Will a website redesign hurt your AI visibility?

A website redesign usually does hurt AI visibility, and it tends to hurt it harder and for longer than it hurts Google rankings. Answer engines cite a specific URL, and the sentence they lifted sits at a specific place on that URL. Change the address, the wording, or the structure of the page, and the citation that pointed at it has nothing left to point at.

The reason the damage is worse than in classic search is that the recovery loop is slower and darker. Google gives you Search Console, a sitemap ping, an index coverage report and a rough sense of when a change has been processed. No answer engine offers any of that. You cannot submit a URL to ChatGPT, there is no coverage report, and the only way to know whether the new pages have been picked up is to ask the engines questions and see who gets named.

Most teams discover this after launch, because the launch checklist was written for search. Rankings hold, analytics look normal, and then a salesperson mentions that the AI answer that used to name the company now names a competitor. Nothing in the redesign was wrong in SEO terms. The page that carried the quotable answer was reorganised into three tabs and the sentence no longer exists.

None of this is an argument against redesigning. It is an argument for treating AI citations as an asset the project is responsible for, alongside rankings and conversion rate, with a baseline taken before anything ships. The rest of this page is what that looks like in practice.

Why do AI citations break more easily than Google rankings?

AI citations break easily because they rest on a URL that a crawler fetched once, on a schedule nobody controls, and the crawlers are already working from stale lists. The Vercel and MERJ analysis of AI crawler behaviour across the Vercel network, published in December 2024, found that 34.82% of ChatGPT crawler fetches and 34.16% of Claude fetches returned 404, against 8.22% for Googlebot. A third of their crawl budget was already being spent on URLs that no longer existed before you moved anything.

Citations are also short-lived by default, which means a broken one is rarely re-earned quickly. Trakkr Research, running identical prompts daily across eight models, logged 108,650 citation URLs across 10,991 brands and found that 73.5% of cited URLs appeared exactly once and never returned, with the median brand falling to half its peak citation count in 31 days. A separate study by Scrunch with Stacker, covering 3.5 million citation events from September 2025 to March 2026, put the median half-life of a source at about 4.5 weeks.

Put those two findings together and the migration risk is obvious. Your citations are decaying anyway, on a clock measured in weeks, and a redesign removes the thing they decay towards. The replacement page has to be found, fetched, parsed and judged relevant again, in a system that spends a third of its fetches on dead URLs and tells you nothing about its progress.

There is one more asymmetry worth naming. Google evaluates a redirect and transfers signals to the new URL as a documented part of how its index works. An answer engine also has to retrieve the new page, and separately the model has to see enough of it to associate your brand with the question again. Retrieval and association are different mechanisms, which is why a page can be crawled promptly after a migration and still be absent from answers for weeks, as our page on how AI engines choose sources sets out.

What actually breaks during a website migration?

What breaks is rarely the whole site. Specific changes cause specific losses, and knowing which is which lets you spend the project budget on the four that matter instead of on a general anxiety about the launch. The table below is the list we work through on a pre-migration review.

Change	What it costs you in AI answers	What to do instead
URL structure changes	Every existing citation points at a dead address	One-to-one redirects, single hop, mapped page by page
Copy rewritten in brand voice	The extractable answer sentence disappears	Keep the answer sentences and the question wording
Content split into tabs or accordions loaded on click	Text is no longer in the fetched HTML	Server-render the content, hide with CSS if you must
Move to a client-rendered framework	Pages return an empty shell to crawlers that do not render	Server-side rendering or static generation for content pages
FAQ and schema dropped in the rebuild	Question-shaped blocks and their markup are gone	Port the FAQ blocks and the JSON-LD with the page
Pages consolidated into one longer page	Several cited URLs collapse into one, most citations lost	Consolidate only where the old page earned nothing

The second row is the one designers and copywriters argue about, and it deserves a straight answer. A sentence that reads as flat and literal on a new homepage, such as a plain definition of what the company does, is often the exact sentence an engine has been quoting for a year. Rewriting it into something more evocative removes the quote and replaces it with prose that answers nothing.

The fourth row is the most expensive and the least visible, because the site still looks perfect in a browser. Our page on whether AI crawlers read JavaScript covers the mechanism and the one minute test that catches it before launch rather than after.

What should you do before the new site goes live?

Before launch, take a baseline you can compare against, because without one you will never know what the migration cost. Write down 20 to 40 buyer questions in the wording people actually use, run them across the engines that matter to you, and record which URLs are cited and for which questions. That file is the only evidence you will have that a page used to earn something.

Then pull the list of pages that carry citations and mark them as protected. Protected means the URL does not change, or changes with a one-to-one redirect; the answer sentences survive the rewrite; the headings keep their question wording; and any FAQ block and structured data comes across with the page. Everything else on the site can be redesigned freely, which on most sites is nearly all of it.

Check the new build against the crawler before it ships, on the staging environment if it is reachable, or on the first day if it is not. Fetch a protected page, read the raw HTML that comes back, and confirm the answer text is present, the headings are real headings and the structured data is there. The full sequence is in our guide to checking whether AI can read your website.

One thing not to do: do not treat the redirect map as the whole job. A redirect keeps the address alive and does nothing about the content. If the page it lands on has been rewritten into brand copy with the specific claims removed, you have preserved the URL and lost the reason anyone cited it, which is a worse outcome than a 404 because it looks like success.

How long does it take to get cited again after a migration?

Recovery takes weeks rather than days, and nobody can give you a reliable number because the engines publish no crawl schedule. What the evidence supports is a range and a shape: the new URLs have to be fetched, which happens on the crawler's own cadence, and then the association between your brand and the question has to re-form, which is slower than the fetch.

Use the decay research to set the expectation rather than a vendor promise. With a median brand half-life of 31 days in the Trakkr data and a source half-life of roughly 4.5 weeks in the Scrunch and Stacker study, a page that goes dark for a month is competing against its own decay curve when it comes back. Plan for a quarter to return to the previous baseline, and treat anything faster as a good surprise.

Watch the right signal while you wait. Server logs tell you whether GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot are hitting the new URLs and what status codes they get, which is the earliest evidence available to you. Citation checks on your prompt set tell you whether the answers have changed, and those move later. Watching only the answers means a month of no information.

Re-run the prompt set on a schedule rather than when someone asks, because citation counts move week to week regardless of what you did. Trakkr found a typical brand's week-over-week citation count swings by 51.8%, which means one week of silence after a migration proves nothing and six weeks of a downward line is a real problem. The mechanics are in our page on measuring AI citation decay.

How do you protect citations when URLs have to change?

Protect citations by treating the old URL, the old wording and the old structure as three separate things to carry forward. Redirect each old address to the single page that best answers the same question, in one hop, with no chains through intermediate URLs. A chain that resolves fine in a browser costs fetches in a crawl budget that is already being wasted at a third on dead URLs.

Carry the wording next. Keep the headings that are phrased as buyer questions, keep the first sentence under each heading as a direct answer, and keep the specific figures with their sources attached. The Discovered Labs 2026 analysis of more than 2 million citations found vocabulary alignment between page and query to be the only page-level signal that survived domain-level controls, at an effect size of 0.37, so the words you are tempted to smooth out are the asset, as our methodology page explains.

Then carry the structure. Sections that stand alone, a summary block near the top, comparison tables and FAQ blocks are the shapes engines lift, and the ConvertMate GEO Benchmark study of 8,000 domains found 68.7% of cited pages used a logical H1 to H2 to H3 hierarchy. A rebuild that flattens headings into styled divs removes the thing that made the page parseable.

Finally, tell the machines the new shape of the site. An updated XML sitemap with accurate last-modified dates, internal links pointing at the new URLs rather than through redirects, and consistent linking from your own pages all help the new addresses get found. None of it is a submission mechanism, because no submission mechanism exists, which is covered in our piece on whether llms.txt does anything.

Can a redesign ever improve your AI visibility?

A redesign can improve AI visibility substantially, and the projects that gain are the ones that fix a retrieval problem while they are in there. Moving content pages from client-side rendering to server-side rendering or static generation is the biggest single win available, because it turns pages that returned an empty shell into pages an engine can read at all.

Structure is the second win. A rebuild is the cheapest moment to introduce question-shaped headings, self-contained sections of 200 to 400 words, a summary block at the top of every long page, comparison tables where there are options, and FAQ blocks with matching structured data. Those are template decisions, so doing them once in the design system applies them to every page you publish afterwards.

The third win is removing accidental walls. Documentation behind a login, specifications inside a PDF, pricing detail in an interactive widget, and anything loaded only after a click are all invisible to retrieval, and a migration is when those decisions are open for discussion. Our page on whether gated content gets cited by AI covers which of them are worth keeping.

The pattern in projects that come out ahead is simple enough to put in a brief. Protect what already earns citations, fix what was blocking retrieval, and take the freedom to redesign everything else. Projects that come out behind are the ones where nobody knew which pages earned anything, so everything was equally negotiable.

What else do people ask about AI crawlers during a migration?