Will a site migration hurt our AI visibility?

It can, and nobody can honestly promise it won't — the systems that cite you re-crawl and re-index on schedules you don't control. The risk concentrates in changed URLs, dropped or merged pages whose cited passages disappear, rendering shifts that hide content from crawlers that skip JavaScript, and structural rewrites. The defensible approach is to reduce those specific risks and measure the crossing with a locked baseline.

Do redirects preserve AI citations?

They help and are worth doing comprehensively — permanent, one hop, every old URL — but they are not a guarantee. How each AI retrieval system honors redirects, and how quickly, varies and is only partially observable from outside; stale index entries can persist for a while. Treat redirects as necessary risk reduction, not as preservation anyone can promise.

How do I replatform without losing AI answers?

There is no guaranteed method, only risk reduction: keep URLs stable where you can and redirect the rest permanently; inventory the pages and passages currently cited and move them intact rather than reorganizing them in flight; make sure content remains present in raw HTML rather than moving behind client-side rendering; stage changes rather than combining them; and baseline before so you can re-scan after.

How long does AI visibility take to recover after a migration?

Unknowable in advance — re-crawl and re-index schedules are not public, differ across engines, and change. What you can do is distinguish lag from faults: a cited page that no longer resolves, a broken redirect chain, or content missing from raw HTML are fixable problems to hunt immediately, while fault-free waiting is measured patiently against your pre-migration variance band across multiple scans.

How site migrations affect AI visibility

Guide · SEO / AEO / GEO · 5 min read · last verified 2026-07-27

Reviewed before publication Editorial board Independent commercial review

In shortMigrations put AI visibility at risk through changed URLs, dropped passages, and rendering shifts — and nobody can promise preservation. The measurable plan: baseline before, redirect everything, re-scan after.

A site migration is any change to the infrastructure beneath your content — a new domain, a replatform, a URL restructure, a rendering overhaul — where the words may stay the same while everything about how machines fetch and parse them changes. The honest starting position: nobody can promise your AI visibility survives one intact, because the systems doing the citing re-crawl, re-index, and re-weight on schedules you neither see nor control. What a team can do is know which specific changes carry the risk, reduce them where possible, and measure the crossing instead of guessing about it.

Why AI answers care about your infrastructure

AI answer surfaces cite pages, and what they have learned about you is bound to specifics: particular URLs, particular extractable passages, particular structures that made a page easy to lift an answer from. As currently observed, the systems involved refresh their indexes and their retrieval layers on their own cadences — different per engine, undisclosed, and subject to change without notice. A migration asks every one of those systems to relearn where your evidence lives, and none of them commits to a relearning schedule. That asymmetry — you migrate on your timeline, they relearn on theirs — is the core of the risk.

This page is about the infrastructure move itself. Two neighboring situations look similar and are not: changing your name or brand severs a different association entirely, covered in how to rebrand without losing AI recognition, and editing pages that stay at stable URLs is a gentler discipline, covered in how to refresh old content without losing AI citations. If your project combines a migration with a rebrand, you are running both risks at once — stage them apart if you possibly can, because combined causes are unreadable afterward.

The changes that tend to transfer poorly

Changed URLs. A citation is an address, and after migration the address is wrong until each system relearns the new one. Redirects help — more on their limits below — but the relearning is not instant and not uniform across engines.

Dropped and merged pages. This is the quiet one. The specific passage that earned a citation may simply cease to exist: consolidating a set of focused pages into one long page can delete the extractable unit an engine was lifting, even though "the content survived" by human accounting. Before migrating, inventory which of your pages and passages currently appear in AI answers, and treat those as cargo to be moved intact, not material to be reorganized in flight.

Rendering shifts. A replatform that moves content from server-rendered HTML into client-side JavaScript can make unchanged words unreadable to systems that fetch raw HTML without executing scripts — behavior observed across many AI crawlers at present, detailed in how JS rendering affects what AI can read. This one deserves special attention because it fails silently: the site looks identical in a browser while the machine-readable version has gone dark.

Structural rewrites. Headings, question-shaped sections, and self-contained answers are part of what made pages liftable. A redesign that keeps the words but dissolves the structure can lose the properties that earned the citations in the first place.

What redirects can and cannot do

Do them anyway, comprehensively: old URL to new URL, permanent, one hop, no exceptions — it is the cheapest risk reduction available and the closest thing to a standard practice this problem has. But be precise about what they buy. Redirects are a well-understood signal in traditional search; how each AI retrieval system honors them, and how quickly, is only partially observable from outside and varies. Stale index entries can persist for a while, answering with the old URL after it stopped resolving. So treat redirects as necessary and insufficient: they shorten the relearning, on the evidence available, but they do not guarantee it, and no vendor or consultant can promise otherwise on systems they don't operate.

The measurable plan: baseline before, re-scan after

Before the migration, establish the baseline — the full procedure is in how to set an AI visibility baseline. Run enough scans, spaced out, to know your normal variance, because AI answers move between runs even when nothing changed, and you are about to need the ability to tell a migration effect from ordinary noise. Freeze the question set completely; a migration is precisely the moment the set must not change, for the reasons argued in when to change a locked question set — a before-and-after comparison is only readable if the instrument held still. Record which pages and passages are currently cited, and the date each scan ran.

During, change one thing at a time where the project allows: replatform, then restructure, then any brand work — staged, not simultaneous. Every variable you migrate together is a cause you can no longer isolate afterward.

After, re-scan on the same locked set at the same cadence. Compare against the variance band, not against a single pre-migration number. Track three populations separately: citations that moved to the new URLs, citations still pointing at old URLs, and citations that vanished. The first is recovery, the second is lag, the third is where the diagnostic work starts.

Reading the after-picture

A single weak scan after a migration is not a verdict; movement has to clear the variance you measured beforehand, in the same direction, across multiple scans, before it counts as an effect. Recovery time is genuinely unknowable in advance — re-crawl schedules are not public, and any confident promise about them deserves skepticism. What separates waiting from worrying is the fault check: a previously cited page that no longer resolves, a redirect chain that loops or hops, content that vanished from raw HTML — these are infrastructure faults you can fix today, and they should be hunted immediately. Absent faults, the honest posture is patient measurement: keep scanning, keep the set locked, and let the trend — not the anxiety — tell you what the migration actually cost.