How to Qualify Backlink Prospects With TypeSafe AI's Jev

A backlink prospect is useful only when the source page has a credible editorial reason to reference your target page. TypeSafe AI's Jev can help with one part of that decision: classify the relevance of a candidate page from evidence you supply. It cannot visit the page, tell you how Google will value a link, secure a placement, or write a justified outreach email. A practical workflow starts with Screpy prospect and backlink data, filters obvious mismatches with code, and asks Jev a bounded Choice question about editorial fit. The result becomes a review queue for a human link builder. Below is a fictional page-to-page example, a real TypeSafe request structure, and an illustrative response you can adapt without treating the model's label as a ranking score.

Filter the candidate list before using Jev

Do not spend a model call on a prospect that fails a basic, verifiable requirement. Check the current HTTP response, canonical URL, page language, duplicate entries, and whether the page is actually about a related subject. Confirm that the proposed target is live and says what the campaign claims. These are crawler or application checks; Jev has no browser in this workflow.

Screpy's backlink checker exposes source and target URLs, anchor and context for existing links. Its link-building workflow can help collect prospects. Treat those records as starting points, not endorsements of every domain. A competitor backlink might come from a page that would never naturally cite your resource. A high-looking metric also does not prove editorial relevance.

Keep the list finite. For example, if the target is a guide to debugging ecommerce product-page canonical tags, filter out candidates about unrelated app marketing before semantic review. Save why each candidate passed the mechanical filters. This prevents a Jev label from hiding the fact that a URL was broken, duplicated, or misidentified.

Give Jev a bounded evidence packet

For each remaining candidate, provide a short source-page excerpt, its title and audience, the target-page summary, and the editorial context you want to assess. Fetch the source yourself and record the capture time. Do not pass only a URL: Jev evaluates supplied state, so a URL string does not give it the page contents.

Suppose the target is “How to diagnose canonical tag errors on Shopify product pages.” Candidate A is a recent technical checklist that tells store owners to audit canonical tags and cites external diagnostic guides. Candidate B is a generic coupon directory with no SEO explanation. Candidate A has a plausible editorial reason to link to the target; Candidate B does not. This contrast is clearer than asking the model to invent an abstract “domain quality” score.

A useful evidence packet also records whether the source excerpt is complete enough to decide. If the page is behind a login, returns an error, or contains only a thin summary, the correct route is insufficient_evidence or manual research. Never silently substitute a search snippet for the original page without labeling that limitation.

Classify page-to-page fit with Choice

Use a TypeSafe Choice question to pick one operational label. The following request uses fictional pages but follows the API structure: a model, state, and atomic questions with instructions and finite criteria.

{
  "model": "jev-1.13.0",
  "state": {
    "target_title": "Diagnose Shopify product-page canonical errors",
    "target_summary": "A step-by-step guide to finding and fixing canonical mismatches on product pages.",
    "candidate_title": "Technical checks before a Shopify migration",
    "candidate_excerpt": "Before launch, audit redirects, product-page canonical tags, and duplicate collection URLs. Give store owners a diagnostic reference for each issue.",
    "candidate_audience": "Shopify site owners preparing a migration"
  },
  "questions": [{
    "type": "choice",
    "instructions": "Classify the editorial fit of this candidate page for citing the specified target. Use only supplied evidence. Do not predict ranking value or write outreach copy.",
    "criteria": {
      "strong_fit": "The excerpt gives a specific, natural reason to cite the target resource for its audience.",
      "possible_fit": "The topics relate, but the cited use or audience fit is not clear enough for automatic approval.",
      "poor_fit": "The candidate has no plausible editorial reason to cite this target.",
      "insufficient_evidence": "The supplied excerpt cannot support a reliable fit decision."
    }
  }]
}

An illustrative response, not a measured result, could be:

{
  "choice": "strong_fit",
  "probabilities": {
    "strong_fit": 0.82,
    "possible_fit": 0.12,
    "poor_fit": 0.03,
    "insufficient_evidence": 0.03
  },
  "confidence": 0.61
}

The selected label is a triage suggestion. The probability distribution shows competing labels; confidence does not establish the publisher's intent or Google's assessment. The 0.61 in this invented response might trigger human review despite the strong_fit label. Re-run the same question against Candidate B's coupon-directory excerpt: a sensible workflow should reject it or mark the evidence insufficient. Include such negatives in a labeled evaluation set before processing real prospects.

Turn the verdict into a review queue

Combine Jev's label with facts already checked outside the model. A small routing rule can separate clear candidates from research tasks:

function routeProspect(pageChecks, result) {
  if (!pageChecks.live || !pageChecks.targetLive || pageChecks.duplicate) {
    return "fix_or_discard";
  }
  if (result.choice === "insufficient_evidence" || result.confidence < 0.7) {
    return "research";
  }
  return result.choice === "strong_fit" ? "editorial_review" : "hold";
}

The 0.7 threshold is an example; choose thresholds after checking labeled cases for your niche. Even editorial_review is not permission to send an email. A person should verify that the page still exists, the potential citation fits the paragraph, the contact is appropriate, and any outreach accurately describes your resource. Screpy's backlink outreach guide covers the outreach workflow; the Jev verdict here only helps prioritize review.

This distinction also matters for search policy. Google's spam policies identify links created primarily to manipulate rankings as link spam. Using AI to score prospects does not make a paid, forced, or irrelevant placement editorial. Record the reason a human approved each candidate and respect disclosure requirements where applicable.

Review three prospects as a link builder would

A practical prospect list contains more than a model score. The three fictional candidates below all passed a basic URL check, but they differ in their reason to reference a guide about Shopify product-page canonical errors.

Jev can help sort the middle column into strong_fit, possible_fit, poor_fit, or insufficient_evidence. It does not supply the missing evidence. If the marketing roundup's excerpt omits a deeper section about migrations, a reviewer needs to retrieve that section before changing the label. If the migration checklist is itself a copied or abandoned page, the editorial-fit label alone should not approve outreach.

A useful trial is to hand-label a mixed batch of prospects, including obvious negatives and borderline pages, before automating any queue. Compare Jev's labels against the human review, then examine where the model mistakes topical similarity for an actual reason to cite the target. Record the reviewer outcome and whether a placement was eventually earned as separate fields. Placement rate depends on publisher decisions, outreach quality, and timing; it cannot be read from Jev's fit probabilities.

After review, an editor can prepare a concise, truthful note explaining why the target resource helps the publisher's readers. That writing step belongs to the editor or a separate generative tool. A typed Jev response is not an outreach message and should not be pasted into one.

Avoid false precision in link prospecting

A strong_fit output is not a domain authority metric, a probability of winning a link, or a predicted ranking uplift. It is a bounded judgment about two supplied page descriptions. If those descriptions omit the actual article, misstate its audience, or become stale, the label can be wrong even when its confidence looks high.

Review a sample across accepted and rejected prospects. Keep the source excerpt, target version, model version, decision, and reviewer outcome so you can see where the classifier fails. Use a separate policy for blocked publishers, paid placements, sponsorship labels, and contact permission. Do not turn this into automatic mass outreach.

The Jev SEO hub explains the typed-decision pattern across tasks. If your problem is links between pages on your own site, use the separate internal-link opportunity guide; backlink prospecting has different evidence and editorial ownership.