What rises, what drowns on Hacker News

The funnel

84.8% of posts never reach 10 points

The distribution is brutally bottom-heavy. The real gate is the /newest page: a post needs a handful of upvotes in its first hour to touch the front page at all. Past that gate, outcomes spread wide — which is why the 100–499 band is bigger than the 50–99 band. Posts that escape tend to keep climbing.

What rises

Niche depth beats hot topics

Hit rate by topic is almost a mirror of how crowded the topic is. Retro-computing — a tiny niche — converts at more than double the site average. AI, the most submitted topic by far, converts worst. Evergreen HN material (languages, databases, policy fights, Linux) all sits comfortably above average.

Where hits come from matters just as much: HN is a personal-blog engine. Nearly three-quarters of all front-page hits are personal blogs and independent project sites — not media, not GitHub, not corporate blogs. Corporate blogs have the best per-post odds (12.5%) but barely submit; media is a steady middle; platforms drag you down.

The AI paradox

AI dominates the leaderboard and still has the worst odds

AI/LLM stories are 25.3% of submissions but only 18.4% of hits. The genre’s giant top-scorers (model releases, the anti-AI essay of the week) hide a flooded middle: an ordinary AI link converts at 3.6% vs 5.3% for everything else, and “Show HN” plus AI is the deadest combination in the dataset — half the hit rate of a non-AI Show HN.

What still works in AI: lab releases (anthropic.com and openai.com submissions hit ~18% of the time), benchmark-moving open models, incidents and misbehavior stories, and well-written skepticism — “I’m Tired of Talking to AI” (2,013 pts) and “They’re made out of weights” (1,529 pts) both out-scored almost every product launch. What drowns: yet another wrapper, agent framework, or AI productivity tool.

The crypto remnant

Blockchain converts only when the coin is the setting, not the subject

HN is post-crypto: blockchain is 0.6% of submissions (AI is 25%), and what remains faces indifference, residual hostility, and the heaviest flagging on the site. The window’s two biggest “crypto” stories were schadenfreude and a ban — “Crypto in 2026: Oh, This Is the Bad Place” (426 pts) and “Codeberg Bans Cryptocurrency Projects” (365 pts). The best neutral one, a Bitcoin payment processor, managed 114.

Two patterns inside the numbers. First, a voting-bloc mirage: Monero/Nostr posts sitting at 30–67 points with zero to two comments are a community upvoting its own — points without conversation, no front page. Second, the exception: content that would survive with the coin removed. A zero-knowledge-proof post framed as privacy tech hit 231; a crypto-datacenters-vs-Texas-grid post hit 154 as an infrastructure story. “Fitting a 12 GB UTXO set into a kilobyte” is a data-structures title; “Running a Bitcoin Lightning node” is a crypto title. Lead with the engineering problem and the numbers.

Thin sample (n=260) collected in a bear-sentiment window — the ceiling may be understated.

What drowns

The mortality table

Format kills more posts than topic does. Video in any form barely registers. Blogging platforms (Medium, Substack) underperform the same writing on a personal domain. A bare GitHub repo drowns 53.5% of the time — code needs a story wrapped around it. And Show HN is a harsh room: half of all launches die at 1–2 points.

The one boosted format: Launch HN (YC companies, placed by the site) — only 20 in the window, but 40% hit 100+ and none drowned. Not reachable for the rest of us.

Also invisible in this data: flags. Users with 30+ karma can flag, flags outweigh upvotes, and flagged-dead posts never reach the API — so every drown rate here is a slight undercount, worst in the most-flagged categories (politics without a tech nexus, crypto, ragebait). The outrage-adjacent winners below are survivors of a filter this dataset cannot see.

Show HN deep-dive

Niche launches: higher floor, lower ceiling

Show HNs for platform-niche tools — Linux, a specific desktop environment, terminal software — drown slightly less than average and reach 50+ points at the same rate: HN’s audience is disproportionately these people, so the niche finds you on /newest. But they cross 100+ only half as often. Winners cluster at 70–230 points — once the sub-audience that can run the thing has voted, the ceiling arrives.

The strategic fork: plain links about the same Linux/terminal niche convert at 7.6% — among the best segments on the site — because everyone can read about a thing even when only some can run it. The write-up of the project on your own domain is the higher-ceiling play; the Show HN launch is the higher-floor one. If you launch niche: a GIF or asciinema cast at the top of the README, a WASM/browser version if remotely feasible, and the platform named honestly in the title — “for Hyprland” filters in the passionate voters.

Timing

Weekends and late UTC evenings are kinder

Two forces pull against each other: audience (how many voters are awake) and competition (how fast /newest scrolls). Submissions peak 12:00–17:00 UTC — US morning — which is also when the crowd is thickest. The heatmap below shows where the balance lands: the single best cell is Sunday 00:00–04:00 UTC — Saturday evening in the US — at 7.3%, 1.7× the site average and 3.3× the worst cell. Weekends beat weekdays across the board (5.4% vs 4.2%), and the weekday 04:00–12:00 UTC trough — Europe’s morning, America asleep — is the worst place to post anything.

Show HN plays by different rules. A launch needs a big audience more than a thin /newest: Show HN converts best posted 12:00–20:00 UTC (1.9–2.1% hit) and worst 04:00–08:00 UTC (0.9% hit, 61% drown). The mega-hits follow the audience too — 56% of all 500+ point stories were posted between 12:00 and 20:00 UTC.

Caveat: timing is entangled with content. Weekend posts skew toward hobby projects and deep dives (which HN likes anyway), and corporate releases are scheduled for US mornings. Treat the weekend/evening edge as partly the content, partly the clock.

Title anatomy

Specifics rise, questions sink

Correlation, not causation — but consistent. Titles with a concrete number (“$1,600 in ESP32s”, “6.0.0”, “18 years of comments”) outperform. A year in parentheses — the classic “(1998)” repost of a timeless piece — nearly doubles the base rate. Short beats long, first-person (“I replaced…”) helps. Question-mark titles and anything tagged [video] sink.

Leaderboards

The window’s biggest stories

The top of the chart is dominated by AI-lab drama and David-vs-Goliath stories. Note what the #2 post of the entire 60 days is: a Show HN about replacing a $120k bowling-center system with $1,600 of ESP32s — cheap hardware beating expensive software, told first-person with numbers in the title. That is the HN formula in one line.

Domains that convert best (25+ submissions in the window): corporate AI labs, a handful of elite institutions — and personal blogs that have become brands. simonwillison.net converts at the same rate as openai.com.

Method

Notes and caveats

• All stories with tags=story from the Algolia HN Search API, May 25 – Jul 24, 2026 (60 days, 59,711 stories). Flagged/dead posts are not in the API, so true drown rates are slightly higher than shown — most so in heavily flagged categories (non-tech politics, crypto). Never solicit upvotes: the voting-ring detector buries posts silently and penalizes accounts and domains.
• Scores are a snapshot at collection time; stories from the last day or two were still accruing points.
• Topic buckets are keyword regexes on titles — approximate, overlapping, and blind to article content. Retro-computing’s lead comes with n=223.
• Everything here is correlation. A question-mark title doesn’t cause failure — but it does mark the kind of post that fails.