Every replatform begins with the same optimism and ends with the same email. Organic traffic is down forty per cent, nobody can say precisely why, and the new design is getting the blame. It is almost never the design. It is the URLs.
Tim Berners-Lee wrote Cool URIs don’t change in 1998. It resurfaced on Hacker News again this week, which tells you something: twenty-eight years on, the industry still has not internalised a document that fits on one screen. In 2026 the stakes are higher than they were, because search engines are no longer the only things reading your address space.
TL;DR
- Pew Research found that 38% of webpages accessible in 2013 had vanished a decade later. 23% of news pages and 54% of Wikipedia reference sections contain at least one dead link.
- Documented site migrations lose anywhere from 20% to 70% of organic traffic, and the usual cause is incomplete one-to-one redirect mapping rather than content or design.
- Only 301 and 308 redirects act as canonicalisation signals that forward link equity. 302, 303 and 307 do not. Keep chains to three hops or fewer.
- Berners-Lee’s 1998 rules still hold: keep file extensions, technology names, CMS artefacts and organisational structure out of your URLs.
- AI crawlers fetch proportionally more 404s than Googlebot and cite specific URLs in answers. An unstable address space now costs you citations as well as rankings.
The rot is measurable
Link rot sounds like an academic concern until you look at the numbers. Pew Research Center’s 2024 analysis found that 38% of pages that existed in 2013 were no longer accessible ten years later, using a deliberately conservative method that only counted nine hard error codes. Pages from 2023, less than a year old at the time of testing, were already failing at 8%.
Follow-on research puts roughly 8% of links breaking within their first three months and 44% gone within seven years. Nearly 30% of URLs cited in legal opinions no longer resolve, and about half of those cited in US Supreme Court opinions are dead. These are institutions with archival mandates and budgets. Your marketing site, mid-replatform, on a Friday, is not better positioned.
The important detail is that almost none of this content actually disappeared. The organisations still exist, the pages were rewritten rather than deleted, and the words are still on the server somewhere. What broke was the address.
What actually changes URLs
In our experience there are four culprits, and only one of them is a genuine business reason.
The CMS changed. WordPress to a headless stack, Volusion to Shopify, Drupal to something newer. Each platform ships an opinionated default permalink structure and most teams accept it. One documented Volusion-to-Shopify replatform saw organic clicks fall from around 1,200 to roughly 500 a day, caused by long redirect chains and missing 301 mappings rather than anything the customer could see.
The org chart changed. Berners-Lee called this out in 1998 and it has not improved. Putting a department, a product line or a subdomain division into a path guarantees the path becomes wrong the moment somebody reorganises. The resource did not move. The company did.
The technology leaked into the path. .php, .aspx, /cgi-bin/, /index.html, ?p=4231. These encode an implementation detail into a public contract. When the implementation changes, and it always does, either the URLs break or you carry a compatibility layer for the rest of the site’s life.
Somebody wanted better keywords. This is the one that hurts most, because it is usually well intentioned. Rewriting /services/development/ to /custom-software-development-dublin/ for a marginal keyword gain trades a known asset for a speculative one. Search engines treat the new address as a fresh page and re-assess it from close to zero.
Designing an address space that survives
Treat URLs as a public API, because that is what they are. Third parties have written them into documents, emails, invoices, printed materials and now model training corpora, none of which you can reach to update.
A few rules that hold up:
- No file extensions and no framework fingerprints. Your routing layer should be free to change without your address space changing with it.
- Use stable facts, not mutable ones. Creation date is fixed forever. Category, author, status and pricing tier are not. Anything that appears in your URL should be something that cannot change without the resource itself becoming a different thing.
- Short, lowercase, hyphenated, no trailing state. Session identifiers, tracking parameters and sort orders belong in query strings that you explicitly mark as non-canonical.
- Decide the trailing slash once and enforce it at the edge. Serving the same content at both variants quietly halves your signals and doubles your crawl cost.
- Version your APIs, not your pages.
/api/v2/is good engineering./blog-new/is a confession.
When you genuinely have to move
Sometimes the change is unavoidable: an acquisition, a rebrand, a legal name change. The discipline then is entirely in the mapping, and it is dull work that nobody wants to own.
Export every indexed URL from Search Console, your server logs, your sitemap and your analytics, then reconcile the four lists. Server logs matter most, because they include the long tail that analytics never recorded and that still carries inbound links. Map each one to a specific destination. Redirecting unmapped pages to the homepage is the same as deleting them; Google treats a mass homepage redirect as a soft 404.
Use 301 or 308 exclusively. Google’s own documentation is explicit that 301 and 308 act as canonicalisation signals that forward link equity, while 302, 303 and 307 do not. Then flatten the chains. If a 2019 migration already redirects A to B, your new mapping should send A straight to C, not A to B to C. Google advises against chains longer than five hops and the practical ceiling is three.
After launch, update internal links to point at final destinations rather than relying on the redirect layer, regenerate the sitemap, and keep the redirect map in version control as a tested artefact rather than a spreadsheet somebody emailed. On a domain change, keep the old domain live for at least two to three years. Renewal is cheap; the backlinks are not.
The new reason this matters
Until recently the argument for URL stability was ranking equity and user trust. There is now a third consumer.
AI assistants cite specific URLs in their answers, and analyses of crawler behaviour show AI bots fetching proportionally more 404s than Googlebot, apparently because they probe archived and inferred paths more aggressively. A stale citation in a model’s answer sends a qualified visitor to your error page. Worse, a URL that has been rewritten twice since it entered a training corpus is effectively a dead entry that nobody will report to you.
The same applies to agent traffic. Coding agents, procurement agents and research agents all need stable, dereferenceable identifiers to work against. If your documentation URLs shuffle with every release, every agent that has cached a path against your service breaks silently. This is the reason we push clients towards durable resource identifiers and explicit deprecation windows on public endpoints: it is the same discipline as URL stability, applied one layer down.
The llms.txt proposal has attracted attention here, though the empirical picture is unflattering: crawlers overwhelmingly skip the file and read HTML directly, and Google has confirmed it does not support it. Chasing the new file is far less valuable than making sure the URLs you already publish keep resolving.
A short audit you can run this week
- Crawl your own site and count 404s and redirect chains longer than one hop. Both numbers should be near zero.
- Pull the top 200 URLs by inbound links and confirm each returns 200, not a redirect.
- Check your CMS default permalink structure for dates, IDs or extensions you would not choose deliberately today.
- Confirm your redirect rules live in the repository and are covered by tests, not in a hosting control panel that one person can edit.
- Add a link checker to CI so a broken internal link fails the build rather than the customer.
None of this is glamorous. It is also the difference between a replatform that holds its traffic and one that spends eighteen months recovering it.
REPTILEHAUS handles migrations, replatforms and API design for clients across Ireland and Europe, and redirect mapping is a line item in our project plans rather than an afterthought in launch week. If you have a replatform coming and would like someone to own the address space properly, get in touch.
📷 Photo by Brendan Church on Unsplash
