← Insights

AEO

Website Redesigns That Break AI Visibility

November 16, 2026 · 6 min read · AEO

A redesign can quietly erase the AI citations you already have. Here is what to inventory first and the five checks that should block your launch.

Website Redesigns That Break AI Visibility

You can rebuild your website without losing search and AI visibility. You just cannot do it by treating visibility as something the team cleans up after launch, because by then the damage is already in the crawl record.

Redesigns are the single most common cause of a sudden drop in AI citations. Not because the new site is worse, but because the four things AI engines depend on, stable URLs, server-rendered text, structured data and plain-language answer blocks, are exactly the four things a rebuild touches hardest. Design systems get rewritten in JavaScript components. URL structure gets "cleaned up." The old FAQ block gets replaced with an accordion. The schema never makes it across. Every one of those decisions is defensible on its own. Together they take the pages an answer engine was quoting and turn them into something it can no longer read.

Why traffic holds for two weeks and then the non-brand queries go quiet

The failure pattern is consistent enough to predict.

Launch week looks fine. Sessions hold. Brand search is untouched, because people typing your name still find you. Leadership sees a flat line and signs off on the migration as clean.

Then weeks three through six happen. Non-brand queries soften. Pages that used to get pulled into AI answers stop appearing. Nothing on the analytics dashboard says "citation lost," so the decline gets blamed on seasonality or an algorithm update.

What actually happened is that crawlers reconciled. The content an engine was quoting no longer exists in retrievable form, either because the URL moved, the text moved into a client-rendered component, or the paragraph that did the answering was rewritten into something shorter and more "on brand."

Two realities make this worse than a traditional migration hangover.

The first: the major AI crawlers do not execute JavaScript. Vercel and MERJ's December 2024 analysis of over a billion crawler requests found that the OpenAI, Anthropic, Perplexity, Meta and ByteDance crawlers do not render JavaScript, with ChatGPT and Claude explicitly verified as not executing it. Gemini runs on Googlebot's infrastructure and does render JavaScript, and AppleBot renders through a browser-based crawler. When they fetch a client-rendered page, they get the initial HTML shell with a root div and script tags. That is all they ever see. Googlebot renders JavaScript eventually. The assumption that if Google can index it, AI search can read it, is simply false. A headless rebuild can rank fine and be an empty room to ChatGPT, Claude and Perplexity.

The second: AI crawlers are already bad at stale URLs before you change anything. In the same December 2024 study, ChatGPT spent 34.82% of its fetches on 404 pages and Claude 34.16%, against Googlebot at 8.22%. ChatGPT spent another 14.36% of fetches following redirects, versus Googlebot at 1.49%. A third of those requests are already hitting dead ends. Ship a redesign with an incomplete redirect map and you are handing a crawler that already fails constantly a fresh set of dead URLs, and it will not come back on Googlebot's patient schedule.

This is not a fringe audience. Across Vercel's network in that month, Googlebot made 4.5 billion fetches across Gemini and Search, GPTBot 569 million, Claude 370 million, AppleBot 314 million and PerplexityBot 24.4 million. The four AI crawlers together came to nearly 1.3 billion fetches, a little over 28% of Googlebot's volume.

And recovering your rankings is no longer the same job as recovering your citations. Ahrefs data as reported, across 863,000 keywords and 4 million AI Overview URLs, shows the share of AI Overview citations coming from top-ten ranked pages fell from 76% in July 2025 to 38% by March 2026. Other analyses put the overlap higher: a June 2026 CiteLens study of 320 buyer queries found Google's AI Mode drew 93% of its citations from Google's top ten and Perplexity 89%, though ChatGPT only 30%. The exact magnitude is contested. The direction is not. Getting your rankings back does not automatically get your citations back. You are now running two recoveries.

Inventory before a single screen gets designed

Most redesign briefs start with a sitemap and a moodboard. Start with evidence instead.

Before design begins, capture four things:

  • Which pages are currently cited or ranked. Not just organic winners. Which URLs are being surfaced in AI answers, and for which prompts. If nobody has measured this, you are about to rebuild blind.
  • The exact URLs. Full strings, including trailing slashes, parameters and pagination patterns. This becomes the redirect map, and it is built before the new site exists, not after.
  • The schema on every important page. Organization, Article, FAQPage, Product, Breadcrumb, whatever is there. Screenshot the markup. "We will add schema later" is how parity gets lost.
  • The specific paragraphs doing the answering. This is the one teams skip. Go into the page and find the two or three plain-text passages that directly answer a question. Those are the sentences an engine quotes. Copy them into the inventory verbatim, so a copywriter rewriting for the new design knows which words are load-bearing.

That last one is the difference between a migration audit and a visibility audit. You are not protecting pages. You are protecting retrievable answers.

The five checks that should block a launch

Treat these as launch gates. If one fails, the date moves.

  1. Server-rendered text. Disable JavaScript and load every template: homepage, service page, article, product, collection. If the body copy disappears, AI crawlers see nothing. Fix the rendering, not the content.
  2. Preserved URL and redirect map. Every old URL resolves to its closest new equivalent with a permanent redirect. No chains, no mass redirects to the homepage. Tested before the DNS switch, not after.
  3. Schema parity. The new page carries at least the structured data the old page carried. Anything you drop, you drop on purpose and in writing.
  4. Answer and comparison blocks intact. FAQ sections, comparison tables, definition paragraphs and spec lists survive in plain text. If the new design hides them behind tabs or accordions, confirm the content is still in the served HTML.
  5. Crawl access. Robots and firewall rules do not block AI crawlers by accident. Staging-environment blocking rules have a habit of riding along to production. Decide which crawlers you want, then verify the live file matches that decision.

Replatforming a store adds product-level risk

Online store migrations fail in a different place. The redirect map usually gets attention because revenue is attached to it. Product data does not.

The parts of an ecommerce site that answer engines actually source from are the parts nobody assigns an owner during a replatform: product descriptions with real specifications, review content and its markup, and collection pages that function as category answers. Teams migrate the catalogue as data, push descriptions through a template, lose the long-form spec copy, and launch with review history disconnected. The SKUs survive. The quotable content does not.

Three rules for a store rebuild. Migrate product copy as content, not as a database field. Keep review markup live and attached to the same product URLs. Treat collection pages as editorial pages, because a thin collection page is invisible to a shopping answer.

The first 30 days

Monitoring starts the day you launch, not the week someone notices a dip.

  • Days 1 to 3: crawl the live site with JavaScript off. Spot-check redirects from your inventory list. Confirm the sitemap submitted and the robots file is the one you approved.
  • Days 4 to 14: watch server logs for AI crawler 404s specifically. This is your earliest warning, and it moves before any traffic metric does.
  • Days 15 to 30: track non-brand queries and AI citations against your pre-launch baseline, prompt by prompt. Brand traffic will lie to you. Non-brand will not.

Be honest about what comes back. Fix a broken redirect or restore server-rendered text and crawlers can re-reconcile. But content you deleted does not get re-cited, because there is nothing to cite. A citation is earned by a specific passage on a specific URL. Remove the passage and the recovery is not a wait, it is a rebuild.

We build and rebuild websites and online stores, and we run AI visibility tracking against the same properties through BeFound AI, which means we see the before and after on the same site rather than guessing at it. If you have a redesign or replatform coming, the cheapest time to protect your visibility is before the design starts.

Frequently asked questions

Will a website redesign hurt our search and AI visibility?

It can, and the risk is highest when visibility is handled after launch instead of before design starts. The common causes are changed URLs without a complete redirect map, body copy moved into client-rendered JavaScript components, dropped schema, and deleted answer paragraphs. Each is preventable if you inventory what is currently cited and ranked before a single screen is designed.

Why did our traffic look fine for two weeks after launch and then drop?

Brand search holds up after a migration because people typing your name still find you, so the overall line looks flat. The decline shows up in non-brand queries and AI citations once crawlers reconcile the new site, usually in weeks three through six. Track non-brand performance against a pre-launch baseline rather than total sessions.

If Google can index our new site, can ChatGPT and Perplexity read it too?

No, that assumption does not hold. Vercel's analysis of over a billion crawler requests found the major AI crawlers do not render JavaScript, with ChatGPT and Claude verified as not executing it, so a client-rendered page returns only the HTML shell. Googlebot renders JavaScript eventually, which is why a site can rank normally and still be unreadable to AI engines.

What makes replatforming an online store riskier than a normal redesign?

Product-level content is where shopping answers get sourced, and it is usually the part without an owner during a migration. Product descriptions get pushed through templates and lose their specifications, review markup gets disconnected, and collection pages ship thin. Migrate product copy as content, keep review markup attached to the same product URLs, and treat collection pages as editorial.

How long does recovery take after a redesign damages visibility?

It depends on what broke. Technical faults like missing redirects or client-rendered text can be fixed and re-reconciled by crawlers once corrected. Content you deleted is different, because a citation is earned by a specific passage on a specific URL, so if the passage is gone there is nothing to cite and the work becomes a rebuild rather than a wait.

Chat with Emma
Chat with Emma