If Google's AI Overview cites your page today and your URLs change next month, you keep the citation by giving the machines a clean forwarding address: a permanent (301) redirect from every old URL to the closest live equivalent, the same core facts still present in plain text on the new page, and nothing in your robots rules or rendering that stops an answer engine from reading it. Do those three things and the answers usually re-point to you within a crawl cycle or two. Skip any one of them and you become the restaurant that moved without telling the delivery drivers.
Which is exactly the scene I want you to picture. You've moved buildings. Same menu, better kitchen, nicer sign. But every driver's GPS still has the old address burned in, and half of them, when they hit a locked door, don't call you. They just mark you closed and drive to the place across the street.
Why the answer engines lose you in the first place
Here's the thing people forget. Google AI Overviews, which rolled out broadly in the US in May 2024, don't invent their sources. They pull from pages Google already crawls and indexes, and they show links to a handful of them. Perplexity crawls with its own agent (PerplexityBot). ChatGPT's search reads the live web through OAI-SearchBot. Every one of these systems is, underneath the fancy summarizing, doing the same unglamorous thing your grandmother's browser did: fetching a URL and reading what comes back.
So when you replatform and the URL that earned the citation returns a 404, you haven't just lost a page. You've told every one of those drivers the building is gone. And an AI answer engine, unlike a patient human, will happily substitute your competitor mid-sentence and never mention that anything changed.
That's the part that stings. Classic SEO breakage is loud (rankings drop, you notice). AEO breakage is quiet. The answer still appears. It just isn't yours anymore.
The forwarding address has to be permanent
A 301 redirect tells crawlers the move is permanent and to transfer the old URL's earned reputation to the new one. A 302 says "temporary, don't bother updating your records," which is the last thing you want when the whole point is that the records must be updated. Redirect chains (old to interim to final) bleed signal and burn the crawl budget you're trying to conserve. Point old straight at final. One hop.
And redirect to the relevant page, not the homepage. Bulk-redirecting a thousand retired URLs to your front door is the single most common way teams vaporize their AI citations. The answer engine cited a specific paragraph about, say, warranty terms. Send it to your homepage and you've answered a question nobody asked. Google treats those homepage dumps as soft 404s anyway, so you get the worst of both.
If you want the mapping discipline that survives launch, we walk through it in the migration planner. The short version: every URL that ever earned a citation gets an owner and a destination before cutover, not after.
The facts have to still be on the page
Redirects get the machine to the new door. What it reads once inside is what keeps the citation. Answer engines quote sentences, so the sentence had better still exist.
Migrations love to quietly launder facts. The old page said "processing takes three to five business days" in plain body text. The new design turns that into an icon, or a tab that loads on click, or an accordion that's collapsed (and empty) in the initial HTML. To you it looks identical. To a crawler that reads the first response, the fact is gone. If the answer engine can't find the sentence it used to quote, it stops quoting you.
Keep the extractable stuff extractable: real headings that name the question, the direct answer in the first line or two under them, and structured data (FAQ, Product, HowTo) that survives the move instead of getting rebuilt "later." Boring, checkable, present in the raw HTML. That's what gets lifted into an answer.
Don't lock the door behind you
New platforms ship new defaults, and some of them wall off the very bots you need. Check that your robots.txt and any WAF or bot-management rules still allow Googlebot, OAI-SearchBot, GPTBot, and PerplexityBot if you want to appear in those systems. A staging environment set to noindex that ships to production with the flag still on is a genuinely common, genuinely catastrophic mistake.
On llms.txt, the proposed file for pointing language models at your content: it's fine to add, but don't count on it. Google has publicly said it isn't using it, and no major answer engine has confirmed reliance on it. It's a business card left on the counter, not a forwarding address.
What actually moves the needle
| Lever | Effect on AI citations | Effort |
|---|---|---|
| One-hop 301 to the relevant page | High: preserves earned reputation | Medium |
| Same facts in raw HTML | High: keeps you quotable | Low to medium |
| Crawler access in robots and WAF | High: prevents total blackout | Low |
| Preserved structured data | Medium: helps extraction | Medium |
| llms.txt | Unproven | Low |
What I'd actually do
Treat your existing AI citations as inventory, the same way you treat backlinks and top-traffic URLs, and freeze that list before anyone touches a redirect map. My honest position: the teams that keep their answers aren't the ones with the cleverest AEO tactics. They're the ones who did the dull migration hygiene properly and then didn't strip the facts out of their own pages in the redesign. Everything glossier than that is a rounding error.
So before you flip the switch, go stand at your old address and knock on the door yourself. If a crawler would find a locked building and a competitor across the street, fix that first. The AI didn't forget you. You changed your address and forgot to tell it.