Every decent migration checklist covers the same ground, ours included. Crawl the old site. Map redirects one to one. Baseline rankings and traffic. Keep the staging environment out of the index. Watch Search Console for two weeks after launch and compare against what you recorded before.
That last part is the whole point. A migration baseline exists so that when traffic moves, you can prove whether you caused it. Without one you are guessing, and guessing during a post-migration panic is how six-week problems become eighteen-month recoveries.
The checklists are still good. They're just now incomplete, because there is a second visibility layer that migrations break in their own specific ways, and almost nobody is baselining it.
Until June 2026 there was a decent excuse: you couldn't measure it. That excuse is gone.
What migrations break in the answer layer
Classic search recovers from migration damage through a mechanism we understand. Redirects pass signals, Google recrawls, rankings settle. Slow, but predictable, and diagnosable when it stalls.
AI citations don't work the same way, for three reasons.
Different engines, different indexes, different refresh rates. Perplexity runs its own crawler. OpenAI runs its own crawler layer. Google's AI features run off Google's index. A migration that Google reprocesses in three weeks may take considerably longer to propagate everywhere else, and none of those refresh schedules are published or predictable.
Worth flagging one piece of received wisdom here, because it's still repeated everywhere and it's no longer safe: the claim that ChatGPT's browsing simply runs on Bing's index. It was a reasonable description in 2023–24. Since then, independent tests have repeatedly found ChatGPT citing pages that appear in Google's index and not Bing's, and nobody outside OpenAI can currently say what the retrieval stack looks like. Plan for opacity rather than for a specific dependency.
Citations attach to passages, not domains. Classic ranking evaluates the domain: authority, links, history, all of which a clean redirect substantially preserves. Generative retrieval evaluates the specific block of text that answers the question. Rewrite the page during the migration, as people almost always do, and you may have removed the exact passage that was being cited, even though the URL redirects perfectly and the topic is unchanged.
Bot access gets rebuilt from scratch. New host, new robots.txt, new CDN configuration, new WAF rules. This is where most of the silent damage happens, and it's covered in detail below.
The technical failure that nobody catches
If you take one thing from this: a migration is the single most likely moment for you to accidentally block AI crawlers, and nothing in your standard post-launch checks will tell you.
Rankings hold. Search Console looks fine. Googlebot is allowed because everyone remembers Googlebot. Meanwhile GPTBot, OAI-SearchBot, PerplexityBot and ClaudeBot are being refused by a robots.txt that got rebuilt from a default template, or a hosting provider's boilerplate bot rules, or a security configuration set by whoever handled the DNS. Worth being clear that llms.txt does not fix this: it is not an access-control file, and shipping one on the new domain does nothing if robots.txt is refusing the crawler.
Cloudflare made this both better and more dangerous. Since 1 July 2026 its bot controls split into three categories (Search, Agent, and Training), replacing the old blanket "Block AI Bots" switch. The granularity is genuinely useful: you can permit retrieval for search and answers while declining to feed model training. But it also means there is now a specific, easy-to-get-wrong setting that removes you from AI answers while leaving your Google rankings entirely intact.
Put a second date in the diary too. From 15 September 2026, Cloudflare applies new defaults on ad-monetised pages: Training and Agent blocked, Search allowed. If your migration lands near that date and the site runs ads, you will have two changes to your bot posture in the same window and no way to tell them apart afterwards unless you recorded the before state.
Set either wrong and the symptom appears weeks later as a slow decline in a channel you weren't monitoring, with no obvious cause and no alert.
The baseline: what to record before you touch anything
Four things, none of which take long.
1. Search Console generative AI impressions. The Search Generative AI performance reports have been live since 3 June 2026, one for AI Overviews and AI Mode, a separate one for Discover. Export impressions by page, by country, by device, for a clean 28-day window before the migration. This is your control. It doesn't give you clicks (see our guide on filling that gap), but as a before-and-after comparison it's exactly the right shape.
2. GA4 AI Assistants sessions by landing page. Live since around 13 May 2026, covering third-party assistants: Google names ChatGPT, Gemini, DeepSeek, Copilot and Grok, and explicitly excludes its own AI Overviews and AI Mode. Export the same window. Add a Perplexity custom channel group first if you haven't, since it isn't among the named sources and in practice still lands in Referral.
3. Bing indexation and Citation Share. Bing Webmaster Tools, plus the AI visibility insights added on 16 June 2026: Intents, Topics, Citation Share and Compare. Record which of your key pages are actually indexed. Bing's post-migration recrawl is consistently slower than Google's, so this is the layer most likely to lag, and Citation Share gives you a clean before-and-after for Copilot visibility specifically.
4. A manual citation spot-check. Pick your fifteen most commercially important queries. Run them through ChatGPT, Perplexity, Gemini and Claude. Screenshot what gets cited. It's crude and it doesn't scale, but it's the only record you'll have of which passage was being quoted, and that's the thing that tells you whether a rewrite broke something.
Post-launch checks, and when to run them
| Check | Timing | Looking for |
|---|---|---|
| robots.txt on live domain | Day 0, before DNS propagates fully | GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot all permitted |
| CDN / WAF bot rules | Day 0 | Cloudflare Search and Agent categories allowed; Training is your call |
| Bing sitemap submitted | Day 1 | Confirm accepted, not just uploaded |
| GSC AI impressions | Weekly to week 8 | Decline steeper than classic Search decline |
| GA4 AI Assistants sessions | Weekly to week 8 | Same, on the referral side |
| Bing indexation of key pages | Week 2 and week 6 | Coverage gaps versus baseline |
| Manual citation spot-check | Week 4 | Same fifteen queries, compared to screenshots |
The signal to watch for is divergence. If AI visibility and classic Search decline together, you have a normal migration dip and the standard playbook applies, which is the ground covered in diagnosing a traffic drop after a migration. If AI visibility falls while classic Search holds, the cause is almost always access or passage-level, and both are fixable quickly once identified.
One content-side thing while you're in there
If your migration includes a content rebuild, two notes.
FAQ rich results stopped appearing in Google Search on 7 May 2026, with the Search Console filter and Rich Results Test support removed the following month and API support ending this month. Existing FAQPage markup is harmless but inert. If you're carrying it forward purely for the rich result, it's dead weight in Search, though the structured Q&A format itself still helps generative retrieval, so this is a reason to reconsider the markup rather than abandon the content pattern. Our guide to FAQ and HowTo schema covers what that deprecation did and did not change.
And be deliberate about rewrites. If a page is currently earning citations, the specific sentences doing that work are an asset. Migrating the URL correctly while rewriting the passage away is a real way to lose visibility you had, and it will look like a technical problem when it isn't. Your week-four spot-check against the baseline screenshots is what catches it.
The point
The reason we baseline before a migration has never been the numbers themselves. It's so that when something moves, you can tell the difference between "we broke it" and "it was already moving." That's what turns a panic into a diagnosis.
There are now two visibility layers, they fail in different ways, and only one of them is on your checklist. Adding the second costs maybe an hour before launch. Not having it costs you the ability to diagnose an entire channel for as long as the problem persists, which, given how few people are monitoring this, tends to be a while.
Sources
- Google Search Central, "Introducing Search Generative AI performance reports in Search Console," 3 June 2026
- Google Analytics Help, AI Assistants default channel definition, May 2026
- Bing Webmaster Blog, "New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare," June 2026
- Cloudflare, "Content Independence Day" AI traffic options, 1 July 2026, and September 2026 default changes
- Google FAQ rich result deprecation, May 2026