There aren't a lot of Pearson correlations on HN's front page

Let’s take a minute to appreciate that the front page of Hacker News is extremely hard to hack. What appears there is unpredictable, varied, and put there by humans. 

As a blog and copywriter for software companies, I (along with Matt Guay) am often asked to publish articles and pages that hit the front page. But I’ve learned over the years, from experience and from data, that there’s no time of day or day of the week with disproportionately more “winners.” There aren’t any magic words to put in your title. Or domains to post from. 

To do well on Hacker News, you cannot appeal to an algorithm or a pay-to-play scheme (unless you’re a YC company, but let’s ignore them). You have to enthrall and entertain the kind of users who choose a forum with a design that hasn’t changed in 18 years. People like that can smell a min-maxxed post a mile away.

The cruft

Based on the BigQuery dataset, there were 384,871 Hacker News posts between Nov 1, 2024 and Nov 1, 2025, more than 1,000 per day. The average score across all of those posts was 15.64 points. The average number of comments on a post was 10.63. But if you visit /new often, you know those numbers are much higher than what you typically see.

The top five posts of the past year (SQL-free approximation here) scored over 3,000 points each, pulling up the average score and number of comments by a lot. How much? Using APPROX_QUANTILES(score, 100)[OFFSET(50) tells us the median score–line up all the post scores and pick the score in the middle of the whole ordered series–is two points. For comments it’s zero. More than half of HN posts get two or fewer points and zero comments.

Let’s move it to the top quartile with APPROX_QUANTILES(score, 100)[OFFSET(75)`. That returns four points and one comment to get you in the top 25%.

If one of your posts scored 66 points or more between Nov 1, 2024 and Nov 1, 2025, you were in the top 5%. Even then, you might not be on the front page for more than an hour, if at all. Between the excellent voting ring detection and human moderation (welcome tomhow!), it’s nearly impossible to break into that cohort with brute force or pattern conformity. 

If you can get certain users to post something for you from their account, you have a better chance of scoring well. But all that proves is that amidst a drumfire of AI-related nothingburger posts, knowledge of the community and discerning taste are by far the most important ingredients to hitting the front page.

Karma refactoring

Two years ago, Matt and I put together a guide to what had, anecdotally, worked well for us on Hacker News. No data, just vibes. Matt explained that,

“I tend to post early morning Europe time (around 4PM Singapore / 8AM London / midnight Pacific), in hopes of it being read + upvoted during the Asian afternoon and European morning ... and if so, and it keeps momentum, then it’ll start the East Coast morning with votes and discussion.”

Unfortunately, the numbers don’t really bear that out. If you select by the top 100,000 posts for the Nov 1 - Nov 1 date range, extract the hour and adjust it to US West Coast time, median scores and comments are fairly evenly distributed.

The per-hour averages are a bit more random but still clustered between 50 and 60 points. 7am to 3pm have slightly lower median scores but higher averages, which makes sense since they’re both the noisiest and the highest traffic. Similarly, Sunday has significantly fewer posts and very slightly higher positive signals.

“Middle of the week is the best time to post on Hacker News,” we wrote. “People are busy on Mondays.” Whoops. Still, these are somewhat large datasets and the dimensions are too vague to reveal anything beyond the fact that you have a decent shot at any time, as long as you have something humans want to read and respond to.

Titles and flames

Most of what I’m doing here involves over-relying on score as a metric for success, which is…problematic. You can get on the front page fairly easily by tossing a gasoline-soaked headline into a submit field. That’ll yield a bunch of karma and, if it’s a company or personal domain, flames that might be impossible to douse.

I had a python script pull out one- and two-word pairs that appeared more often than others in the top 100k posts. AI was the most frequent standalone keyword over the past year (ugh), appearing in 8,705 out of the top 100k posts (and more than double the number two spot). Trump was fifth and Elon Musk was the the most frequent bigram, appearing in more than 3,000 posts, combined. Social media, China, and White House were also in the top 15 most frequent words or phrases. These are keywords that came up frequently and often scored highly, but were not positive conversations.

Looking at the top 5,000 posts, the Pearson correlation between common words/phrases and scores didn't turn up anything more than more than +0.12 (pebble was near the top at .084, with thank hn, and honda standing out). And nothing went lower than -0.04 (rust and ask hn both coming in at -.036, as well as notables like 2020, and framework). Word count too was negatively correlated with score.

You could post low-wordcount titles stuffed with devwhistles, but the data seems to indicate that they don’t move the needle that much. 

Posting users and domains

The data seems to indicate that there’s no meta that’ll win you a coveted front page slot. What about data that shows the opposite? Do certain people and domains consistently outperform? Kind of. When extract domains that have at least 10 posts over the past year, then pull their median score, only 23 of them tend to score above the four points that put a post in the upper 25%:

Simonewillison.net is really the only non-journalism site on the list. Big news domains are the only ones showing up frequently and consistently scoring well. That, compared with youtube.com links that were the second most common domain by post count, but not on this list because of their atrocious median and average scores. 

[Apropos of nothing, my SQL query counted 3,110 twitter.com vs 33 x.com and 1,293 old.reddit.com vs 101 reddit.com, which is sincerely hope is an affirmation of HN’s stubbornness and not my horrible querying.]

What about individual users? It’s here that we find the humanity in Hacker News: 341 users posted around once per month and with median scores performing in the top 25%. A smidge under 50 users had median scores 10 or higher.

You cannot (easily) hack your way to the Hacker News front page. You can, however, participate in the community, read the room, and learn what it wants and what it deplores. Unlike so many other high-traffic internet hangouts, the algorithm does not play kingmaker. It, alongside HN’s incredible human moderators, prevent ranking manipulation and encourage rough consensus. 

As someone who accepts payment for helping software companies earn positive attention from developers, Hacker News is one of the most fair, human places to hang out.

There are still people who I downvote into oblivion. And people who have downvoted me. But at least it’s still real people doing the burying. And the exalting. 

Photo by Evgeniy Smersh on Unsplash