A canonical asks which version to keep, and the engine decides. Fintry Fasteners supplies bolts and fixings from Stirlingshire, has 610 products at about 3,700 addresses because of sort and filter parameters, and declared a canonical on every duplicate in January 2026. 2,470 of 3,100 declarations were accepted and 630 were not.
That ratio is not a failure. It is what the instruction is: a hint, weighed against everything else the engine knows, and where they disagree, the chosen page usually has links and the declared one does not. Fintry's 630 disagreements were nearly all cases where an old filtered address had been linked from a forum post in 2022 and the plain product page had not.
The three rules that make the rest work
Every page should carry a canonical pointing at itself. That sounds redundant and it is the single most useful line here: a page with a self canonical is a page saying which of its own address variants is the real one, so a visit arriving with a tracking parameter on the end does not create a second candidate.
Core rules
Three rules that keep canonicals from being ignored
Most canonical faults come from relative URLs or chained tags created automatically during migrations.
Canonical tag hygiene
- YesSelf-canonical on every pageStops tracking parameters from creating a second candidate.Yes
- YesAbsolute scheme and domainRelative canonicals are ignored or misread by processors.Yes
- NoTwo-hop chains (A to B to C)Dropped outright; Fintry had 140 chains from a 2024 migration.No
Every page pointing to itself with a full URL prevents parameter clutter, while chained targets cancel the whole hint.
The canonical has to be an absolute address, with the scheme and the domain in it. A relative canonical is ignored by some processors and misread by others, and it is one of the two ways a plugin gets this wrong.
A canonical across a chain of two is ignored. Page A pointing at B, and B pointing at C, is a request nobody follows: the whole arrangement is dropped and A is treated as its own page. Fintry had 140 of these, made by a migration in 2024 that added canonicals without checking what the targets already declared.
What it does not do
A canonical does not stop a page being crawled. All 3,700 addresses were still fetched every month after the canonicals went in, which is the difference between this and what to let bots see, and what to keep out. A canonical consolidates; a disallow saves the fetch; and a shop with 3,700 addresses and 610 products usually needs both, in that order.
Scope boundaries
What a canonical consolidates, and what needs another control
A canonical merges duplicate signals without stopping crawls, and it must never merge genuinely separate products.
- Use both in order
- Never block real SKUs
- Do not merge distinct items
- Canonical tag
- Robots disallow
- Separate products
- Canonical tagConsolidates duplicate addresses; all 3,700 URLs still get crawled.
- Robots disallowSaves crawl budget across 3,700 filter URLs for 610 products.
- Separate productsTwo different bolts must keep distinct pages even with similar text.
- Canonical tag — Robots disallowUse both in order
- Robots disallow — Separate productsNever block real SKUs
- Separate products — Canonical tagDo not merge distinct items
Auto-canonicalising by title similarity deletes real products from search, while forgetting disallow leaves bots crawling every filter.
It also does not make two different pages into one. Two product pages for two different bolts, canonicalised together because their descriptions were similar, is a shop deleting one of its own products from search, and that is the second way a plugin gets this wrong: a rule that canonicalises by title similarity.
Where this stalls: declaring and never checking
Read the declared and the chosen in the page indexing report. Search Console lists them separately, which is the only place the 2,470 and the 630 exist, and almost nobody opens it because the declaration feels like the end of the job.
Canonicals
What you declared against what was chosen
Only one of these four cells is the outcome anybody intends, and the cell below it is where 630 of 3,100 declarations landed.
The engine agreed
The engine chose something else
The right hand column is not an error state. A site with no canonicals still has one version chosen per group, and the only difference is that nobody made a request.
The stall is that a canonical produces no feedback. Nothing errors, nothing warns, and a site can run for years with a canonical pointing at a page deleted in 2024 while everybody involved believes the duplicates are handled. Fintry had eleven of those and found them in the same report.
Canonical the sort and filter versions at the plain page, and check afterwards. The second half is the whole of the advice that is usually missing, and it is a report to open once a quarter rather than a piece of work.
What to do about the 630
The 630 split two ways, on whether anything outside the site links to the old address. Where an old filtered address has links from outside, the honest fix is a redirect rather than a canonical, because a redirect is not a hint and it removes the choice; that is moving a page without losing what it earned. Where it has none, the disagreement costs nothing and can be left.
Fixing disagreements
What to do when Google picks a different canonical
Not every rejected canonical needs work; only duplicates carrying outside links require a redirect.
- Yes (external links exist)Replace the hint with a redirectA redirect removes the choice instead of leaving an old forum link in charge.
- Export chains with Screaming Frog
- Redirect linked filter URLs to clean pages
- No (no external links)Leave the disagreement aloneAn unlinked parameter page overridden silently costs nothing in search.
- Check declared vs selected quarterly
- Fix deleted targets and chains only
Screaming Frog finds chains and relative tags across the site, while Search Console reveals where the engine disagreed.
Screaming Frog SEO Spider reports the canonical declared by every address and flags the chains and the relative ones, free up to a few hundred addresses, and it replaced reading the source of each page by hand. The export that matters is canonicalised addresses whose target is itself canonicalised, which is the chain case and is invisible any other way.
Search Console's page indexing report gives the other side, free, replacing counting pages by hand and believing the total. Its user declared canonical and Google selected canonical fields are the two numbers on this page, and they are reported one address at a time, which is why the 630 was a sample rather than a census.
Whether those 3,700 addresses should exist at all is readable, stable and shallow, and which of the 610 products is worth a page is pages that dilute the ones that matter. Both are decisions taken above the layer that decides whether any of the rest counts, and a canonical is what is left when they were not taken.