What happens when a website is rebuilt so AI assistants can read it?
In this case, the assistants went from naming the business in about half their answers to nearly all of them, in five weeks. The one assistant that could not read the site at all went from naming it in one answer in five to nineteen in twenty. Here is the measurement, before and after, and what it does and does not prove.
The starting point. In July 2026 I measured a martial arts academy in Zürich. Its website looked complete to a person and served 955 bytes to anything that did not run the page's code: a title and one line. Across the questions that named no academy, the four assistants named it in 47 per cent of English answers and 57 per cent of German ones. That put it fourth of fourteen local academies in English and first in German. The assistant that searches the web before answering, and so depends on being able to read a site, named it in 19 per cent of English answers. On the question about MMA, which the academy teaches, it was named in none of 24 answers. With web search switched off, so the assistants answered purely from memory, it was named in none of 180.
What changed. The site was rebuilt so that its content is present in the page as delivered, before any code runs. The same homepage now returns about 70,000 bytes of readable text. The sitemap went from 2 addresses to 22. A visitor sees the same design as before.
The re-measurement, 1 September 2026. Same questions, same four assistants, same two languages, both modes.
| July | September | |
|---|---|---|
| Named, English, questions naming no academy | 47% | 87% |
| Named, German, questions naming no academy | 57% | 93% |
| Rank among 14 local academies, English | 4th | 1st |
| The searching assistant, English | 19% | 94% |
| MMA question, named | 0 of 24 | 9 of 12 |
| Answering from memory, no search | 0 of 180 | 0 of 90 |
Movements of 36 to 40 points, against an instrument that repeats to within about two or three points. The size is not in doubt.
What this shows. When the assistants search, they can now read the site, and what they read changes what they say. The searching assistant is the clearest case: it depends entirely on what it can fetch, and it went from rarely naming the academy to nearly always.
What this does not show. Two things, stated plainly. First, the website was not the only thing that changed between July and September. The summer break ended, the timetable grew, and the directory listings were worked on. The measurement shows what moved; it cannot say which change moved it. Second, memory did not move at all. With search switched off, the assistants named the academy in none of 90 answers, as they named it in none of 180 in July. What an assistant already knows updates slowly, and no website rebuild touches it. The gains here are all in what the assistants find when they look, which is also where the fixes are within any business's reach.
The owner reports enquiries rising from roughly 10 a month to roughly 40 over the same period. That is his figure, not mine, and it sits alongside the same caveat.
Details of the academy are available on request to anyone who wants to check the working.
If you would like to know whether the assistants can read your site, and what they say when a customer asks, email contact@signal-lock.com. The first look is free.