The first quarterly read of HooklyAI’s scoring corpus: 177 scored hooks, 61 scored posts, and 65 accounts — every number traceable to the aggregate, every small sample withheld instead of published. No extrapolation, no LinkedIn-wide pretense, nothing borrowed from other studies.
Generated by one pipeline end to end — SQL over the production corpus, JSON committed to the repository, page rendered from the JSON. To update the study we re-run the query, not the marketing.
The strongest style in the corpus with n=33 — a full 9.0 points above story hooks, the weakest published style.
70.2 vs. 61.2, both at n=33. A claim lands faster than an anecdote in this corpus — the reading below explains why.
39 of 177 hooks reach the two best bands — earned, not gifted: the rubric is calibrated so the average post sits at 45–55.
The two lowest bands are empty. Weak openings die in the editor — hard gates and scored variation make “ship it anyway” the rare failure mode.
Short, explicit, and dated — the terms every number on this page obeys.
The corpus is HooklyAI’s scoring store: every generated hook variation, analyzed post, and repurposed draft receives a calibrated 1–100 score at creation time, and only the metadata survives — score, style label, rough length, source tool, and timestamp. Hook text is never stored and neither is any user identity, which is why the aggregate can be published at all: there is nothing here to leak. This snapshot aggregates 177 hooks and 61 score samples from 65 accounts, produced on October 8, 2026.
Scores come from the published rubric — seven equally weighted dimensions including human-voice AI-detection resistance, with hard gates capping external links and engagement bait at 60. Because the rubric is calibrated to be honest rather than flattering (the target average for a typical post is 45–55), a mean of 66.3 means something: it describes content that was generated, scored, and in many cases revised upward before it was kept.
One caveat belongs in the same breath as every headline number: this is a corpus of tool usage, not a random sample of LinkedIn. People who run their drafts through a scorer are selected for caring about scoring, and generated drafts start from formulas the generator already believes in. The analyzer rows — real posts users chose to score — are the closest thing to platform-normal content here, and they average 61.6, lower than the generated hooks, which is exactly the direction selection bias would predict. Read everything below as “what scores well,” never “what is typical.”
Publish floor: no statistic appears on this page with fewer than 10 observations behind it. Styles and cuts below the floor are named, counted, and withheld — see the last section. The study re-aggregates quarterly; the next snapshot is scheduled for January 2027.
Score bands across all 177 scored hooks. Band edges are the rubric's published calibration, not a bucket of convenience.
The distribution is top-heavy and bottom-empty. The modal band is 60–74 (81 hooks), 22.0% of the corpus reaches 75 or better, and the two bands below 45 hold zero observations — not one scored hook landed there. The mean is 66.3 against a median of 62.0: the small upper tail (26 hooks at 90+) is what pulls the average above the middle.
The empty bottom is a feature of the workflow, not a miracle of the writing. Scored generation returns multiple variations at once, hard gates force revisions on violations before anything ships, and nobody saves a 38 to their swipe file. The honest reading: in a loop that shows you the number, bad openings get edited rather than published.
Mean score by style, published only where n ≥ 10. Medians shown because averages lie at small tails.
Bold-statement hooks are the corpus’s strongest published style at a mean of 70.2 — 9.0 points clear of story hooks (61.2), the weakest published style, with both styles holding 33 observations. A bold claim compresses a complete argument into the fold; a story spends its first lines establishing setting before the promise arrives, and the fold does not wait.
The medians tell the subtler story. Bold statement’s median is 62.0 — statistically indistinguishable from the medians of mixed and question hooks — which means its higher mean comes from a heavier upper tail, not a higher floor. Done well, a bold claim is the most likely style in this corpus to reach for a 90+; done lazily, it scores like anything else. The style is a high-variance bet that skilled writers win.
Mixed-style hooks — openers that blend a claim with a question or a concrete detail — are the corpus’s workhorse at 45.8% of all scored hooks (81 of 177), averaging 68.4. That combination of volume and above-average score is itself a finding: users converge on blends, and the blends hold up.
A gap with an obvious mechanism and a small-sample asterisk the size of a house.
Mean hook score · n=177
Mean full-post score · n=16 — small sample
Generated hooks average 66.3; full posts average 61.6 across 16 observations — enough to publish, few enough that the margin of error is wide. The posts’ median (64.0) actually sits above their mean, and their distribution is narrower than the hooks’. Nothing here is conclusive; it is a pattern worth naming so the next quarter can confirm or kill it.
The mechanism, if it holds, is mechanical rather than mysterious: hooks are judged inside the 210-character fold, where every dimension is concentrated, while full posts must survive two thousand additional characters of formatting, CTA, and algorithm exposure — more surface area for a dimension to sag. The practical implication costs nothing to adopt: the standards your hook must meet are the ones the first two lines of your post must meet too.
61 score samples by source tool — the denominator behind every claim above.
Hook generator — scored variations, saved by users (45), post analyzer — real posts users chose to score (12), and blog repurposing before/after pairs (1 + 3). The hook-heavy mix mirrors how the product is used — hooks-first is the thesis, and the corpus reflects it — which is also why style-level claims are made about hooks and not about full posts. As the analyzer side grows, this page grows sections for it.
The withheld list is part of the dataset. If a number is not on this page, the sample refused to support it.
statistic (n=7), contrarian (n=3), bold claim (n=2), historic reference (n=1). Their raw means exist and range from 55.0 to 57.5, but at these sizes a single additional hook moves the average by points. Publishing them would manufacture conclusions out of noise; they graduate to the charts the quarter their samples clear the floor.
This was the planned headline insight: do high-scoring hooks run longer or shorter? It is withheld because exactly 1 hook in this snapshot reached 70+ with a recorded length. The mid-band average of 87 characters ( n=44) is publishable as a descriptive fact, but a one-observation comparison is not. The 2027 snapshot answers it properly.
The corpus stores per-dimension scores where the model returns them, but the industry and tone labels are far too sparse at this corpus size for any cut to clear the floor. As coverage thickens, quarterly editions will add dimension-movement analysis (which dimension improves between draft and final) and — if volume justifies it — industry-level style benchmarks. Nothing on this page borrows another study’s numbers to fill the gaps.
Scope, sourcing, and the citation format — answered the way the methodology would answer them.
No — and we will not pretend otherwise. Every number here comes from HooklyAI’s own scoring corpus: hooks and posts that ran through our generator, analyzer, and repurposing tools between beta and October 8, 2026. It is a census of content our users chose to score, not a random sample of all LinkedIn posts. That makes it ideal for answering "what works in content that scores well" and wrong for claiming anything about the average post on the platform.
Every scored item is graded 1–100 across seven equally weighted published dimensions — hook strength, formatting and readability, emotional resonance, call-to-action effectiveness, shareability, LinkedIn algorithm optimization, and human voice (AI-detection resistance). Two hard gates cap the score at 60 regardless of writing quality: external links in the post body and engagement bait. The full rubric, bands, and calibration targets are published on the scoring methodology page.
Because their sample sizes are below our publish floor of ten observations. Four styles in this snapshot fall under it: bold claim (n=2), contrarian (n=3), historic reference (n=1), and statistic (n=7). Their means exist and fluctuate wildly at that size — publishing them would be decoration, not data. They appear the moment the corpus supports them, on a quarterly re-aggregation.
One observation is one scored item: a generated hook variation, an analyzed post, or a repurposed draft. Only the score, style label, and rough length are stored — never the text itself and never your identity, which is what makes aggregate publication possible at all. If you scored a hook or analyzed a post with HooklyAI, your number may be inside an average here; your words are not.
Quarterly. The numbers on this page are frozen to the October 8, 2026 corpus snapshot and will not drift silently — the next re-aggregation lands around January 2027 with a fresh snapshot date and larger samples. If a number matters to your argument, cite it with the snapshot date and the sample size shown next to it.
Yes — that is why they are published. Cite them as "HooklyAI scoring corpus, October 8, 2026 snapshot" with the sample size attached, and link to this page. The only thing we ask is that you keep the sample size visible: a mean of 70.2 from 33 observations says something different than the same figure from 3, and the honesty is the point.
Every score in this study came from the free tier — 30 scored generation requests a month, hard gates and written rationale included. Score your hooks before you publish, and the next quarterly snapshot carries better data for everyone.
The methodology behind every number, and the tools that produce it