Orviqa

Did my launch post work? How to read the days after an open-source launch

By Bright Oparaji10 min read

Write down what you posted and when, then compare the few days after it with the same few days in each of the previous four weeks. If the post's window is above most of them, something unusual happened. Whether the post caused it is a separate question, and the answer usually depends on what else you did that week.

Most maintainers judge a launch post by looking at the repository the next morning. Views are up, so it worked; views are flat, so it failed. Both readings are usually wrong, because "up" and "flat" only mean something against a normal day for your project, and few people have written that down.

This is a method you can run in a spreadsheet. It will not tell you what caused a change. It will tell you whether there was a change worth explaining, and what the candidate explanations are.

Start with what you did, before any numbers

Open a sheet with one row per action and three columns: the date (in UTC, to match GitHub), where it went, and the link. A Show HN submission, a Reddit post, a forum reply, a release, a README rewrite: anything that could change who looks at the project goes in.

Do this before you open the Traffic page. Once you have seen a spike, it is easy to remember the post that fits it and forget the three other things you did that week. The six-week digline case study built this kind of record, 253 dated actions over six weeks, and it is what made the record useful even where it could not settle the question.

Add one more column if you can: what you expect, in one sentence, written before you look. It is the cheapest protection against reading every result as confirmation.

Compare against your own usual range

Comparing against zero ("I got 200 views") ignores the views you would have had anyway. Comparing against yesterday ignores that yesterday might have been a Sunday or itself unusual. The fairer comparison is your own recent history, on the same days of the week:

  1. Pick a window starting on the day of the post. The post day and the two days after is a reasonable default.
  2. Total the repository views in that window.
  3. Go back one week and total the views for the same three weekdays. Then two weeks, three weeks and four weeks. You now have four comparison totals.
  4. The post's window is unusual only if it is above most of those four, and more convincing if it is above all of them.

Same weekdays handle weekends, and four weeks stop one strange week setting the bar.

Here is an invented example, with invented numbers, to show the arithmetic. A project's post went out on a Tuesday. Its Tuesday to Thursday views in each of the four previous weeks were 41, 55, 38 and 62. The post's Tuesday to Thursday window had 140. That is above all four, so the window is unusual. If the window had been 58, it would sit inside the range of ordinary weeks, and the honest reading would be that nothing distinguishable happened, not that the post failed.

You need the history for this to work. GitHub keeps daily views for only 14 days, so four weeks of comparison means you must already be saving them. The post on keeping GitHub traffic past 14 days has a script and a scheduled workflow for it. If you launched yesterday with no saved history, you have at most two weeks, and your notes should say so.

Read several signals, and know their shape

A post that worked usually moves several numbers, and a spike in one alone is more often noise or a bot. These are the signals worth putting side by side:

SignalWhere it livesCan be read per day?What to watch for
Repository views and unique visitorsGitHub TrafficYesThe main number for the usual-range comparison
Clones and unique clonersGitHub TrafficYesScripts clone too; only interesting if views moved as well
Referring sitesGitHub TrafficNo, only as a 14-day windowDoes the venue you posted to appear at all?
StarsRepositoryOnly roughlyGitHub no longer gives a timestamp for a star
Site visitors and tagged linksYour site analyticsYesArrivals through a link tagged for that one post
DownloadsYour package registryYesRelease days dominate; read them with care

GitHub reports referrers as the top 10 sites over the last 14 days, so you cannot say how many came on the day of the post. You can compare a snapshot taken a few days after the post with one taken before it, and check whether the venue appears at all. Bear in mind that links shared in chat and email often arrive with no referrer: in the digline record, 39 percent of site visits in one week landed as "unknown".

Tagged links are the most direct evidence a small project can collect. Give each post its own tag (UTM parameters, or whatever your analytics reads as a campaign), then click it once and check that the click shows up. The digline links were tagged for six weeks with a parameter the analytics treated as a source rather than a campaign, so the tags recorded nothing beyond the referrer.

Stars and downloads are the numbers most people check first, and the least suited to this job. GitHub stars vs active users explains why a star measures attention rather than use. Downloads have a bigger problem, covered next.

List everything else that happened nearby

The post is only one possible cause of a rise. Before you credit it, write down everything within a few days either side:

  • Releases. In the digline record, 9,415 of 10,378 downloads, 91 percent, landed on release days. The median release day had 343 downloads and the day after had 71. A new version is fetched by bots, scanners and CI jobs whether or not anyone read your post. If the post went out on release day, downloads cannot separate the two, and measuring adoption goes into why release-day counts are mostly machines.
  • Other things you did around the same time. Over 40 days, the digline maintainer released on 25 and promoted on 31, with 23 days having both. Only 7 days had neither. With that overlap, nothing could be attributed to any single effort.
  • Weekends and holidays. The same-weekday comparison handles weekends; a public holiday can still shift a week.
  • Someone else linking to you. A mention in a newsletter, a blog or a thread you never saw. One digline essay scored a single point on Hacker News and appeared in Python Weekly ten days later. If you had only looked at the Hacker News submission, you would have called it a failure.

Put these in the sheet too, marked as things you did not do.

Investigate spikes with nothing recorded

Sometimes a window is well above the usual range and nothing in your sheet is nearby. Those are worth chasing. Check the referrer snapshot from the days after, search for your project's name and repository URL, and look at popular content to see which pages people opened. An unexplained spike is often a link from somewhere you did not know about, which tells you about a room where your users already are. If you never find the cause, write "unexplained" and leave it. That is more honest than guessing.

Be honest about what the comparison can say

The method tells you whether a window was unusual for your project, not that the post caused it. A few limits belong in your notes:

  • A rise after a post is not proof. Something you did not record may have happened. The list of nearby events narrows the candidates; it does not choose between them.
  • One post on one day is one sample. A good window could be luck and a flat one bad timing. You learn about a venue from several posts.
  • Small projects have noisy baselines. At a few dozen views a week, ordinary weeks vary a lot, so only a large change stands out. The digline docs had about 30 pageviews a week, and at that volume no single intervention was expected to show up as a rate.
  • Too little history means you cannot judge yet. If you do not have four comparable weeks, say "cannot judge" rather than reaching for yesterday. It is a real answer.

The digline case study has a cautionary example: an early regression on its downloads reported that releases multiplied them six to eight times, until the comparison group turned out to be three days and the figure was withdrawn. A clever method on too little data is still too little data.

If you would rather not keep the sheet by hand

Orviqa's Analytics page has a growth timeline covering the last 60 days, with each day that has something recorded beside that day's stars, views, clones, site visitors, tagged arrivals and newly seen referrers. Each action's repository views over its day and the two days after are compared with every three-day window in the previous four weeks (the usual range is the 10th to 90th percentile of those totals), and described as above, within or below it. Other actions recorded close by are listed beside it, a day with unusual views and nothing recorded is shown as unexplained, and with too little history it says it cannot judge. It never says an action caused a change, and it calls no model and costs no tokens.

Either way, pair the record with a plan from how to promote an open-source project or the promotion overview: decide where to post, then decide in advance how you will read the result.

A checklist for your next post

  • Start saving daily GitHub views at least four weeks before you post
  • Record the post's date (UTC), venue and link before you look at any numbers
  • Give the post its own tagged link and test that a click is counted
  • Keep releases off the day of the post, or record both
  • Compare the post's three-day views with the same weekdays in each of the four weeks before
  • Call the window unusual only if it beats most of those four
  • Compare referrer snapshots from before and after, as 14-day windows
  • List every release, action and outside mention nearby, and chase unexplained spikes
  • Write "cannot judge" when the history is too short

← All posts