You Redesigned the Site. AI Lost You.

A lighting factory in Dongguan rebuilt its website last year. The new site looked better, the product categories finally made sense, and the whole company shared the launch on WeChat Moments.

Two weeks later an old customer messaged them on WeChat: is your site down? I clicked through from Doubao and got a blank page.

The technical check was blunt: every old product URL returned 404, the new pages used a different address structure, and no redirects had been configured in between. The pages that AI had been citing all pointed at addresses that no longer existed.

In AI search optimization this counts as a self-inflicted injury. Not one piece of content was lost, yet the key AI was holding no longer opened the door.

Boundaries first. This piece is about protecting existing AI recognition when URLs change. It is not about front-end design and it takes no view on whether a redesign was worth doing. If your URLs never change, skip it. One last limit: GEO can connect old addresses to new ones. It cannot bring back content that was thrown away, which depends on whether you kept a backup.

What AI sees during a redesign

Working on GEO means holding two perspectives at once. From the buyer's side, a redesign is the site getting upgraded. From the retrieval chain's side, it is something else entirely.

What it remembers of you is a set of URLs, each mapping to a piece of content. After the redesign, that set becomes a run of 404s and the new site is a batch of addresses it has never seen.

Give it no signal and its conclusion is simple: the old content disappeared, new content appeared, and the two are unrelated.

This is not stupidity. It judges only on signals. A person would ask whether the address changed. A machine does not ask.

So in generative engine optimization, the right posture for a redesign is not observe after launch. It is build the bridge before launch.

Four common mistakes, ranked by cost

First, using 404 where a 301 belongs. This is the most severe and the most common. An old URL returning page not found tells the retrieval chain the content is gone. A 301 says it moved here.

Second, redirect chains that run too long. The old address points to an intermediate page, that page points to the new address, and sometimes it runs three or four hops. Google states it will follow up to ten hops, but explicitly recommends pointing straight at the final destination, with chains of no more than three hops and never more than five. The longer the chain, the more likely one link fails.

Third, the canonical still points at the old address. The tag was never updated, so the page keeps declaring that the authoritative version lives at an old URL which now returns 404. That is the site telling the retrieval chain, on its own initiative, to look somewhere that no longer exists.

Fourth, the sitemap was never resubmitted. The site launched and the sitemap is still the old one, or a new one went in without the old one being replaced. Discovery slows down accordingly.

Ranked by cost in AI search optimization: the first and third are fatal, the second and fourth slow you down. In a real redesign project these four items sit with four different people, which is exactly why they get missed.

What Google actually says

Google's documentation on moving a site contains three conclusions worth isolating, because the popular versions of them are frequently wrong.

First, a 301 does not strip equity. The documentation states that 301 and other permanent redirects do not cause a loss of PageRank. In other words, when the redirects are configured correctly, the signal travels with them. Internal discussions often turn this into the claim that changing a domain always costs you half your visibility, and on that basis some teams skip the change entirely, which is a different loss.

Second, keep redirects for at least a year. The documentation recommends retaining them as long as possible, typically at least a year, because during that window the retrieval chain gradually transfers signal to the new addresses, including from links on other sites that still point at the old ones. Removing them earlier takes the bridge down ahead of schedule.

Third, the timeline is weeks or longer, with no fixed number. The documentation states that for a medium-sized site, Google tends to take a few weeks or more to start showing new URLs instead of old ones, and longer for large sites; the move itself proceeds URL by URL, and it is only complete once every URL on both the old and the new site has been visited at least once.

Those three statements describe search engines and they are also the foundation of GEO. The answer layer adds another variable on top: each AI provider runs its own indexing and refresh cadence, unannounced and outside your control. So the question of how long recovery takes in AI has no answer anyone can promise. All you can do is build the bridge correctly and wait.

One acceleration is worth mentioning: beyond submitting a sitemap, push the changed URLs through IndexNow. It guarantees nothing about indexing, but it announces the change rather than waiting to be discovered.

Five things to do before and after

Listed in the order an AI search optimization project runs, usable directly as a checklist.

First, build a URL map before anything ships. List every old URL that carried traffic and pair it with its new destination, one to one. That table is the foundation of the whole migration; without it everything downstream is guesswork. To decide what carried traffic, use two sources: the landing page report in your search console, and the addresses that appear in your own server logs.

Second, configure 301s at the server, one for one. Do not take the shortcut of redirecting the whole site to the home page, because that tells the retrieval chain every old page became the home page and effectively voids all of them. Rules belong at path level.

Third, check canonicals and internal links. Every new page must declare its own new URL as canonical, and every internal link that still points at an old address must be rewritten. This is the most tedious step and each omission adds one more break.

Fourth, resubmit the sitemap and keep the old one for at least a month. The new sitemap tells the retrieval chain which pages exist; the old one, left in place for a while, helps it discover which addresses changed.

Fifth, keep the redirects for at least a year and resist the cleanup instinct. Many teams run a technical debt sweep three months after launch and delete the redirect rules along with everything else. That is the most common way a migration rolls itself back.

Before the launch window, run one spot check most teams skip. Take ten old URLs that matter, one from each product family, and walk them in a staging environment with the redirects enabled. The point is not to confirm that the browser lands somewhere. It is to see how many hops it takes, and whether the landing page is the precise counterpart rather than a category page. Category pages are the usual culprit: a product URL redirects to its parent category because the product was merged, and the buyer who clicked expecting a spec table arrives at a grid of twelve items. That redirect technically works and converts nothing. The same check catches trailing-slash mismatches and case differences, which look harmless in a browser and can loop a crawler indefinitely.

Finish those five and GEO has done what it can. The rest is time.

One more layer: old addresses inside AI answers

One point runs against intuition. Search engines recrawl; answer layers may not. The same batch of changes can take effect weeks apart in ERNIE Bot and ChatGPT, because each provider's indexing cadence differs and none of them publish it.

AirOps reports in State of AI Search 2026 that 48% of citations in AI answers come from community platforms and 83% of commercial citations come from pages updated within the previous year. Both point at one thing: AI prefers pages that were updated recently. After a redesign, if the new pages carry no clear update signal, they look to AI like a batch of new pages with no history.

That suggests one move during a redesign: important pages should carry an explicit publish date and update date, visible on the page, with a machine-readable version alongside it to save the parser a step. Not for appearance, but to give recency something readable to attach to.

Another move: after launch, go update the pages other people write about you. Industry directories, marketplace storefronts, company registry listings — if they still carry old addresses, buyers arriving from AI hit a 404 anyway. You do not control their content, but you do control the URL field.

Both are closing work in GEO. Skipping them causes no immediate problem; doing them makes a difference two or three months out.

Three questions that come up

We only changed the category structure, not the domain. Do we still need 301s?

Yes. The documentation is explicit that the change of address tool is only needed when moving across domains or subdomains. A path change inside one domain still needs 301s, because to a retrieval chain the URL is the identity: change the path and you have a new address.

The old page content is gone for good. Can we just return 410?

You can, provided you distinguish the cases. A 410 means permanently gone, which is clearer than a 404 and lets the retrieval chain give up sooner. If the page truly offers no further value, 410 is correct. If it merely moved, you need a 301, or you are voluntarily abandoning recognition you already earned.

Rankings dropped after launch. Should we roll back?

Not yet. The documentation notes that ranking fluctuation during a move is normal, because pages are being recrawled and reindexed. The test is not how much you dropped, but whether every old URL 301s to its correct new destination. If the mapping is right and the redirects are live, the rest is waiting. If the mapping is wrong, rolling back will not save you; fixing the mapping will.

Closing

Back to the lighting factory. They added three things: an old URL list rebuilt from server logs, with a 301 for each entry; canonicals rewritten to the new addresses; and a resubmitted sitemap, pushed once through IndexNow as well.

Six weeks later, the same link clicked from Doubao landed on the new product page.

The description inside the AI answer was still the old one, but at least a human could get in. The bridge is dull work, and it is the only part of this fully within your control.

2024 was the year Yihang shifted to GEO; before that the work was cross-border B2B and Alibaba.com services. Marketplace storefronts have one characteristic in their favour: the URL structure is set by the platform and sellers cannot change it, so the addresses can never be lost in a redesign. A company website is different. The right to change it sits with you, and the freedom to get it right is the same freedom that lets you get it wrong. Work down to this layer of AI search optimization and the most expensive failure is rarely a missing optimisation. It is an upgrade that zeroes out everything accumulated so far, in one afternoon.

Data and effect boundaries: statements about 301s not causing a loss of PageRank, redirect chains of no more than three hops with a maximum of ten, retaining redirects for at least a year, medium-sized sites taking weeks or longer for new URLs to replace old ones, migrations proceeding URL by URL, and the change of address tool applying only to cross-domain moves all come from Google's documentation on moving a site, in its own terms. That document describes Google Search crawling and indexing; it does not mean other search engines or AI answer layers follow the same cadence. The 48% and 83% figures from AirOps are vendor-published observational statistics, not independently reviewed, and reflect a non-Chinese market snapshot. The Dongguan lighting factory is an illustrative situation and describes no specific client. Every figure cited is a mechanism description or an observational statistic; none of it predicts results for any business. The methods preserve an existing indexable state. They promise no recovery timeline, no ranking, and no volume of inquiries.

References

1. Google Search Central: How to move a site (301 and other permanent redirects do not cause PageRank loss; chains followed up to 10 hops, recommended no more than 3; redirects kept at least a year; medium-sized sites take weeks or longer; moves proceed URL by URL). https://support.google.com/webmasters/answer/6033049
2. Google Search Central: Change of address tool (only for cross-domain or subdomain moves; same-domain path changes do not need it). https://support.google.com/webmasters/answer/83106
3. AirOps: State of AI Search 2026 (48% of citations from community platforms; 83% of commercial citations from pages updated within a year; vendor-published, observational).
4. What Gets Cited, accepted at SIGIR 2026 (252,000 controlled experiments; topical relevance and context position dominate, recency of timestamp consistently positive; peer reviewed).