Because most of a migration is not building. It is inventory work. Someone has to list every old URL, decide where each one goes, move the content, and check that nothing broke. On a site with a thousand pages that is thousands of small decisions, and the building part is almost the easy bit.
This is exactly the shape of work language models are good at. Repetitive, text heavy, pattern based, and tedious enough that humans start skimming halfway through.
So the promise is real. The catch is that a migration is also where mistakes are most expensive, because a bad redirect map costs you rankings you spent years earning.
AI is genuinely useful for matching old URLs to new ones, drafting meta titles and descriptions at volume, cleaning messy exported content, and spotting gaps between two lists of pages. All of these are pattern matching over text. None of them require judgement about your business strategy.
The common thread is that a human can check the output quickly. You can scan a redirect map and see that a blog post is pointing at a pricing page. That is a thirty second check on work that would have taken an hour.
The parts AI is bad at are the ones where checking takes as long as doing. Deciding which pages deserve to survive the migration is a business call, and no model knows which of your service pages the sales team actually uses.
We think of it as a fast intern with no context. Brilliant at volume, worth trusting only where you can verify the answer at a glance.
Export the old URL list and the new URL list, then ask a model to match them and flag anything it is unsure about. The flagging matters more than the matching. A model that returns 900 confident matches and 100 marked as uncertain gives you a real work list instead of a false sense of completeness.
Get the redirect type right before you start. Google's site move documentation says to use server side permanent redirects if technically possible, and recommends HTTP permanent redirects such as 301 and 308.
Google also tells you how long to keep them. Its documentation says to keep the redirects for as long as possible, generally at least one year, so Google can transfer all signals to the new URLs including recrawling and reassigning links from other sites.
The context windows now make this practical at real scale. Anthropic's model documentation lists Claude Opus 5, Claude Sonnet 5, and Claude Fable 5 with 1 million token context windows, which means a large URL inventory fits in a single pass rather than being chopped into batches that lose the thread. We covered the manual version of this work in our guide to setting up redirects without losing traffic.
For metadata, yes. Generating title tags and meta descriptions for 400 pages is a legitimate use, because each one is short, checkable, and follows an obvious pattern. For body content, we are far more cautious, and mostly we say no.
The reason is that a migration is not the moment to change what your pages say. You are already changing the URLs, the templates, and the internal links. Changing the words too means that when rankings move, you cannot tell which change caused it.
Where AI helps on body content is cleaning rather than writing. Exported HTML from an old CMS is usually full of inline styles, empty spans, and stray formatting. Asking a model to strip that back to clean semantic markup while preserving the words exactly is safe and saves days.
Anything the model does generate has to be checked against a source before it goes live, which is a whole discipline in itself. We wrote up our process for fact-checking AI-generated content separately.
Point it at the differences rather than the content. Give it the old sitemap and the new sitemap and ask what is missing. Give it the old page and the new page and ask what content disappeared. Difference finding is where a model earns its keep, because humans are terrible at it and machines are not.
We run a crawl of the staging site with Screaming Frog SEO Spider first, then feed the export into a model with a specific question rather than a vague one. Asking which pages lost their h1 gets a useful answer. Asking whether the site looks good does not.
The other high value check is internal links. Old body content is full of hardcoded links to old URLs, and those will chain through your redirects and slow everything down. A model reading the exported content can list every internal link that still points at an old path.
Do not skip the human pass on your top pages. We check the handful of pages that drive the most traffic by hand, every time, no matter how clean the automated output looks.
Never let it decide what to delete, never let it invent a redirect target for a page it cannot match, and never let it write the canonical rules. Those three decisions have consequences that only show up in search rankings weeks later, when the cause is hard to trace and the damage is already done.
The invented redirect is the dangerous one, because models want to be helpful. Asked to match a page with no equivalent, a model will often pick something vaguely related rather than saying it does not know.
The fix is in the instruction, not the model. We tell it explicitly to return an unmatched marker when confidence is low, and we treat a high unmatched count as a good sign rather than a failure.
Deleting is a business decision. A page with no traffic might still be the one your biggest client links to from their intranet. No model can know that, and no crawl will tell you.
We use a large context model for the URL matching and content cleaning, a crawler for the raw data, and Google Search Console for the truth after launch. The model choice matters less than the data you feed it and the instruction you give it.
On the model side, Anthropic's documentation lists Claude Opus 5 for complex agentic and enterprise work, Claude Sonnet 5 for the balance of speed and intelligence, and Claude Haiku 4.5 as the fastest option with a 200k token context window. For bulk metadata generation the faster model is usually the right call, because the task is simple and the volume is high.
On the data side, a crawler gives you what actually exists rather than what the CMS thinks exists. Those two lists differ more often than anyone expects, especially on an old WordPress site with a decade of plugins behind it.
If the destination is Webflow, the Webflow Data API is what turns a cleaned CSV into real CMS items without anyone pasting for a week.
You check Google Search Console, and you give it time. Google's own documentation says that as a general rule a small to medium sized website can take a few weeks for most pages to move, and that larger sites take longer. Panicking in week one is the standard mistake.
Google also asks you to submit the new sitemap in Search Console so it learns about the new URLs, and says you can remove the old sitemap at that point since Google will use the new one going forward.
If you moved domains, Google says to submit a Change of Address in Search Console for the old site. It also notes that if you are only moving from HTTP to HTTPS, you do not need the Change of Address tool at all.
Our full pre-launch and post-launch sequence sits in our guide to migrating a website without losing SEO, which covers the parts of this that have nothing to do with AI.
It makes them faster, which makes them riskier if you skip the checking. The danger is not that a model produces bad output. It is that it produces a lot of plausible output quickly, and volume discourages review in a way that a slow manual process does not.
We have a simple rule about this. If using AI on a task means we check less than we would have, we are using it wrong. The time it saves should go into more verification, not less.
Used that way it genuinely helps. The hours saved on mechanical matching become hours spent on the twenty pages that carry your revenue, which is where a migration is actually won or lost.
Before you touch a model, export a complete list of your current URLs and a list of what each one earns in traffic. Almost every migration failure we see traces back to an incomplete inventory rather than a bad tool. Get the list right and the rest gets much easier.
Then use AI on the matching and the metadata, keep the deletion and canonical decisions with a human, and hold your redirects for at least the year Google asks for.
If you are planning a migration and want a second pair of eyes on the plan before anything goes live, we are happy to walk through it. Reach us at phoenix.studio and tell us what you are moving from and to.
Tell us where you want to go. We'll tell you how we'd get you there.