brand_present: the focal brand appears in the extracted page content.
• competitor_present: a defined competitor appears.
• supports_claim: the page supports the answer claim attached to the citation.
• outreach_safety: ordered scale from prohibited/unsafe to normal editorial contact.
• evidence_quality: ordered scale based on sourcing, method and specificity.

Do not use a confidence score as proof that the label is correct. Confidence shows how concentrated the model’s option distribution is. Validate thresholds against your own labelled sample.

How do you encode the taxonomy as Jev questions?

Jev accepts a state plus a map of typed questions. A choice question can define up to 255 options and returns one option with probabilities and confidence. Multiple questions run in one request and share the state.

{
  "model": "jev-1.13.0",
  "state": {
    "url": "https://example.com/best-payroll-tools",
    "domain": "example.com",
    "title": "Best Payroll Tools for Distributed Teams",
    "h1": "The 12 Best Payroll Platforms for 2026",
    "excerpt": "We independently reviewed each platform using...",
    "prompt_segment": "commercial_comparison",
    "supported_claim": "Product A supports contractor payments in 120 countries"
  },
  "questions": {
    "publisher_type": {
      "type": "choice",
      "instructions": "Classify who publishes this page. Use other_or_unclear when the evidence is insufficient.",
      "criteria": {
        "company": "Commercial organization publishing on its own property",
        "editorial_media": "News, magazine, or trade publication with an editorial model",
        "community_forum": "User-created discussion or answer community",
        "review_marketplace": "Listings, reviews, or product-comparison platform",
        "reference": "Encyclopedic or general reference source",
        "academic_research": "University, journal, conference, or research repository",
        "government_public_body": "Government, regulator, or public institution",
        "social_video": "Social network or video publishing surface",
        "other_or_unclear": "No defensible match from the evidence"
      }
    },
    "page_format": {
      "type": "choice",
      "instructions": "Choose the page's primary document format, not its publisher type.",
      "criteria": {
        "product_or_service": "One offering, its features, pricing, or value",
        "category_or_listing": "A list or directory without substantial evaluation",
        "comparison_roundup": "Two or more options compared using criteria",
        "review": "Evaluation of one named product, service, or organization",
        "how_to": "Ordered instructions for completing a task",
        "explainer_definition": "Definition or conceptual explanation",
        "research_data": "Data or findings with a stated method",
        "news_announcement": "Time-bound event, release, or announcement",
        "documentation_support": "Official documentation, help, or troubleshooting",
        "discussion_thread": "User-led posts, replies, or answers",
        "profile_entity": "Profile for one entity, product, person, or location",
        "other_or_unclear": "No defensible primary format"
      }
    },
    "intent": {
      "type": "choice",
      "instructions": "Classify the page's primary reader job.",
      "criteria": {
        "informational": "Learn a concept, fact, or process",
        "commercial_investigation": "Compare options before a decision",
        "transactional": "Complete a purchase, signup, download, or booking",
        "navigational": "Reach a known entity or resource",
        "mixed": "Two intents are equally central",
        "unclear": "Insufficient evidence"
      }
    },
    "tourism_source_role": {
      "type": "choice",
      "instructions": "Classify the role this page plays in a traveller's decision. Classify the page, not the domain.",
      "criteria": {
        "official_property": "First-party hotel, resort, attraction, or tour-product page",
        "ota_booking_platform": "Multi-supplier search, comparison, or booking intermediary",
        "review_platform": "Ratings and traveller reviews are central evidence",
        "destination_marketing_organization": "Official or delegated destination promotion body",
        "travel_editorial": "Independent destination advice or travel journalism",
        "traveller_community": "User-led travel discussion or experience thread",
        "government_advisory": "Public safety, entry, health, visa, or regulatory guidance",
        "transport_operator": "Airline, rail, ferry, airport, or local transport operator",
        "research_reference": "Statistical, academic, library, or standards source",
        "social_video_creator": "Social or video creator page",
        "other_or_unclear": "Insufficient evidence or no defined role"
      }
    },
    "needs_human_review": {
      "type": "noul",
      "instructions": "Does ambiguity, missing content, conflicting evidence, or policy risk require human review?"
    }
  }
}

Pin the Jev version after you tune thresholds. A moving jev-latest target can change label distributions even when your inputs and taxonomy remain fixed.

Should every field be classified by Jev?

Use deterministic rules where the answer is already known.

This hybrid design gives you an audit trail. A reviewer can see whether a label came from a domain rule, Jev decision or manual override.

How do you build the human-labelled test set?

Sample pages before running the full inventory. Include common classes, rare classes and hard boundary cases.

1. Draw a stratified sample across engines, prompt segments and high-frequency domains.
2. Have two people label at least the ambiguous subset independently.
3. Record disagreements and revise definitions before measuring Jev.
4. Freeze a held-out test set that will not be used to tune instructions or thresholds.
5. Keep unclear and manual_review examples in the test set; deleting them creates a misleadingly easy benchmark.

Start with enough examples to expose definition failures. A small pilot of 100–200 pages can reveal overlapping labels, but it does not establish production accuracy for rare classes. Expand the sample until every class that drives an action has a meaningful test count.

Which metrics should you report?

Overall accuracy hides failure on small but valuable classes. Report:

• confusion matrix for every axis;
• per-class precision, recall and F1;
• macro-F1 so large classes do not dominate the score;
• coverage at each confidence threshold;
• error rate after human-review routing;
• manual-review volume and minutes per 100 pages;
• label distribution by engine and prompt segment;
• cost and latency per page.

Set thresholds by consequence. A false editorial_outreach label wastes analyst time. A false monitor_only label can bury a valuable opportunity. Those classes may need different automation thresholds.

How should confidence-gated routing work?

Use a three-lane queue after validation:

1. Auto-accept: label and all rule checks meet a tested threshold.
2. Human review: confidence falls below the threshold, two axes conflict, or needs_human_review crosses its gate.
3. Reject and refetch: page content is empty, blocked, stale or mismatched to the URL.

Store the full probability distribution, resolved model version, taxonomy version and final reviewer decision. These fields let you recalibrate thresholds and find drift later.

How do the labels become SEO and GEO work?

Join the classified page table back to prompts and citation events.

• owned + how_to + informational shows which first-party procedures already earn citations.
• competitor_owned + explainer_definition reveals first-party content gaps.
• third_party + editorial_media + comparison_roundup forms an evidence-led outreach pool.
• third_party + review_marketplace + commercial_investigation points to profile and review operations.
• third_party + community_forum + discussion_thread shows where real category conversations shape answers.
• government_public_body + research_data identifies standards and data the brand should align with.

Calculate the distribution by engine and prompt segment. A source mix averaged across every prompt can hide that commercial questions depend on review platforms while technical questions depend on documentation.

What should you publish with the results?

Publish the taxonomy version, definitions, sample construction, model version, thresholds, confusion matrices and manual-review policy. Keep client-sensitive URLs private if necessary, but expose enough method for another practitioner to evaluate the claims.

Do not publish a source-type percentage without naming the classification unit. State whether the denominator is citation events, unique pages, unique domains, answers or prompts. Re-run a labelled audit when the taxonomy, extractor or pinned Jev version changes.