At the end of September about 35 to 47 people a day reached the site from a Google search. Since October 2 it has been 3 to 13 a day, and this week mostly under 7. We checked the plumbing first: the domain, the certificate, robots.txt, the firewall (no Googlebot blocked in seven days), and the pages as Google sees them. All fine. The problem was not that Google could not reach us. It was what we showed it.

We doubled the site in a week. On September 30 we added the Knesset archive to the sitemap - tens of thousands of new pages at once, many of them short. The sitemap grew to about 137,000 addresses. To a search engine, a small site that suddenly publishes that many thin pages looks like a content farm, not like a library.

We had two addresses for the same page. The site was also reachable at a second, technical address on Cloudflare (workers.dev), serving the same pages. One of our own pages even embedded that address for seven weeks. Google saw two copies of everything and had to guess which one was real. The second address has only redirected to the real one since October 9.

We told Google the wrong address, 84,000 times. On October 5 a publish command, rebuilt by hand instead of running the real job, wrote the technical address into the "this is the original" line of 84,000 live pages. It was fixed the same evening. The drop had already started, but this did not help.

We renamed pages without forwarding them. Fixing bad protocol titles changed their addresses. The old addresses, the ones Google knew, returned "not found". The forwarding by document number went live on October 9.

27,962 pages had the same title as another page. Meetings of the same committee all said the same thing. Google treats that as duplicates. Since this morning every one carries its date.

Google also ran a spam update through September, so not all of the drop is ours. But every item above is something we did to ourselves, and we did all of them in the same two weeks. That is the real mistake: so many large changes at once that we cannot tell which one did the damage.

What changes: a large change to what Google sees - a new section, a rename, a sitemap jump - goes out alone, and we watch the search arrivals for a week before the next one. We count those arrivals every morning now, instead of noticing three days late.

Recovery from this takes weeks, not days. We will report the number here.