Software companies ship on a three-day week
We read the changelogs, blogs, docs and release feeds that 263 software companies publish on their own domains. Between 1 January 2026 and 14 September 2026 those companies put out 25,999 dated updates, and 60.3% of them landed on a Tuesday, Wednesday or Thursday. A flat week would be 42.9%.
The half we did not expect: Friday is the quietest working day, not Monday. It runs 22.6% below Wednesday, and 104 of the 154 companies we could measure individually publish less on it.
What this measures, and five things it does not
Sample: 25,999 updates, 263 companies, 1,070 first-party sources, published between 1 January 2026 and 14 September 2026. One row is one update a company published on a page it owns. Computed from the live database on 14 September 2026.
Counted: pages served from the company’s own domain or a subdomain of it, whose own markup states a publication date. Excluded: news coverage about a company, undated pages, and our own domain.
The limits, in the order they would bite
- This is not a random sample of companies. It is the set our customers asked us to watch, which skews to business software and developer tools. Read every number below as a statement about those 263 companies and nothing wider.
- Only dated pages can be in a date study. 59.1% of the pages we hold state a publication date; the rest are absent from this entirely. Coverage is not evenly spread: release feeds nearly always carry a date, pricing pages and product catalogues never do, so this sample is software companies and contains no ecommerce catalogue at all.
- Third-party coverage is thrown away. A news article about a company is a newsroom’s publishing decision, not the company’s. Requiring the page to live on the company’s own domain is the single filter this finding rests on.
- Timezone is not recoverable per publisher, and the error runs against us. Dates are stored as each company published them. A Friday evening post on the United States west coast can land on Saturday once normalised, which inflates the weekend. The true weekend share is therefore at or below the 6.6% below.
- There is no claim about time of day here. 29.5% of the non-midnight timestamps in this sample sit at exactly 12:00:00, which is what our extractor writes when a page gives a date and no clock. An hour-of-day chart from this data would be an artefact of that default. We are not publishing one.
Wednesday carries 6.5 times what Sunday carries
| Day | Updates per calendar day | Per day | Share |
|---|---|---|---|
| Mon | 118.4 | 16.8% | |
| Tue | 145.8 | 20.2% | |
| Wed | 147.3 | 20.4% | |
| Thu | 138.8 | 19.8% | |
| Fri | 114.0 | 16.2% | |
| Sat | 23.7 | 3.4% | |
| Sun | 22.6 | 3.2% |
Updates per occurrence of that weekday, so an eight and a half month window holding more Mondays than Tuesdays cannot read as a Monday effect. Raw share is in the last column, and the two orderings agree.
60.3% of updates in the window fall on the three midweek days, against 42.9% for a flat week. 6.6% fall at the weekend, against 28.6%. The pooled figure is partly a statement about the handful of companies that publish most, so here is the whale-immune version: taking one figure per company and then the median across the 154 companies with at least 20 updates, the typical company puts 67.4% of its output into those three days. 72 of those 154 companies published nothing at all on a weekend in the entire window.
Friday is the quiet day, not Monday
The received idea is that nothing happens on a Monday. In this sample Monday runs at 118.4 updates per Monday and Friday at 114.0, against 147.3 on Wednesday.
Paired per company, so no single publisher can carry it
Comparing each company against itself across the 154 we could measure: 104 publish less on Friday than on Wednesday, 47 publish more and 3 are level. A two-sided sign test on that split gives p below 0.0001.
Monday is below Wednesday too, for 94 of 154 companies against 50 above and 10 level, p below 0.001. Both ends of the working week are quiet. Friday is the quieter of the two.
Is this just our crawler’s own schedule?
It is the first thing we tried to prove, because if our crawler ran harder on weekdays it would manufacture this exact shape. Four checks, each of which could have killed the finding.
One: the dates are theirs, not ours
weekend 11.0% by our clock vs 6.6% by theirs
Every figure on this page uses the date the company stamped on its own page. Bucketing the very same rows by the moment our crawler wrote them gives a flatter, smeared distribution, which is what a uniform sampler observing a lumpy stream looks like. A weekday-skewed crawler would give a sharper one, not a flatter one.
Two: the crawler has no concept of a weekday
scheduling is an interval in hours
Sources are due for a crawl on a per-source interval and a loop that runs around the clock. There is no day-of-week branch anywhere in it to produce a day-of-week effect.
Three: sources we visit at least twice a day
weekend 6.3%, midweek 59.9% on 19,320 updates
Restricting to sources on an interval of 12 hours or less removes the possibility of missing a weekend item and catching up on Monday. The shape does not move.
Four: the archive, where our schedule could not have selected anything
weekend 5.6%, midweek 63.0% on 4,235 updates
These are back-catalogue pages published before the study window across 238 companies, pulled in during a single first visit to each source: the median one was already 441 days old when we first saw it. A recurring schedule had no opportunity to select them by weekday, and they replicate the live result. This is the strongest of the four.
The attacks we could not make stick, and the one we could
A few companies dominate the row count
True. Dropping the 4 highest-volume publishers outright (pabau.com, google.com, zoho.com, cisco.com) leaves 16,526 updates across 259 companies, with a weekend share of 3.7% and a midweek share of 64.5%. The effect gets stronger without them, not weaker.
Would it flip in a quarter?
Across the 11 quarters we hold first-party dated pages for, the weekend share stays between 2.1% and 8.5%, and the midweek share between 56.6% and 67.8%. It never inverts and never comes near the 28.6% and 42.9% a flat week would give. Earlier quarters rest on far fewer rows, so they are shown with their counts rather than smoothed away.
| Quarter | Updates | Midweek | Weekend |
|---|---|---|---|
| 2024Q1 | 159 | 57.2% | 7.5% |
| 2024Q2 | 317 | 64.7% | 5.4% |
| 2024Q3 | 272 | 59.9% | 5.5% |
| 2024Q4 | 233 | 59.7% | 2.1% |
| 2025Q1 | 378 | 56.6% | 5.6% |
| 2025Q2 | 490 | 67.8% | 8.4% |
| 2025Q3 | 609 | 67.3% | 4.8% |
| 2025Q4 | 882 | 63.2% | 3.7% |
| 2026Q1 | 4,625 | 62.9% | 6.8% |
| 2026Q2 | 10,610 | 58.9% | 8.5% |
| 2026Q3 | 10,764 | 60.7% | 4.7% |
The companies that disagree, named
A reader who checks one company and finds a counter-example should find it here first. These publish the largest weekend share of any company with at least a hundred updates in the window. The last column is the share of that company’s dates our extractor filled in from a date with no clock, which is the reason to discount some of these rather than treat them as evidence.
| Company | Updates | Weekend | Date filled in |
|---|---|---|---|
| aftership.com | 325 | 28.9% | 56.9% |
| zoho.com | 1,653 | 24.3% | 5.9% |
| yahoo.com | 409 | 22.7% | 39.4% |
| baidu.com | 117 | 22.2% | 96.6% |
| clickcease.com | 108 | 20.4% | 0.0% |
| google.com | 2,758 | 19.2% | 14.5% |
The attack that lands
The sample. These companies are here because somebody paid attention to them, not because they were drawn at random, and a page that says a company ships midweek is describing a company that publishes a changelog in the first place. If you want to know when companies in some other category ship, this page does not tell you, and no amount of arithmetic on this sample will.
The aggregate as JSON, and the shape of the query
Every number on this page is read from the database when the page is built, not typed into it, and the page prints the date it was computed: 14 September 2026. The same aggregate is served as JSON so it can be cited and diffed rather than transcribed.
It carries the per-weekday counts, the calendar-day denominators, the per-company split, every quarter in the stability table, the four confound checks and the same caveats in machine-readable form.
The selection, in words
Take every stored update that carries a publication date, join it to the source it came from and the company that source belongs to, and keep it only when the page’s host is that company’s domain or a subdomain of it. Ceiling the date at the present moment, because a small number of pages advertise a future year in their own text and our extractor believes them. Then count by ISO day of week, and divide each by how many times that weekday occurred in the window.
This is a by-product. Spyingbee watches the pages your competitors publish and tells you what changed, and the archive that makes this page possible is the same archive a customer reads one company at a time.