Redesigning without losing your AI citations

The citations you already have are an asset nobody puts on the redesign checklist. You lose them the same two ways you lose rankings: URLs that change without a redirect, and the deletion of the exact text that got quoted. Baseline what you have before anyone opens Figma, preserve the URLs, and keep the quoted passages as text.

Oskar Mieta

Founder, Designer & Developer

Summarize with AI

Category

Measurement

Reading time

6 min

The citations you already have are an asset nobody puts on the redesign checklist. You lose them the same two ways you lose rankings: URLs that change without a redirect, and the deletion of the exact text that got quoted. Baseline what you have before anyone opens Figma, preserve the URLs, and keep the quoted passages as text.

An AI citation is a link to your page inside a generated answer. The source card under a Google AI Overview, the numbered footnote in ChatGPT or Perplexity. It lives in a different report than your rankings do, which is most of why it falls off the plan.

Desses is a design and development studio in Rotterdam and website design is one of the things we sell. Read the rest knowing where my money comes from. Plenty of the sites reading this have nothing worth protecting, and I'll say which ones.

Baseline before anyone touches the design

Four instruments, all four on the same day, before the first wireframe.

Instrument

What it tells you

What it can't tell you

Search Console, generative AI performance report

Impressions by page, country, device and date

No query dimension, no position, no click metric. 1,000-row cap

Bing Webmaster Tools, AI Performance report

Total citations, cited pages per day, page-level activity, grounding queries

No ranking, no prominence, no clicks

Server logs, filtered to verified bot IPs

Which URLs OAI-SearchBot, Claude-SearchBot, PerplexityBot and Googlebot fetched

Nothing about whether a fetch produced a citation

A saved prompt set, run and recorded

What the assistants say about your category today, and whether you appear

Sampled prompts, never real user queries

Google announced the generative AI performance reports on 3 June 2026, global by 31 August 2026. Export the page-level impressions and keep the file. Verify server-log bots by IP, since user agents get spoofed and Anthropic publishes its ranges.

For the prompt set, write down the twenty questions a buyer asks an assistant before finding you, run each one seven or eight times, and save the answers. Schulte et al. measured a daily Jaccard around 0.35 inside a single engine: about 35% of yesterday's cited sources are cited again today. One run tells you nothing. The full setup is article 11.

Bing's grounding queries are the best redesign brief you'll get

Grounding queries are, in Bing's own label, sample query phrases AI systems used to retrieve and cite publisher content. Microsoft shipped them in the Bing Webmaster Tools AI Performance report on 10 February 2026. It's the only place any provider shows what was asked before your page got used. No third-party tracker reconstructs it: those tools sample prompts they wrote themselves.

Read that list as a design input and it changes shape. It's an inventory of the questions your site answers well enough that a machine picked it up and handed it to somebody. Half will be questions nobody chose to answer. Put the wireframe next to it, and if the new architecture drops the page carrying four of them, that's a decision made with open eyes. Finding out in week nine is the version that hurts.

The caveats travel with it. Bing reports citation frequency only, and it's Copilot's grounding, directional elsewhere.

URLs are cheap to keep. Expensive to get back

Preserve the URL first, and redirect only what you couldn't preserve. A URL that survives untouched is one you never think about again. Every redirect is a rule somebody maintains and eventually forgets to test.

Google's AI Overviews and AI Mode are grounded in the Google Search index, so a redesign that holds its rankings holds most of its AI Overview citations. It holds much less of ChatGPT's and Perplexity's, because those assistants fetch the URL live. How much ranking actually predicts citation is its own argument, and the honest answer sits between two studies rather than on either one.

The Framer sub-path trap

Changing a CMS slug in Framer creates no redirect. Framer's documentation says it plainly: "If you change a sub-path from the canvas or CMS, it won't automatically update in your redirect settings." The old URL starts returning a 404, nothing warns you, and the CMS item looks healthy in the editor.

This is the most common way a Framer redesign drops a cited page. Somebody tidies /insights/ai-search-b2b-2025 into /insights/ai-search-b2b during a content pass and nobody writes the redirect.

Framer's redirects permanently route traffic from an old path to a new one, support wildcards and capture groups, and start at the Pro plan. Basic has none, which makes a Basic-plan redesign a URL-preservation exercise with no net. Framer generates and maintains sitemap.xml, so pull the URL list out of it before you start and diff it after launch.

Keep the text that got quoted as text

A page can survive a redesign while the paragraph that earned the citation doesn't.

Google confirmed it ranks passages inside a document, rolled out in February 2021. Martin Splitt: pages are "considered as one document with multiple annotations and in ranking can be scored independently". Above that sits query fan-out, Google's mechanism for AI Overviews and AI Mode, issuing multiple related searches across subtopics.

The thing at risk is smaller than a page. Four redesign moves delete a passage and leave the page standing: copy set inside an image, three paragraphs folded behind a tab, a section hidden under a scroll reveal, and the plain explainer paragraph cut for a tighter hero. Why each of those costs you is its own article.

The rule is short. Any passage in a Bing grounding query, or on a page with real generative AI impressions, moves to the new design as visible HTML text or gets deleted on purpose, with a name next to it.

Re-measure at 2, 6 and 12 weeks

When

What you check

Normal

A real problem

Week 2

Server logs and crawl errors

Bot hits redistributing across the new URLs, scattered 404s in the long tail

404s on the pages that carried citations, or total bot volume down across the whole site

Week 6

GSC generative AI impressions by page, Bing citation counts

Individual pages down, roughly flat at site level

Site-level impressions down and still falling

Week 12

The saved prompt set, re-run seven to eight times per prompt

Back at or above baseline

Still absent from answers you were present in before

Week two catches the expensive mistake while it's still cheap. Week six is noisy: Google gives impressions with no queries and no clicks, and page-level movement inside a flat total is mostly the index reshuffling. Twelve weeks is where you make a call. If the pages came back and the citations didn't, another pass over the copy won't find the answer.

Your baseline moves while you measure against it

Seer Interactive tracked 53 brands across 5,471,127 queries from January 2025 to February 2026. Organic CTR on informational queries averaged 3.35% with no AI Overview present, 2.07% when the brand was cited and 0.94% when it wasn't, all three full-year 2025 averages. The same study tracks the uncited number climbing from a floor of 1.3% in December 2025 to 2.4% in February 2026.

Read that before you blame your redesign. A dip measured against a 2025 annual average is compared to a market that has already moved, and on this measure it moved up. Save the raw monthly numbers and compare month against month.

When the citations aren't worth protecting

If Search Console shows a handful of generative AI impressions and Bing shows no citations, there's nothing here to defend. Everything above costs a week better spent on the site itself. Most B2B sites are there. Skip the baseline, ship, measure after launch.

A clean break is the right call more often than the preservation instinct suggests. A pivot to a different buyer. A blog archive built for a keyword strategy nobody believes in. Holding those URLs alive points a machine at an answer you no longer want to give.

The citations worth losing are the ones on questions you can't sell against. A citation on a question that never converts is a small number getting smaller.

We turn down Webflow rebuilds, for what it's worth. Nobody here is going to tell you the answer is a migration.

What we do on a redesign

Export the sitemap URL list, save the Search Console impressions per page, pull the Bing grounding queries, take 90 days of server logs filtered to verified bot IPs, and run the prompt set. Day one, before kickoff.

Wireframe against the grounding-query list. Every question gets an owner: a page in the new site, or a line in the retire column with a reason. Build the URL map before any visual design is signed off, three columns, keep, redirect and retire. On Framer that means Pro at minimum.

Ship, then run the two, six and twelve week checks against the saved baseline. Whoever ran the baseline runs the re-measurement, because the prompt set only works if it's the same one.

If you want that run end to end, our website design work covers it, from $8,000 (€7,000) over three to four weeks. Paired with the Framer build it runs four to six weeks against seven in sequence.

If you have a redesign scheduled, export your Search Console generative AI impressions by page today. Takes five minutes and you can't get it back later. Send me the list and I'll tell you which pages I'd protect.

Questions people ask about this

Will a redesign cost me my AI citations?

It costs you the ones attached to URLs that change without a redirect and to passages that get deleted or hidden. Pages that keep their URL and their text usually keep their citations. Ranking is the variable nobody controls, and AI Overviews are grounded in Google's index, so a redesign that damages rankings damages citations too.

How long should I wait before worrying?

Check server logs at two weeks, generative AI impressions at six, and re-run your prompt set at twelve. Page-level movement inside a flat site-level total in week six is normal. Site-level impressions still falling at week twelve is a problem worth money.

Can I see which prompts caused my site to be cited?

Partly. Bing Webmaster Tools shows grounding queries, its own label for the sample query phrases AI systems used to retrieve and cite publisher content, in a beta released 10 February 2026. Google's reports carry no query dimension. Nothing exists for ChatGPT, Claude or Perplexity.

Do I need to keep the old URL if the content moved to a new page?

Redirect it. A Framer redirect permanently routes traffic from the old path to the page now carrying the content, and preserves most of what the old URL earned. A 404 drops the page from live-fetch assistants immediately and from the index over the following weeks.

Does a rebrand hurt more than a redesign?

It has a longer tail, and a rename resets the half that was winning. Amsive, across about 10,000 AI Overview-triggering keywords, found overall CTR down 15.49%, branded keywords up 18.68%, non-branded down 19.98% and keywords outside the top three down 27.04%. A new name starts from zero on the branded side. Ahrefs, across 75,000 brands, found branded mentions correlating with AI visibility far more strongly than link metrics, correlation-not-causation attached. Bing's February 2026 guidelines ask for clear, consistent entity names, which a rebrand breaks on purpose. Keep both names in visible text on one URL and give it quarters.

Questions people ask about this

Will a redesign cost me my AI citations?

How long should I wait before worrying?

Can I see which prompts caused my site to be cited?

Do I need to keep the old URL if the content moved to a new page?

Does a rebrand hurt more than a redesign?