On 18 October 2026 we sent the same crawler check to the same 120 AI-built sites as on 20 September, without changing a line of the script. Of the 104 sites that had answered in September, 98 gave exactly the same result on all 16 checks. Of the 60 that were still online and had looked empty to a crawler, 59 still did. The 3 sites that were gone had each disappeared in the month its domain reached the anniversary of its registration.
Our first report ended on a condition: if AI app builders began to render pages on the server by default, its main finding would vanish overnight, and we would look again. This is the second look. The finding did not vanish. What a month changed was mostly which of the sites still exist.
What we measured
- The same addresses. All 120 from September: sites from the Made with Lovable showcase,
projects their makers listed publicly as made with Cursor, Claude Code or v0, apps on
lovable.appand a few Vercel templates. The same 5 server-rendered sites served once more as the control group. - The same probe. The September script, byte for byte, under the same bot name: one plain HTTP request per check, nothing rendered and no JavaScript run. It scores the same 16 problems, from an empty page and a soft 404 to a missing sitemap or canonical link.
- Site by site. Each address was compared with its own September result. Counted as populations, the live sites went from 104 to 105. That tally hides 3 sites that vanished, 3 dead addresses that answer again and a parking page that counts as live, so every figure below follows one site across the month.
- The noise. Minutes after the main run we sent the whole probe again. 105 of 105 live sites gave identical answers both times, so a difference between September and October is a change on the site, not a hiccup in the network.
- Two records from outside the page. The registration dates of the 72 domains that the sites use as their own, from the public registry service RDAP; 59 came back with dates. And the file name of each site’s main JavaScript bundle, which we had written down on 27 September.
The results
| 20 September to 18 October | AI-built sites, n = 104 | Control sites, n = 5 |
|---|---|---|
| The same answer on all 16 checks | 98 (94%) | 4 |
| Fewer problems than in September | 2 | 1 |
| More problems than in September | 1 | 0 |
| Gone: no answer, or a parking page | 3 | 0 |
| Average problems per site still online, September → October | 5.1 → 5.1 | 1.2 → 1.0 |
| Sites that passed every core check, September → October | 6 → 6 | – |
The September average here is 5.1, not the 5.2 of our first report, because it covers
only the 101 sites that are still online. In the control group the one change was a site that added the
og:title it had lacked.
The only AI-built site with more problems got them by moving. It now runs on a different host, and on the way it lost its
canonical link, its robots.txt, its sitemap and its structured data. It was an empty page before the move and it
is one now. Of the 2 that improved, one added the sitemap it had been missing. The other is the one exception in
the next section.
The empty pages stayed empty
On 20 September, 60 of the sites that are still online gave a crawler less than 300 characters of text to read. On 18 October, 59 of them still did, and not one site went the other way and lost text it used to have.
The exception is worth a look. In September its HTML was a bare shell; in October it carried 9,287 characters of text, a canonical link and a real 404 for an address that does not exist. Its build tool and its host were the same as before. The platform had not changed. Someone had done the work on that one site.
Were the rest simply left alone, or published again with the same outcome? Vite, the build tool behind Lovable projects, writes a fingerprint of the code into the name of the main JavaScript file, so any change to the code gives the file a new name. On 27 September we had recorded that name for 76 sites. On 18 October 57 of them still loaded exactly the same file: no new code had reached them in three weeks. 52 of those 57 were empty pages.
Another 16 had a new file name, so their owners had published changes. 8 of them had been empty in September, and 7 were still empty after the new release. Publishing again did not fill the page. Had the builders moved existing projects to server rendering, this is where it would have shown.
How the sites disappeared
3 sites that answered in September did not answer in October. 2 of them no longer accept a connection; the other one shows a domain parking page where the site used to be. All 3 had been empty pages.
The registry tells more than the pages do. Every one of the 3 domains reached the anniversary of its registration between our two runs, and the registry record of each was changed between that anniversary and our second visit. 2 of them now carry the status “auto renew period”, which a registry sets when a registration reaches its end date and is extended automatically while the registrar waits to learn whether anyone will pay. Among all the sites that were live in September, only 4 had an anniversary in this window, so 3 of 4 went dark. The registry record of the fourth had last been changed in July, well ahead of its date, and the site gives the same answers as in September.
The 16 addresses that were already dead on 20 September mostly stayed that way: 13 still do not answer, and for 2 of them the registry has no record of the domain at all. Of the rest, 2 came back, one of them as an empty page, and 1 had been registered afresh on 27 September and now redirects to another site that offers the domain for sale.
On this evidence an AI-built site that nobody maintains does not decay. It stays exactly as it was, empty page included, and in this sample the event that ended such a site was its domain coming up for renewal.
If your site was built this way
Nothing in the month suggests the problem will be solved for you. The builders did not change existing projects, and publishing a new version left 7 of 8 empty sites empty. The one site that became readable did it with work on the site itself, and its changes are the first steps in our guide to the five problems.
The test is one line. Pick a line of text that shows on your home page; if the command below counts zero, a crawler that skips JavaScript does not get that line either:
curl -s https://your-site/ | grep -c "a sentence from your home page"And if the address matters to you, switch on automatic renewal for the domain. Of everything we measured this month, a domain anniversary was the most dependable way for one of these sites to disappear.
What we did not measure
New projects. An existing site is not rebuilt on its own, so if a builder changed its defaults for new projects during October, these 120 addresses could not show it. Only a fresh sample of new sites can.
Our 16 checks are yes-or-no questions. A site that gives the same answer to all of them may still have new text, new pages or a new design; “the same” above means the same to these checks, not untouched. The bundle file names are the stronger sign that nothing was published, and they cover only the 76 sites built with Vite.
An anniversary is not a reason. We see dates in a registry and what each address answers, never why an owner let a
domain go. Dates came back for 59 of the 72 domains; the others use endings without a public RDAP
service or are no longer registered, and a site on a lovable.app address has no domain of its own to lose.
And this is still one sample, drawn mostly from a single builder’s showcase, followed for a single month.