Headlines Analytics Console

Headline Variant Tool

Paste a headline + primary keywords (all required in every variant). Optionally paste secondary keywords and the full article body. Get 5–10 alternative phrasings (rule-following variants score ≥85 on the production grader) plus one ⚡ wildcard for the rolling off-pattern cohort study.

… … … ?
How to use this tool click to expand
  1. Paste your headline into Seed headline. This is the version you wrote and want alternatives for.
  2. List your primary keywords—the load-bearing SEO terms. Every variant the tool offers will contain ALL of them. Comma-separate (e.g. vagus nerve, anxiety).
  3. Optional but recommended: paste the article body. With it, the tool unlocks "maximize difference" mode—variants surface different facets the article actually covers (ingredients · mechanism · audience · cost), and the grader judges accuracy against the article instead of the headline alone. Without article body, the tool stays conservative.
  4. Click Generate variants. Wait 5–15 seconds. Each card shows its grader score, similarity to your seed, and structural breakdown.
  5. Click Keep on every variant you'd ship—multi-pick is supported. Click Kept ✓ again to unkeep. Kept state is server-stored and survives regenerating the same seed, so you can come back to it. Or Copy to clipboard if you want to compare across runs.
  6. Don't love any of the options? Click ↻ Regenerate for a fresh batch with the same inputs.

The ⚡ wildcard variant is one experimental option per call—structurally off-pattern, intentionally different. Its job is to surface candidate new headline formats. Verify wildcards editorially before publishing (they bypass some of the LLM judgment criteria).

?
? All primary keywords are required in every variant.
? Including these adds bonus signal but isn't required.
? Pasting the article unlocks far more diverse, article-grounded variants.
?
? Optional—a name, a study, or a plain description all work; the tool matches it to your article and never invents one. Leave blank to let it pick.
Advanced (optional)
?
?
?
?
?
?
Reference: what the tool guarantees

Every rule-following variant scores ≥85 on the production headline grader. ("Rule-following variants" are the standard options the tool returns: the ones that score 85+ on the grader and pass every hard rule. The single ⚡ wildcard is the exception, covered below.) Hard structural rules (named-entity lead, formula, primary keywords, banned patterns) are non-negotiable; candidates failing those drop silently. Length is enforced with one editorial tolerance: an option a few characters over the 90–104 target (up to ~115) soft-passes carrying an amber ⚑ editorial-judgment flag instead of dropping, so you get a fuller set and decide whether to trim; only options past ~115 or under the 75-char floor drop. The 85-point floor tolerates one stylistic LLM miss (e.g., a borderline active-voice or curiosity call) but still rejects substantive issues like lead burial or accuracy. Each variant card shows its actual score.

Every variant contains every primary keyword you supplied. Per the content team lead's editorial rule (2026-05-06), the load-bearing keywords are hard-required in every option. Secondary keywords add bonus signal but aren't gated.

Article body unlocks the richest variants. When you paste the article text, the tool switches to "maximize difference" mode (you'll see a blue article-mode pill in the results header). Pairwise diversity gates kick in between rule-following variants, and the LLM grader judges accuracy against the article itself—variants may name specific techniques, mechanisms, findings, sources, or comparisons present in the body even if absent from the headline. Without article body the tool stays conservative: anything beyond the headline's stated scope gets flagged.

Wildcard variant (1 per call): tagged ⚡ wildcard as an experimental off-pattern option. Wildcards skip the rule-following LLM grader (active voice / lead burial / curiosity) since those criteria fight the structural-novelty objective; they still pass every hard structural rule plus a single accuracy safety-net check. Verify wildcards editorially before publishing. Keeping wildcards over time feeds the long-term cohort study that surfaces candidate new headline standard formats.

Active gates (article-body mode values in parens when different):

  • Cosine < 0.97 (0.93) vs your seed headline—catches near-literal duplicates; tighter in article-mode where the article gives variants room to legitimately diverge.
  • Pairwise cosine < 0.85, pairwise word-overlap < 0.45 between rule-following variants (article-mode only)—forces facet diversity across the set, so the variants use different wording.
  • Wildcard cosine < 0.95 vs seed—wildcards have their own ceiling; cohort study values structural fingerprint, not semantic distance.
  • Formula cap (max 2 per formula)—no formula hogs the option set across the credit-worthy formulas in the canonical taxonomy (formula_taxonomy.py, shared with the Findings classifier and the grader); lowest-scoring excess get dropped. Worker-side pattern list is synced manually against the canonical taxonomy.
  • Char band 90–104 target, +11 over-tolerance—options up to ~115 chars soft-pass with an amber ⚑ editorial-judgment flag (variant-tool-only; the daily Grader hard-fails over target). Below 75 or above ~115 still drops.
  • Grader ≥85 on every rule-following variant; the other structural rules (named-entity lead, formula, keywords, banned patterns) pass at 100% on every variant including wildcards.

Optional advanced inputs (under the "Advanced" disclosure above the Generate button):

  • Also avoid—paste any previously-selected headlines from the same article cluster (one per line). The tool will gate new variants against each of these by the same cosine ceilings, so sibling bridges and deepeners don't collide. Useful when you've already published a deepener + 2 bridges and need a third bridge that's distinct from all three. Skip if you don't have prior picks to avoid.
  • Format / Persona / Platform / Outlet—narrow rule application by editorial format (Everything to Know, Recipe, etc.), persona, distribution platform (Apple News, SmartNews), or outlet (US Weekly, Woman's World, Life & Style). Default is general H1 (80–100 chars, no platform constraints).

What gets stored. Every offered candidate (rule-following + wildcard, picked or not, with the gate that dropped it if dropped) is persisted to feed the long-term cohort study. Kept variants additionally land with their full grader breakdown + structural fingerprint, and unkeeping soft-deletes (preserves the decision history). Generation audit (cost, rejection counts, regen passes, calibration version, model assignment, prompt hash) is logged for reproducibility.