Technical foundations for AI search

When the same content is reachable at several URLs, the signals that would have accumulated on one page are split across all of them. A canonical tells engines which address is the real one and consolidates the rest onto it.

This matters more than usual on multilingual sites. If an untranslated page still renders under every language prefix, you have three addresses serving identical text. Pointing the canonicals at the source-language version keeps one indexable copy instead of three competing ones.

A language cluster must name every version the page genuinely has, including itself. The failure mode catches people out: if a cluster points at a translation that does not exist, engines discard the entire cluster rather than ignoring the bad entry.

So a site that claims German and French copies it has not written is worse off than one that claims nothing. Generate the alternates from a list of what actually exists rather than from the list of languages the site supports.

A sitemap is a statement about which URLs you want crawled. Listing pages that redirect, 404 or only exist in one language wastes crawl budget and muddles the signals around them.

Generate it from the same source of truth that defines your routes, so a page added to the site is added to the sitemap automatically and a page removed disappears from it. Hand-maintained sitemaps drift within weeks.

Content published by nobody is harder for any system to weigh. A named author with a real biography, a stable page of their own and links to their profiles elsewhere gives both search engines and assistants something to attach credibility to.

This is the practical end of E-E-A-T. It is not a score you can check; it is a set of signals — who wrote this, what do they know, has anything here been updated — that either exist on the page or do not.

A visible last-updated date helps readers and machines judge currency, and freshness carries real weight on fast-moving subjects.

Setting that date without changing the content is a short-lived trick. Systems compare what the page says against what it said before, and a page that claims constant updates while saying the same thing teaches them to discount the signal.

None of this is novel. It is the base layer, and it is worth confirming before spending a quarter on content.

• Retrieval crawlers are allowed in robots.txt, named explicitly rather than left to a wildcard
• Every important page returns 200 and is reachable without JavaScript execution
• Canonicals point at one real address per piece of content
• hreflang names only languages that genuinely exist
• The sitemap is generated, not hand-written, and lists only live URLs
• Articles carry a named author and a genuine modification date
• Structured data describes what is actually rendered on the page