The sample
- 46 public WordPress sites, drawn at random from the Tranco top 1M list of popular websites (list GQG5K, 19 October 2026). We put the list in a random order and kept the first 170 sites whose home page loaded WordPress files, which took checking 1,451. Each was crawled once by Silotree’s crawler, the same crawl as the free audit, reading up to 500 pages. 112 sites were bigger than that. A partial crawl makes posts look unlinked when the posts linking to them weren’t read, so those sites are left out, and this edition describes the 46 sites we could read in full, which are the smaller ones. 12 sites couldn’t be crawled (they refused, failed, or asked crawlers to wait minutes between pages) and are left out too.
- Addresses that are not public sites are left out before anything is counted: local and private addresses, reserved test names, example domains and Silotree’s own site.
- No site is named. Every figure is an aggregate across sites, and none is published from fewer than 10 sites.
What counts as a post and a link
- A post is an indexable page in the crawl: it answered normally, isn’t marked noindex, isn’t an archive or a listing, and isn’t the home page. These are the posts the Silotree Score counts.
- A link is a link inside a post’s article text. Menus, footers, sidebars and archive pages are left out, because they link everything equally and say nothing about which posts the writing points readers to.
- An unlinked post has no link in from another post. One with one or two is counted separately; the score treats 3 or more as linked.
The score figures
- Each site is scored with the Silotree Score, formula version 1, the same arithmetic the product uses. Sites with fewer than 5 posts get no score and are left out of the score figures; the report says how many.
- A part that can’t be measured on a site (Organised, when its posts weren’t grouped into topics) is left out for that site rather than counted as zero, so each part’s figures cover only the sites it was measured on.
- Medians and percentiles are across sites, interpolated between the two nearest values (the method spreadsheets use by default). Totals of posts pool every site, so big sites weigh more there; the report gives the per-site figure beside each pooled one.
The search comparison
The question: within the same site, do posts with more links in from other posts get more clicks from Google?
- Data. Search Console clicks and impressions over 28 days, from sites whose owners connected it.
- Which sites. A site joins with at least 30 posts, at least 10 of them seen in Search Console, and some spread in how many links its posts have. Sites left out are counted with the reason.
- The comparison. A regression with a separate baseline for each site, so posts are compared only with posts on the same site. Links in are used two ways: as unlinked, one or two, and 3 or more; and as a continuous count. Clicks are on a log scale, because a few posts get most of them.
- Held level. Word count, click depth from the home page, and the post’s age. Because links shorten click depth, the same model is also reported without depth, and once with nothing held level, to show how much the controls change the answer.
- Uncertainty. 95% intervals, clustered by site once there are 10 or more sites. A second check with fewer assumptions ranks posts within each site and pools the sites’ correlations.
Thresholds set in advance
These were written down before any production data was seen, so a result can’t be dressed up by moving them afterwards.
| Result | Needs |
|---|---|
| Nothing reported | Anything less than below |
| Indicative, worth reading | 10+ sites and 198+ posts |
| Adequate, published as a finding | 30+ sites and 787+ posts |
Where these come from: detecting a small association (a partial correlation of 0.1) at the usual 80% power and 5% significance takes about 787 posts; a moderate one (0.2) about 198. Posts on one site resemble each other, so those are floors. 30 sites is the common rule of thumb for intervals clustered by site to be trustworthy, and stops one big site carrying the result.
What we will and won’t claim
- A result is stated with its interval and the number of sites and posts behind it, or not at all.
- A result whose interval includes zero is published as “no clear difference”, as prominently as a positive one.
- We won’t claim that adding links causes more traffic. This is observational data: authors link to their best and most popular posts, so a positive association is expected even if links did nothing.
- We won’t claim the Silotree Score predicts a site’s traffic. Sites differ in subject, age, backlinks and brand, none of which the score sees.
- No figure from a subgroup picked after seeing the results, unless it is labelled exploratory.
Search Console leaves out anonymised queries, so per-post clicks undercount. Publication dates are often missing; where they are, age is filled in with the site’s median and flagged. A link counts once the crawl finds it, which can be later than it was written.
Checking a figure
The aggregates behind every figure are downloadable as JSON, with the date of the run and a fingerprint of the data it read (836879eba5e6). Each edition is a new run of the same code, and its output is archived with it.