What actually makes a technical headline work

There is no universal headline formula. The rules invert across platforms — a question mark costs 19.8 percentile points on Dev.to and gains on Medium. This tool scores your draft against a model fitted on real engagement data, and tells you what it is reacting to.

— posts analysed
— platforms
— best rank correlation
— last updated

Your headline

Score

Enter a title to score it.

Attention axes

—

Axis values are measured properties of your text. Weights are platform-specific and fitted on the data — green means the model rewards it here, red means it penalises it.

Diagnostics

—

Generated candidates

Enter a topic and press Score & generate.

Subtitle suggestions

—

A subtitle has one job: expand the promise without repeating the title. Target 10–18 words and under 25% word overlap.

Recommended tags

—

Ranked by mean engagement percentile among tagged posts. Aim for one broad, one mid and one specific — never four specific.

What this platform rewards

Prompt for an LLM

Enter a topic to generate a prompt loaded with this platform's measured rules.

Four rules that held on every platform

  • One concrete number with a unit. "Cut p99 from 900 ms to 40 ms" — not "dramatically faster", and not a listicle count.
  • Evidence of first-hand work. "we measured", "I built", "in production", "postmortem", "after three years".
  • Ornamentation is a tax. ALL-CAPS, stacked acronyms and adjective stacking correlate negatively everywhere.
  • Never restate the title in the subtitle. Overlap above 25% loses the one chance you get to expand the promise.

Everything else — length, questions, listicles, negativity, colons — flips sign depending on the audience. Any tool that applies one global formula across platforms is, on this evidence, wrong about half the time.

Method & limits

Every post is converted to a within-platform engagement percentile, because points, claps, reactions, likes and upvotes are not comparable numbers. Each platform has an explicit engagement floor so the corpus contains genuine successes, with the full range above that floor kept intact so the model still has variance to learn from.

What it cannot do. The best per-platform model explains roughly 13% of the variance in relative engagement; most explain 3–7%. Headline text is a small lever. This is a good tool for choosing between your own drafts and a bad one for predicting reach — the body, the timing and the author's existing audience are much larger factors.

Two further caveats. Only high-performing posts were collected, so every effect size here is a lower bound. And five of six platforms do not expose view counts, so "engagement" means votes, claps, reactions and likes rather than attention itself.

Full write-up: methodology · platform comparison · the model · frameworks · improving it with more data.

Who made this

Built by Kailash Rathod (itsarises.com) and published alongside Connecting Dots, an evidence-based resource for SEO, AEO and GEO. The design system is shared with Connecting Dots.

Code is MIT licensed. The research — the fitted weights, the statistics and everything in docs/ — is CC BY 4.0, so you may reuse it commercially provided you credit the source.