Not the ones that read best. Migration shortlists are usually drawn up on editorial judgement, which quietly discards pages that earn citations and links. Decide on data instead: what has inbound links, what earns impressions, and what is currently cited in AI answers. Then apply editorial judgement to what survives.
The short version
- Editorial judgement alone will drop pages that are quietly earning links and citations.
- Decide on three data sets: inbound links, search impressions, and current AI citations.
- Being off-message is a reason to rewrite a page, not to delete its URL.
- Everything not carried over still needs a redirect to something genuinely relevant.
- Do this before the URL structure is signed off, because afterwards it is expensive.
Why the usual method loses things
The standard approach is a content lens: somebody reads the back catalogue and keeps what is evergreen, on-brand and still true. That is a sensible starting filter and it is not sufficient, because it cannot see the two things that matter most.
It cannot see links. An old, unglamorous post can be the single most linked page on your domain, often because it answered a narrow technical question years ago. It reads as dated. It is load-bearing.
It cannot see citations. A page can be the source an assistant currently uses to answer a question about your category. Nothing in the page tells you that. You have to go and look.
There is also a framing error worth naming. When a business repositions, old content is judged against the new proposition and much of it fails that test. But being off-message is an argument for rewriting the page, not for deleting the URL. The URL holds the accumulated authority. The words on it are replaceable, and replacing them on an established URL is far cheaper than earning that position again from nothing.
The three data sets that should decide
Inbound links. Any page with credible external links is carried over or redirected precisely. Never dropped. This is not negotiable and it is the cheapest check on the list.
Search impressions. Pull the last twelve months from Search Console. Pages with meaningful impressions are earning attention even where clicks are modest, and on question-shaped queries low clicks can mean an assistant is answering in place using your content.
Current AI citations. Run your priority questions across the assistants and record which of your URLs get cited. Those pages are actively working. Losing one is losing a position you will have to win back.
Only once those three lists exist should editorial judgement decide what is rewritten, merged or retired among what remains.
How to check it yourself
- Export your top pages by inbound links. Anything with real links goes on the keep list.
- Export twelve months of Search Console impressions by page, not just clicks.
- Run 20 to 30 buyer questions across ChatGPT, Perplexity, Gemini, Claude and AI Overviews and record every URL of yours that appears.
- Overlay the three lists against the proposed migration shortlist. The gaps are what you were about to lose.
- Check the redirect map covers everything else, one hop, to a genuinely relevant destination rather than the homepage.
What to do about it
- Do this before URL sign-off. After that, changes are expensive and often refused.
- Rewrite rather than delete where the URL has equity but the words are off-message.
- Merge near-duplicates into the strongest URL rather than carrying both.
- Redirect everything else specifically. A bulk redirect to the homepage passes almost nothing and tells a model nothing.
- Record a baseline before launch. Once the old site is gone, you cannot prove what you had or diagnose what you lost.
The full migration sequence is in keeping your AI citations through a site migration.