Blog

AEO During a Replatform: Surviving the Answer Box

· 6 min read · By the Replatform Radar team

When you replatform, an AI answer engine treats you like an old friend saved in its phone under a number you no longer have. It knows your name. It trusts what you said last year. And when someone asks a question you can answer, it dials that number without checking — and gets nothing. To keep Google's AI Overviews, ChatGPT search, and Perplexity pointing at you after the URLs change, you need three unglamorous things: clean permanent redirects the AI crawlers will actually follow, structured data and entity signals that say this is still the same company that knew the answer, and enough patience to let each engine re-dial the new number on its own schedule.

That's the whole article, really. But nobody believes it until they watch it happen.

Why the phone-contact metaphor is annoyingly accurate

Here's the scene that got me thinking about this. A site does everything right on the classic SEO side of a migration. Redirects mapped one-to-one, rankings dip for a week, then claw back. Everyone exhales. Then someone notices the brand stopped appearing in AI Overviews for a set of questions it used to own, and it stays gone for a good deal longer than the blue links did.

Because the answer engine isn't Google's live crawler racing to re-index. It's a friend holding a saved contact card. When you moved (changed URLs), a normal search crawler eventually knocks on the door, sees the forwarding note, and updates its address book fast. The AI layer is slower and grumpier. It cached what you said, it saved the URL it said it from, and it will happily quote a page that now 404s until something forces it to look again.

Call forwarding is your 301 redirect. Set it up and the call reaches you. Miss one, or use a temporary 302 (the equivalent of "I'll be back in five minutes"), and the caller assumes you're just out, keeps the old number, and never bothers to save the new one. John Mueller has been clear for years that a 301 passes full signal with no PageRank loss. Use them. Every single one.

Who's actually dialing you

The confusing part is that "AI" isn't one caller. Several different bots want to reach you, and they follow different rules and re-crawl on different clocks. Blocking the wrong one, or assuming they're the same, is where people quietly delete themselves from answers.

Bot / tokenWhat it feedsBlocking it means
GooglebotSearch and the grounding behind AI OverviewsYou disappear from search and the AI answer
Google-ExtendedGemini API and Vertex model trainingYou opt out of training, but AI Overviews are unaffected
OAI-SearchBotCitations shown in ChatGPT searchChatGPT stops linking to you as a source
GPTBotOpenAI model trainingYour content isn't used for training runs
PerplexityBotPerplexity's index and answersPerplexity can't cite the new URLs
ClaudeBotAnthropic crawlingClaude's context won't see the move

The one that trips people up is Google-Extended. It sounds like it governs AI Overviews. It does not. AI Overviews are grounded through ordinary Googlebot access, so if you block Google-Extended in a fit of "keep the robots out of my content," your Overviews carry on regardless — you've only opted out of a different thing. Read your own robots.txt before cutover as if a stranger wrote it, because after the migration a stranger sort of did.

Prove you're still the same person

Redirects get the caller to the right door. Structured data is how you convince them it's still you behind it. Answer engines lean on entity signals to decide who to credit, and a replatform is exactly when those signals get scrambled: your Organization schema gets half-rebuilt, author markup drops off, the sameAs links to your profiles vanish because nobody ported the JSON-LD, canonical tags point at staging.

So the identity work is boring and it matters. Keep the Organization and Article schema intact on the new templates. Keep sameAs pointing at the same social and knowledge-panel profiles. Keep author entities consistent. Make canonicals point at the live production URL, not the CMS's internal preview path (a classic new-platform own-goal). If the answer engine can look at the new page and match it, byte for concept, to the entity it already trusts, you keep the citation. If the page suddenly looks anonymous, you're a wrong number.

One honest note about llms.txt: it's a proposal from Jeremy Howard's team, floated in late 2024, for a plain file that hands models a clean map of your site. It's a sensible idea. It is not, as of now, honored by Google, OpenAI, or Perplexity in any load-bearing way. Ship one if you like tidiness. Don't ship it instead of the redirects and the schema, because that's decorating the mailbox while the phone rings unanswered.

The part you can't rush

Even with everything perfect, re-discovery isn't instant, because these engines cache. Search links usually recover in days to a few weeks. The AI layer lags behind that, sometimes noticeably, because it's quoting a stored snapshot and only refreshes when its crawler comes back around. Knowing which of your URLs feed answers before you move — so you can watch exactly those and re-fetch them the day you cut over — is the difference between a two-week gap and a two-month one. That mapping is a big part of what a pre-migration crawl like ours is for: flagging the pages that are already answering questions so they don't fall off the map silently.

So what would I actually do

I'd treat AEO as a first-class part of the redirect plan, not a nice-to-have after go-live. Before cutover: inventory the URLs that already show up in AI answers, keep their schema and canonicals identical, map them to hard 301s, and audit robots.txt so I'm not accidentally hanging up on OAI-SearchBot or PerplexityBot. After cutover: request re-indexing on the answer-earning pages first and expect the AI layer to lag search by weeks, not days — and not panic about it.

Because the friend with your old number isn't ignoring you. They just haven't heard you moved. Leave a forwarding note they can't miss, keep your name on the door, and they'll call the new place soon enough.

Questions and discussion

If you've replatformed recently, how long did your AI Overview or Perplexity citations take to come back compared with your regular search rankings, and did anything you tried actually speed it up? If you never measured the AI layer separately because nobody asked you to, say that too, because I suspect that's most of us and I'd rather know than guess.
Steven Solano, who wrote this — and reads every reply

Loading discussion…

Want this analysis for your exact site before you migrate?

Request a scan →