Rebuilt to be found
A marketing agency in Israel whose site nobody could find: not people, not search engines. Rebuilt from the ground up: 50 pages, under 800 KB each, not one third-party script. Here is what the server logs measured in the five days that followed.
AdFree Tools: design, build, deployment and monitoring.
Visit the site ↗Where it started
Naturally Digitale trains entrepreneurs, produces video content and runs social accounts, from Tel Aviv to Ashdod. The business was doing well. The site was not helping.
It ran on an off-the-shelf platform, with a home page weighing 9.7 MB, no heading structure, and the same description repeated across four pages. Nothing on it told a search engine, or a language model, what the company sold, to whom, or where.
The starting point, measured
| Before | After | |
|---|---|---|
| Pages Google knew of | 5 | 47 crawled out of 50 |
| Home page weight | 9.7 MB | under 800 KB |
| Main heading | none on the home page | one per page, 50 pages |
| Page descriptions | the same one on 4 pages | one per page |
| What Google understands of the offer | nothing | services, prices, cities, questions |
| Third-party scripts | loaded on arrival | none before consent |
What “under 800 KB” means
A web page is a download. Every image, font and script it pulls counts against the visitor's data and against the time before anything appears on screen.
The old home page pulled 9.7 megabytes, about the size of a three-minute song, for a page of text and photos. On a phone with a weak signal, that is several seconds of blank screen, and most people leave before it ends.
We set a hard ceiling of 800 kilobytes for the first load of any page, and refused anything that broke it. That single rule is what killed the YouTube player loaded on arrival, the social widgets, and the oversized images.
What was done
- The site was rebuilt from scratch: 50 pages, each with its own subject, title and description.
- Every page declares what it is in a format machines read: the company, the services, the prices, the cities, the questions and answers.
- The old addresses were kept alive. That detail turned out to matter more than any other, as explained below.
- During the rebuild the site answered “temporarily unavailable” rather than showing a holding page, so the search engine kept its existing entries instead of replacing them with a page about nothing.
What happened in five days
These figures are not estimates. They come from the server's own access logs, filtered to the official address ranges of each crawler.
| Who | What the logs show |
|---|---|
| 1,641 pages served, 47 of the 50 crawled | |
| GPTBot (OpenAI) | about 40 pages in 90 seconds, the night the site reopened, and it re-reads the site map every morning |
| OAI-SearchBot | 16 pages in 26 seconds, then new pages the same day they went live |
| ChatGPT, live | 8 fetches, from 8 different sessions: real people asking real questions |
| Perplexity | read the training page the day the site reopened |
The detail that did the most work
On 12 September, someone asked ChatGPT a question and it went looking for a page at the site's old address, the one from the previous platform, which it still had in memory.
That address had been kept, pointing at its replacement. So the assistant landed on the right page instead of a dead end. Had we let the old links break, it would have concluded the offer no longer existed.
Preserving old addresses is the least glamorous part of a rebuild, and on this project it is what carried five years of accumulated knowledge across.
How it was built
- No third-party script. Nothing loads before the visitor agrees to it, with no exception made for convenience.
- A weight ceiling per page, enforced on every change rather than measured at the end.
- Images served at the size actually displayed, in modern formats, and never before they are needed.
- A structure a machine can read, because a page a machine cannot summarise is a page it will not quote.
None of this is decoration. A site with nothing to load is a site that crawlers read to the end, and that is the whole story above.
Something similar in mind?
Rebuilds, performance work, and sites that stay readable by both people and machines.
Get in touch