Can this work for an arbitrary website?

Yes as an assisted archive viewer. This experiment does not establish that a URL can reliably become a complete timelapse without review.

What we observed

We targeted January and July from 2000–2025 for four large, long-lived sites. Of 208 targets, 93 produced reviewed examples: Google 27, Apple 25, Amazon 32, and Yahoo 9. The other 115 were unavailable or rejected. These are observations from this dataset, not a success-rate estimate for the web as a whole.

A separate Feeder run on 28 September 2026 selected 56 candidates from 127 monthly-indexed Wayback captures across 2012–2026. It yielded 49 reviewed stills, five visually rejected renders and two failed archive requests. This denser sequence is not capped at one capture per half year. Repeated layouts at different dates remain real observations; 49 captures does not mean 49 redesigns. The resulting collection contains 142 reviewed captures across five sites. Feeder holds each still for two seconds, twice the original examples.

The limiting factor was usually the archive render, not making a video. A replay could return successfully yet show a consent page, an empty feed, missing CSS or missing principal images. Later Google captures often used a basic Portuguese layout. Amazon also changed locale. Yahoo's newer news pages often lost their main content. A smooth transition cannot repair those gaps.

What should generalize

The discovery command already accepts a custom URL. Once reviewed screenshots exist, the fixed-size capture, dated gallery, speed controls and crossfade export can work without knowing the site's design. Static, public, mostly server-rendered pages are a plausible good fit; that is an engineering inference, not a measured result on a representative sample of sites.

Estimated layout morphing is less general. This prototype matches unique short link/button/heading text or IDs, requires four matches and rejects very large movements. Those matches can still be incidental navigation. They do not prove that two images or sections have the same meaning. Redesigns, different languages, duplicate labels and missing DOM geometry need a plain crossfade. The untouched stills remain the source of truth; intermediate shapes are only visual estimates.

What an arbitrary-URL version would need

• Normalize the requested page and retain redirects, URL changes and exact replay timestamps. Decide explicitly whether the user wants one page or a whole site.
• Discover first, then show the available date range and gaps before rendering. A requested six-month cadence is a target, not an archive guarantee.
• Offer a small preview for review before producing a long video. Surface missing imagery, language changes, consent screens and likely archive error pages.
• Pace archive requests, cache completed work and resume after failures. Rate limits must stop requests; do not retry around them with extra workers.
• Keep the current archive-only network boundary. A hosted URL-input service would additionally need private-address/redirect protection, browser isolation, resource limits and job quotas. The local script is not that service.
• Keep exact dates and source links, allow gaps, and make crossfade the fallback. Never fill missing dates with invented historical pages.

Login-only, personalized and script-heavy pages are poor candidates for faithful historical reconstruction. We have not tested authenticated archives, a broad sample of small sites, full-site crawling, or a production multi-user queue. We also have no measured per-site operating cost, throughput or reliability claim.

A useful next experiment would take a varied set of user-supplied public URLs, record discovery/render/review outcomes separately, and compare static pages with script-heavy ones. That would test generality instead of extrapolating from four well-known homepages. The current result supports a reviewed research tool; it provides no basis for promising complete histories of arbitrary sites.

Feeder collection

The collection includes Feeder’s 49 reviewed stills and a 3:34 video, with playback speed controls and dated source links. The original four examples remain available.