What this check inspects
Shipwork crawls a set of pages, embeds the text of each, and groups the ones that read as near-duplicates using a strict similarity threshold, deliberately higher than the near-duplicate content check so a merge is only suggested for pages that genuinely repeat each other. From each group it picks a keeper: the page with the most inbound links from the crawl, and where that ties, the most complete content by words, headings, images and structured data. Every other page is given an action, merge or redirect, a list of the unique content to move across first, and a numbered plan ending in a 301 and a sitemap removal. When a search property is connected it corroborates each group with the one fact similarity cannot see: whether the pages both rank for the same query. It also reports cannibalisation, a query two of your own pages both get impressions for, which embeddings cannot detect because the pages read differently.
What a failure means
This check needs a set of comparable pages and a working text model, so with too few readable pages or no model it reports a reason rather than a plan. Redirected and noindex pages are skipped, because a redirected page is a copy of its target and a noindex page is not competing. A group of near-duplicate pages is a warning with a keep, merge and redirect plan. A query two of your own pages both rank for is a separate warning, since one is splitting the signals of the other. Without a connected property the plan is exactly the same; the only loss is the corroborating search evidence and the cannibalisation list.
How to fix it
- Keep the page the plan names as the keeper and make sure it carries a self-referencing canonical tag.
- Before removing a page, move the unique content the plan lists into the keeper.
- 301 the other pages to the keeper and remove them from the sitemap, rather than leaving redirects and dead entries.
- Update internal links so they point at the keeper instead of the pages that were merged away.
- For a contended query, pick the page you want to rank and give each remaining page a clearly different intent.
A typical failure, worked through
The setupA store has three collection pages that say almost the same thing under slightly different names. All three are in the same crawl, and two of them both appear in search data for one query.
What the check reportsThe check groups the three pages, names the best-linked one as the keeper, lists the unique sections to move across before the others are redirected, and adds that two of the pages both rank for the same query.
The pointA similarity score tells you pages are alike. This turns that into which page to keep, what to rescue first, and the exact sequence of edits, which is what a person needs in order to act.