# How the analysis changed Update, 16 October 2026, by Dirišwago. https://palo.dirishwago.co.za/reports/2026-10-16-how-the-analysis-changed Its figures are as of 2026-10-16 06:26 SAST, and are not updated. Licence: CC BY 4.0; credit "Palo, Dirišwago (palo.dirishwago.co.za)". > On 16 October 2026 Palo moved its graph analysis to a faster library. Betweenness is now exact, PageRank is now fully worked out, and reach was corrected for more than 32,000 sites. What changed, and by how much. On 16 October 2026 Palo changed how it computes the measures of the link graph: PageRank, betweenness, reach and communities. How the census counts sites and links did not change; the counts of sites, links, islands and unrecognised sites, and the largest connected group, are computed as before. Three measures changed: one by choice, two by correction. The figures below compare the analysis of 04:40 SAST on 16 October, the last under the old method, with the analysis of 06:26 SAST the same day, the first under the new one. In between, the crawlers added about 585 sites to the graph (202,127 before, 202,712 after), so a small part of each difference comes from new data; most of it comes from the method. ## Betweenness is now exact Betweenness measures how often a site lies on the shortest chains of links between other sites. It was estimated from a sample: the shortest paths starting at 5,000 sites, about 2.5% of the graph. It is now computed exactly, from every site. The estimate was good at the top and poor below it. 93 of the 100 sites with the highest betweenness are the same under both methods, and 9 of the top 10. Further down, a site's estimated rank depended partly on which sites happened to be in the sample: in a test on an earlier copy of the census, the median site with any betweenness moved about 4,100 places once betweenness was exact. Exact betweenness gives the same answer for the same census, and a rank that means something for every site. ## PageRank is now fully worked out PageRank is computed by repeating a calculation until it settles. The library Palo used before stopped repeating once the values changed, in total across all sites, by less than a millionth for each site in the graph. With more than 200,000 sites, that limit is loose, and the values had not fully settled. Computed until they do, the values change, and so does the order of the sites near the top. - **Of the 100 sites with the highest PageRank,** 80 are the same as before, and 9 of the top 10. - **The top three, with all links:** domains.co.za, status.domains.co.za and sacoronavirus.co.za. Before: domains.co.za, sacoronavirus.co.za and google.co.za. - **Without the links only ever seen in a page footer:** domains.co.za is first, where sacoronavirus.co.za was; sacoronavirus.co.za is third. domains.co.za and status.domains.co.za are two sites of one organisation. domains.co.za links to one other .za site, its status page, which links only back to it, so the two pass their PageRank to each other. The census counts every host name as a site of its own. From the next analysis, Palo also publishes [PageRank between registered names](/census/top/pagerank-between-names), which leaves out the links between sites under one registered name, such as x.co.za and shop.x.co.za. Statements made from the earlier PageRank figures used values that had not fully settled. [The .za web, September 2026](/reports/2026-10-13-the-za-web-september-2026), which lists the sites with the highest PageRank, keeps the figures it was published with. ## Reach was corrected A site's reach is how many other sites it can reach by following links. Sites in the same strongly connected group can all reach one another, so Palo works out reach a group at a time. The earlier calculation left out the other sites of a site's own group, so every site in a group of two or more had its reach understated by the size of its group, less one. In this analysis the 202,712 sites of the graph form 170,454 groups, so more than 32,000 sites belong to a group of two or more. That mattered most for the largest group, 19,739 sites in this analysis: each of them reaches about 44,645 sites, where about 23,979 was shown before. Sites outside any group of two or more were counted correctly, and changed only with the data. Reach is estimated, not counted exactly: a typical figure is within about 5% of the true number. ## Communities Communities, groups of sites more densely linked to each other than to the rest, are now found with a different implementation of the same method, and the result no longer varies from run to run. The groups differ a little from before: 6,543 groups of two or more sites, where the previous analysis found 6,604. ## What did not change The number of sites and links, each site's status, the links in and out of each site, islands (64,290 of the active sites in this analysis), unrecognised sites (40,266) and the largest connected group are computed as before. The census summary, the timeline and the site records now also read every figure from the same moment, so they agree with one another. The analysis now takes about 7 minutes, where it took about 26.