The canonical tag: a real signal, never a rule Google has to follow.

What the tag actually promises

A rel=canonical link tells Google that the page it sits on is a duplicate, or a close variant, of some other URL, and names that URL as the one worth showing in search results and crediting with the page's signals: links, authority, and the content Google uses to judge quality. That's the entire job. It is not, and has never been documented as, a ranking factor in its own right. Search Engine Journal put it plainly in an October 2023 piece reviewing Google's own guidance on canonicalization: "The rel='canonical' tag is more of a suggestion than a directive, and definitely not a factor for rankings." The same piece flagged a change Google had made to its canonical help documentation two months earlier: as of May 2023, Google's own text says the canonical link element "is not recommended for those who wish to avoid duplication by syndication partners," closing off a use case site owners had been leaning on for years.

One signal among more than twenty

The mechanics behind that suggestion were laid out by Gary Illyes of Google's Search Relations team on the Search Off the Record podcast, in an episode covered by Barry Schwartz at Search Engine Roundtable on 4 November 2020. Illyes described duplicate detection as the first step: pages are stripped of boilerplate, navigation and footers, anything that isn't the actual content, and what's left is reduced to a checksum. Pages with matching or near-matching checksums get grouped into a cluster. Canonicalization is a separate step after that: Google picks one page from each cluster to actually show in results, using, in Illyes' account, more than 20 signals. He named content, PageRank, HTTPS, sitemap inclusion, server-side redirects and rel=canonical as part of that set, weighted by a machine-learned model rather than hand-tuned rules, because manual tuning kept causing side effects elsewhere in the system. Illyes was direct about where that leaves ranking: canonicalization, he said, is "completely independent of ranking." The page Google picks gets ranked normally afterward. The selection itself is a separate decision your canonical tag only partly controls.

When Google overrides your canonical anyway

Google has said this for years: the tag can lose. John Mueller said as much on Twitter in July 2018, after a site owner asked why Google had indexed page 3 of a paginated series instead of page 1, which carried the rel=canonical. "A rel=canonical isn't a guarantee, so ultimately that can happen," Mueller wrote, as reported by Search Engine Journal days later. He made a related point about the self-referencing version of the tag a year later: asked in a mid-2019 Google Webmaster Central hangout how important it is for a page to canonicalize to itself, Mueller called it "a great practice" but added "it's not critical," since Google's own signals can usually identify the right URL without it, per a Search Engine Journal piece from June 2019.

The sharpest version of this comes from SEO consultant Glenn Gabe, who ran a crawl audit for a client and found a large number of pages canonicalized to URLs with content that wasn't actually equivalent. In a Google webmaster hangout Gabe cited, Mueller explained the mechanism: when the canonicalized page's content differs meaningfully from its target, Google's algorithms can treat the canonical tag as a mistake and index the page anyway. On that client's site, a simple site: search turned up roughly 1,360 of these unintended indexed pages. A full export from Google Search Console, pulled with a spreadsheet add-in, found the real number: 2,867 indexed URLs that should have been consolidated away. Several were earning real impressions and clicks nobody had planned for. Gabe published the writeup in January 2017, and the underlying behavior, a canonical tag only working when the content actually matches, hasn't been contradicted since.

What changed this year

Google revised its canonicalization troubleshooting guidance again in 2026. Search Engine Journal's SEO Pulse roundup, published 17 July 2026, covered the update: pages can sit inside a duplicate cluster for up to two weeks after a genuine content fix before Google re-evaluates them, less if the revised content is clearly distinct from everything else in that cluster. The guide treats a wrong canonical, a broken redirect and a server misconfiguration as three separate, independently fixable problems rather than one bucket, and its stated starting position is worth sitting with before fighting Google's choice: the canonical Google picked might simply be the stronger page. We found the same dynamic from the other side while covering duplicate content and AI search: Google overriding a stated canonical isn't rare or accidental by design, it's a known, measured outcome of the clustering step described above.

Does the canonical tag matter for AI search?

For Google's own AI features, the honest answer is: probably, since nothing public suggests Google built a second, separate pipeline just for them. The more useful fact is this: Google's own generative AI search optimization guide, published 15 May 2026 and covered by Search Engine Roundtable three days later, tells site owners to reduce duplicate content as part of maintaining a clear technical structure, the same housekeeping job a canonical tag exists to do. Google is pointing site owners at exactly the problem rel=canonical was built to solve, inside guidance aimed squarely at AI search, not classic search.

What we don't have is anything close to the test that exists for hreflang. Our hreflang guide covers a side-by-side comparison an SEO consultant actually ran across ChatGPT, Perplexity, Claude, Google AI Mode and Copilot, and found a real, measurable split in behavior. No one has published that kind of test for canonical tags. OpenAI, Anthropic and Perplexity AI have not said, anywhere we could find, whether GPTBot, ClaudeBot or PerplexityBot read a rel=canonical link at all, let alone whether it changes which URL gets quoted when the same text sits live at several addresses. That is a real, open gap, not a settled no.

What to actually do

Set a self-referencing canonical on every indexable page, including the ones you're sure have no duplicates, since parameter and tracking-tag variants create accidental clusters more often than anyone plans for. Only canonicalize a page to a URL that genuinely represents the same content; Gabe's case study above is what happens when that rule gets ignored at scale. Don't lean on rel=canonical to manage syndication partners anymore; Google's own May 2023 documentation update specifically pulled that use case out from under the tag. Check Search Console's "Duplicate, Google chose different canonical than user" report on a schedule rather than once, since that's the one place Google tells you directly when it picked a URL you didn't. None of this is a guess-and-check job on a site of any real size. It's exactly what a technical SEO audit is built to map systematically, alongside the duplicate URL variants that waste crawl budget on the same sites that tend to get their canonicals wrong.

What we could not verify

Illyes' "more than 20 signals" figure, and the specific weighting between them, has never been published as a full list; we know the category names he gave, not the complete set or their relative weight. Gabe's 1,360-versus-2,867 case study describes one client site in 2017, not a benchmark that generalizes to how often this happens industry-wide. Neither OpenAI, Anthropic nor Perplexity AI has published documentation on how their crawlers treat rel=canonical, so the AI-search section above describes an absence of evidence, not a tested result, and should be read that way.

Sources

The 20-plus-signal canonicalization process comes from Barry Schwartz, "Google Goes Deep On Dupe Detection & Canonicalization," Search Engine Roundtable, 4 November 2020, reporting Gary Illyes on the Search Off the Record podcast. The 2018 "isn't a guarantee" quote and the 2023 ranking-factor verdict come from Search Engine Journal: "Google: Rel=Canonical Isn't A Guarantee," 24 July 2018, and Search Engine Journal's canonicalization ranking-factor guide, 3 October 2023. The self-referencing-canonical comment comes from Search Engine Journal, "Google: Self-Referencing Canonicals Are Not Critical," 14 June 2019. The client case study comes from Glenn Gabe, GSQi Marketing Blog, 26 January 2017. The 2026 troubleshooting-guide update comes from Search Engine Journal's SEO Pulse roundup, 17 July 2026. Google's generative AI search guidance comes from Barry Schwartz, "Google Publishes How To Optimize For Generative AI With Mythbusting," Search Engine Roundtable, 18 May 2026.

Common questions