In this episode, Bill Scully and Bill Hunt explore how website migrations in the AI era influence what language models learn, remember, and repeat about your business. They explain why redirects no longer guarantee continuity, how old content competes with new versions inside AI systems, and why mistakes made during website migrations can become difficult or impossible to fully correct.
The conversation highlights new risks for international websites, the rise of AI knowledge orphaning, and why website migration planning now requires governance, not just technical execution.
Key Takeaways
- Website migrations now carry new risks, as they can lock outdated information into AI systems.
- Search engines can re-rank content, but AI memory persists, so fixing SEO doesn’t fix narratives.
- Redirects preserve traffic, not understanding, leaving AI blind spots even after migrations.
- Every migration publishes a new version of your content, which AI and data sources may capture permanently.
- Old content competes with new, meaning outdated pages can still define your narrative.
- International websites face higher risk, as secondary markets can become the global “source of truth.”
- Migration failures are harder to detect, with AI citations missing or misrepresenting your brand despite normal traffic.
Listen now on your favorite platform:
Youtube: https://youtu.be/q8nRPWjZyvc?si=Ph4uLf1LW3Ttn57m
Spotify: https://open.spotify.com/episode/6nOm3ksRVcdxscltCuNrYv?si=lxUjDJ40S_ST5aIq0OChtA
Apple Podcast: https://podcasts.apple.com/us/podcast/website-migrations-in-the-ai-era/id1755789311?i=1000745781097
Show Note Links
Free International Web Mastery GSC Question Report:
https://lookerstudio.google.com/reporting/9fc44b52-5a7a-4fe7-8509-ac49e302c35d
Referenced Work: AI Share Button URL Creator (CiteMET) by Metehan Yesilyurt
Transcript
Bill Scully (00:01)
Welcome to episode 40 of the International Web Effectiveness Podcast, where Bill Hunt and I, Bill Scully, discuss the strategies and techniques necessary to improve your web presence and manage your global websites.
Today, we’re going to be discussing web migrations in an AI era. But before we get into that, please make sure you subscribe to the podcast so you never miss an episode. And if you do enjoy it, please tell someone about it and leave us a quick review. The reviews help motivate us to continue making more of these episodes.
In our last episode, we spoke with the fantastic guest Adam Gent from Indexing Insights, and we dug into the state of indexing and why it should be a critical part of your search efforts.
In this episode, we’re going to be talking again about the topic of website migrations for an AI era. And I have to say, when we talked about this topic, I thought this was going to be typical migration stuff, which a lot of people make mistakes with, and we’ve done lots of successful ones.
But this really got my brain going. So why don’t you tee this thing up and talk about how these differences in LLMs and search affect migrations?
Bill Hunt (01:37)
Sure. You know, we put this spin on it because, as you said, we both remember the days when migrations struck fear in the hearts of SEOs, because we probably found out about it after the fact or a couple of days before.
What’s interesting is that in all of this craziness of AI, I haven’t seen anything about migrations. And so I thought, wow, it’s odd that Bill wants to talk about migrations when nobody cares about them anymore. We’re all doing our thing.
So as we start thinking about whether there is anything different in the AI environment that we should think about, I think all the other things are true, right?
Are we able to transfer the equity that we built up on the old pages to the new pages? Can we get the pages reindexed? Can we recover any traffic we lose?
And that’s what you alluded to. A successful migration to us is when we don’t lose any of the value that we had before, and hopefully, through the refresh—whatever reason we’re doing a rebuild or migration—that it’s additive, that we actually gain.
So I started thinking about this, and it actually came out of some research I’ve been doing around why hreflang isn’t working, and we’ll circle back to that later. I think in our next episode, we’re going to talk about that specifically.
But one of these was if we’re launching this new content, the idea that it’s new, it’s fresh, it’s freshly being seen, it could, by the very nature of it being newer, having stuff that maybe isn’t in another market, impact our performance in a way we may not have even considered.
And what I mean by that is that before, when we deploy a new page, for whatever reason it’s a new page, we always monitor: did we lose rank and did we lose traffic?
And I think the same thing is true, but at this time, I think what I’m calling it is, are we losing the narrative?
Meaning, if we’ve materially changed—and we know if we materially change the content on a page, like when we went through the phase of content grazing, where we would take five paragraphs of content and the, you know, UX, the smarter UX people, the PhDs, would walk in and say, “Guys, this is better with five bullets because nobody can read anymore. We need five bullets instead of five paragraphs.”
So all the goodness that benefited us, that content went away. And we could point back to, hey, we lost the content that was either a good meta, a good description, a snippet.
And so I think the same thing happens now. So the question is: if we materially change our content, are we losing citation eligibility? Are we losing the very piece of content that, in a RAG model or in a memory model, was the very piece of content that made us stand out?
I think a second thing, too, is when we make a material change, are we now changing the data or the insight, and maybe what the LLM has stored is now old data?
So maybe old data is still being shown, because remember, and as we’ve tried to explain to people, there’s two points of ingestion, two points of action, two points of presentation.
One is in memory. And the first thing all SEOs ranted about with AI search was, “Look at the model. The last time it was updated was 18 months ago. So everything it has is 18 months old. Don’t use it. It’s outdated.”
Then the second piece is RAG, which is the new stuff. So when we say, “What’s the best running shoe for 2026?” clearly an 18-month-old data set is not going to have anything about 2026 running shoes. So that’s where RAG comes in.
So it can be the fact that it hasn’t updated on your new content. And so it’s a whole other problem. And so can the LLM get to this new stuff? Because remember, they’re not crawling like Google.
Bill Scully (06:26)
Right. Yeah, I like the line here. Search engines can recrawl and re-rank. Language models don’t work that way.
Bill Hunt (06:35)
Exactly. And so if we’re thinking about how this works in practice,
Bill Scully (06:37)
Right. I know you’ve done a lot of research on this, so let’s explore it further.
So let’s say that there’s a source of truth. There’s no need to do a fan-out. The old site doesn’t exist anymore, and you have redirects set up.
It’s going to be using the old content for its answers, do you know? I don’t know. Would they continue to get a link back to the old site?
If you have more of the transcript, you can paste it in and I’ll continue using this exact format.
Bill Hunt (07:26)
No. So I’m going to say possibly, because any crazy stuff can happen. If we think about it from an LLM perspective, the moment that cluster or chunk is parsed out, it loses, for all practical purposes, any connection back to a URL. It is now a block of fact, a block of information, especially when we’re talking about in-model responses.
So it loses any connection back to that. Now, if there’s a RAG thing—meaning it went out and fetched recently—or if it believes that this data is still valid, it will still answer the question. Like your point is, it will still cite the old content because it has it.
Bill Scully (07:54)
Okay.
Bill Hunt (08:24)
And then the question is, if it tries to fetch it because it knows this is a good place to go, from what I’ve seen, LLMs will not and do not follow redirects. It won’t follow them, because it didn’t necessarily store the source of where it got that information. It just puts it into its dataset, so it doesn’t necessarily have a link to follow.
It’s not going to represent it as Google may do. It’s either going to go out and try to fetch it. If it knows of a resource and that resource is no longer there, it’s not going to follow a redirect, go to the new one, and refetch it. It’s just not going to do that. I can’t find any evidence that it happens.
Bill Scully (09:10)
Right, right. So that makes it really important to get the first new page done right, because it may never pull that page information again.
Bill Hunt (09:29)
That’s true. And if there is a material change, like when we have an old product and a new product, the new product may have different material, different specs, different sizes, different features.
So if the query is nuanced, and especially if you put in a brand—like Lenovo for anything dealing with a computer—it’s going to go to Lenovo predominantly. So we’re not losing everything, but when there’s no brand affinity, when we’re asking a broader question where anything, any brand could be a result, that’s when we run into this.
Now, one thing I want to introduce, because this is the International Web Mastery, is: what market now is the source of truth? I want to really dig into this a little later, but that’s a whole nother thing. What page is the source of truth?
Now, if you launch a new market before—let’s say you launch, and there’s a case we’re going to talk about in our next episode where Norway updated their content—Google’s AI system fetched it. And because the query fan-out was looking for the newer technology, it actually used the grounding from Norway as opposed to the US.
This creates a whole new dynamic, which we’ll talk about a little bit later, because there are a couple of other nuances. But I want to circle back to this: redirects still make sense to do, but in RAG scenarios, from what I’ve seen and researched—I spent a couple of hours in all the different LLMs asking these questions, looking for examples from sites that had recently been updated—I couldn’t find any case where it followed a redirect.
Even when I was looking to compare a 2025 version of a product versus a 2026 version, everything I found showed a total disconnect, with the embeddings being disconnected from any live URL. These vectors were put into the database, and in a way, anything in the original model—those facts—aren’t necessarily tied to date and time.
An interesting thing I came across was the numerous mentions of how these systems build their corpora or their databases, and it was at the burden of Common Crawl. I can’t find any way to really affect Common Crawl, which is this massive crawl repository out there crawling. And I don’t know if you’ve seen some of the latest complaints from publishers; some have been frustrated because Common Crawl does not respect robots.txt. Cloudflare has been very vocal about this as well.
Bill Hunt (12:53)
They’re just out there crawling away like a rogue bot. And since a lot of these LLMs use Common Crawl, it’s an interesting dilemma. There’s no way to feed it, and it doesn’t necessarily use XML sitemaps.
I wrote another article—published it but didn’t promote it—that was a little bit geeky about my rant against Cloudflare. There has to be some hierarchy of whether or not we go get it. I think there was a guy who posted, he’s been writing a lot about LLMs lately. He wrote about context and content hierarchy.
Basically, he said: the number one thing that was cited in LLMs from his research was Facebook, then number two was Reddit, and number three was Wikipedia. He basically said these are highly influential sources. I sort of disagree because I think the context of the queries people are doing matters.
For example, what if I was a researcher doing something on health? I would never expect any one of those three to be the most cited. I’d expect WebMD or YouTube or something like that. Again, that’s just my view.
Bill Scully (14:18)
Yeah, that’s probably correlated to the type of queries, right? They’re more comparison queries, like “What’s the best pickleball paddle?” There might be a lot of people writing about that on Facebook.
Bill Hunt (14:25)
Yes, exactly.
So to get back to this redirect thing, one thing I would recommend—jumping ahead to recommendations for folks—is that we should look at grounding references as everybody is jumping on AI scoring tools, ranking tools, and prompt tools. We want to see, when in the grounding references, do the different LLMs swap out?
Just like we used to do in the old days. Remember the model we used to do for some mutual clients? We would redirect the old page to the new page, and we would use a report that you mastered in Data Studio. We would get the old URL that was ranking, and then we could take that keyword and see the new URL. We could load our old URL-to-new URL redirect model and basically see: not only is the new URL in, but did the old one get replaced by the new one?
I don’t know if that makes sense to people hearing it. I see the picture in my head: keyword, old URL ranking; keyword ranking today, same, better, worse. And then, did it get swapped out? Did the best-ranking URL actually get swapped out to where we wanted it to go?
Right, so we built that model. Interestingly enough, I couldn’t give it away to people. But it’s the same thing now. We want to see: did the grounding reference for something related to us post-migration change?
Bill Scully (16:32)
Right.
Bill Hunt (16:33)
Right. And that answers your question. Did the grounding reference—meaning, is Google or whatever LLM referencing the new reference over the old reference?
Bill Scully (16:48)
And you’d have to have a tool in order to do that unless you did it manually, which—I don’t think manually is realistic. Even not manually, I don’t know. I don’t have direct experience measuring all that stuff, but I’ve had lots of demos and I’ve been working with clients to try and pick a solution.
I don’t know. I mean, I think the tools will be great to see if the correct information is being displayed, but there is a lot of personalization going on.
Bill Hunt (17:33)
But think, let’s simplify it for people. Let’s say—I don’t know—okay, before I went to Tiger’s Nest, I was looking at potentially new hiking shoes and boots. As I was doing this analysis, there was one that kept coming up. It was the one I was preferring, but I tried it on and didn’t particularly feel comfortable with it. Then they came out with a new version, like three weeks before I left.
What was interesting is that when I was looking at this, I noticed that in some of the messaging, the new claim to fame was starting to show up. That’s what I mean by simplification. If we’re saying our shoe is 22 ounces and it’s Gore-Tex waterproof—maybe that’s the new statement. It’s not great from a marketing standpoint, but it works for me here. A simple statement.
If before it was simply saying it’s 36 ounces and it was Gore-Tex, and maybe it’s not new polymer material—right? It’s an old statement. If we do a query on that shoe, or shoes with similar features and functions, did the statement—quoted generically—adapt to the new terminology, technology, and feature set?
Quick plug for Milestone. Milestone just launched their geo tool. The one thing I liked about it is they came to market not necessarily as a “check your AI rankings” tool, but they positioned it as: is there incorrect information?
Bill Scully (19:58)
That’s the tool I’m looking for. That’s really how I think these tools should be used.
Bill Hunt (20:01)
Right.
What they’re doing is they’ve got a series… or anybody—let’s keep this tool neutral—if you, as a customer, were to load in some statements or some prompts, a set of texts that represent you, and you wanted to see if it’s correct or not…
Every geo workshop I’ve done, one of the top three questions is: how do we detect incorrect information, and more importantly, how do we get rid of it? I tell them, you can only detect incorrect information if you’re actually comparing what’s correct to what’s incorrect. How do you source the incorrect?
So you either have to ask a broad question—something about you, like “Where is our headquarters?” or “What is our mission statement?”—or whatever you think might be inferred incorrectly. In this case, it might be: “Are our hiking shoes waterproof?”
Bill Scully (21:13)
Right.
Bill Hunt (21:14)
And if it comes back and says, “No, they’re not,” or “No, they don’t use Gore-Tex”—which actually is a brand name—then we know it has incorrect or non-current information.
So if we pepper some prompts—potentially, you’re the prompt master, not me—if we write prompts to ferret out inconsistent, incorrect, incomplete, or irregular data, we can actually see what’s returned and then just do a simple delta lookup: did it say X or did it say Y? And we can find it.
I think that’s the interesting thing, sort of point three I had for this, getting back to this episode: every migration event effectively publishes a new version.
Bill Scully (21:52)
Yeah.
Bill Hunt (22:09)
Of this information. And so it’s not like we could do with Google: “Hey, let’s submit this through Search Console to get it fetched.” Google would come get it, and we could see it within 48 hours, right? Come get it, dangle a carrot: “Hey, the checksum has changed, come refresh it.”
We’re at the mercy of either A, in-model training, or B, a RAG trigger. And this has got me thinking: in this content, can we…
Bill Scully (22:18)
Yeah. Right.
Bill Hunt (22:39)
And this gets back to this same guy—I have to think of his name. He came out with this idea: put a block in your blog where people can click it and it actually submits some question to the LLM. You know who I’m talking about? I think he’s Turkish or Romanian. Anyway, he built that module for your blog so that when you click it, it’s sort of a Pseudo submission slash prompt question to ChatGPT. The idea was that if you prompt this, it would effectively ask ChatGPT to fetch something related to the post.
Again, I’m just brain-dumping here, but maybe there are 10 critical pieces of information that have changed that we probably want to prompt the LLMs to figure out. Can we, using his methodology, pre-prompt? Can we physically ask a question that might stimulate that new information as the answer?
So maybe the shoes I like—the Merrill Moab Gen 3…
Right. So if I went out and said, “Break down the features of the Merrill Moab Gen 3 hiking shoes,” that in theory would prompt a series of RAG prompts to Merrill to get their updated data. Does that seed it into the corpus or not? I don’t know. Smarter people than me will test that. But it’s a “what if” scenario.
Bill Scully (24:09)
Yeah
Bill Hunt (24:36)
And so can we, when we do a migration—or even better, going back to my days working with New Balance and ASICS—address this proactively? A big problem we used to have was launching shoes in America for the New York Marathon. Then the next thing you know, markets like Australia and the UK were complaining about losing traffic to the US, because the US had the new shoe, which they introduced the day before the marathon. It wasn’t even available in the other markets.
When it eventually became available, all the link equity was tied to the US. Google was ranking the US page over the Australian page for the new model of marathon or ultramarathon running shoe. So this is one of those things where maybe there’s a pre-seeding—or an immediate post-migration seeding—we can do to get the LLMs to ingest this. Again, thinking out loud. This is like a Mike King-type research project.
Bill Scully (25:43)
Interesting. You could do a search like, “What do you know about Bill Scully, SEO?” and do a deep search. Twenty minutes later, it tells you, “I’m a coach for the Special Olympics.”
Bill Hunt (25:49)
There you go. The most brilliant SEO out there yet.
But people have been doing this. A lot of people have poo-pooed it—Will Reynolds and those guys—all say this works brilliantly well. You create one of these top 10 lists or whatever, you seed it with yourself, put it in your footer, and the next thing you know, it’s gospel in an LLM. So maybe… maybe, let me think out loud, this is a step we should try.
We do a migration, maybe it’s a product refresh or even a product launch. We create a “claim to fame” page—similar to the Bill Scully “world-leading SEO,” your claim to fame for what you’ve added for this. Then we do a prompt asking for it, and maybe that seed page, which paraphrases all the goodness we just deployed across a number of pages, can act as a catalyst for getting it indexed. I don’t know.
Pivoting to the next point: the counter is true, right? What if we deploy and update a lot of stuff—launch a new generation of a shoe or product—and everything is still referencing the old stuff? I think that’s because there hasn’t been anything to force a refresh. It’s not like Google comes in and checks a page nine times after it’s been relaunched to know the content has changed.
We’ve still got the embedded data, content, and content crawl. So it can live almost in perpetuity until something forces a refresh. Again, I think I’m over-generalizing and over-specifying around product information because that tends to change the most frequently.
Bill Scully (28:00)
Yeah, but think about that. I actually have a migration coming up where a company is being absorbed by another brand. I wonder how long it will take for the LLMs to catch up with the product being associated with the new brand.
You have 20 years of information on the web describing the product being associated with the old brand. Now we’re starting with one page in the migration saying that the product is associated with the new brand.
Bill Hunt (28:49)
That just gave me an idea. I’m working with a company that has a solution, and I just came up with a new use for it—so thank you, I’ll give you credit for it. Let me answer one way we might be able to do that.
In helping these guys, I’ve been digging a lot into schema. One thing I’m finding is: 99% of SEOs say schema has no material effect on LLMs. Then you’ve got 1%—typically people selling schema or SEOs who are still attached to schema or have recently done a deployment—who are adamant that it materially does have an effect. I’m in the neutral zone; I can advocate both sides.
Bill Scully (29:43)
Well, but you have to say that Google uses it, right?
Bill Hunt (29:49)
Right. Google uses it. Here’s how everybody cop-outs on this: Google uses it, but we know that Google’s systems are probably the most… I mean, we’ve run to this, did a whole episode about how dysfunctional Google systems are—between crawl and rendering, duplicates, serving.
So imagine how wired they are with all the new AI stuff. Why would the AI people talk to the mere mortal crawl people? I’m skeptical of this, but I’m with you. Let’s take the nuance further.
Bill Scully (30:16)
Complex.
Bill Hunt (30:35)
You wanted to know how we might be able to influence these. Here’s where I’m seeing this. I see the cool people referencing “same as,” and here’s how I just wrote a model for this company. You’re seeing it, right? Think of it as a bow tie: LLMs want fact, and they’re using fact-based sources.
Wikipedia, everybody points to that. I’m a fan of Wiki—WikiGraph, Wikidata, Wikibase. I’ve only written “Wikidata” like 9,000 times in the past three weeks. Google has put, like, $250 million into Wikidata. That’s because they like databases. Here’s one thing I’ve been telling people to do.
Bill Scully (31:13)
Yeah, yeah, we get these.
Bill Hunt (31:34)
“Same as” works. So this is your scenario: we’re doing an acquisition, and it’s a standard acquisition model. We say, “They’re an XYZ company now.” Then, the next thing you know—as you said, risk-aversely—we took the thousand-page website and now it’s jammed into one page. Good luck. Own it now. You’re ours. Who cares about the old brand?
Here’s where I think we need to do something: we need to go back to these databases. I’ve been working with some models around what I’m calling “entity lineage.” What I’m finding are real problems where companies use the trade name, but legally, they’re a different name.
Bill Scully (32:26)
Mm-hmm.
Bill Hunt (32:28)
Let’s take Absolute, for example. Colloquially, it’s known as Absolute or Absolute Vodka. The legal name is the Absolute Company.
Bill Scully (32:38)
Mm-hmm.
Bill Hunt (32:40)
Two different entities. What does Wikibase and Wikipedia have? Wikipedia tends to mirror whoever set it up. If a PR person set it up and used the trade name, it’s probably “Absolute.” But Wikidata probably uses “The Absolute Company,” with references to either the trade name, a trademark, or a DBA.
Bill Scully (32:41)
Yeah.
Bill Hunt (33:09)
So “Absolute” or “Absolute Vodka,” or something like that. When we have a conflict between the trade name and the operating company name, we have to make the connection. That’s my bow tie. There’s an entity in the middle, public databases—everything from the SEC to state-registered company databases, trademarks, Dun & Bradstreet…
Bill Scully (33:35)
Branch stream, there’s all kinds of stuff. Yeah.
Bill Hunt (33:38)
Wikipedia, Wikidata, all referencing the company. On the other side, we’ve got the website and all the public references mentioning this. What I would do is make sure every place we can submit a change references the acquisition. References that the company that got acquired or absorbed is now part of the parent.
Bill Scully (33:40)
A business hero.
Bill Hunt (34:08)
If their website is still available—because they’re in that transition period—I would update the schema. Back to the Google side of this: this company now has a parent reference in the organizational stack. That acknowledges the parent-child relationship on Google.
I found in some research that when all the public databases say one thing, and your “moment of truth” is in schema—it’s funny, I did this in an industry (I can’t even say which)—I looked at 27 companies. Only one of them had a true trade-name to organization-name hierarchy established on their website.
Bill Hunt (35:13)
And of the 27, twelve— all public information was the trade name. All the databases, and then all the front-end marketing—Facebook, Wikipedia, all the socials—were all the trade names. So there was a complete disconnect.
When we looked at the knowledge graph, those twelve that had the disconnect had a disconnected knowledge graph, meaning Google was showing the official business name as the knowledge graph and not the trade name. So there was a little bit of confusion in the marketplace.
What I’m saying is, for you, this is a great version of a migration. It’s one of the nuances that never gets talked about—both acquisitions and divestitures. We’ve actually written a module in our new course. I’m putting on my Bill Scully marketing genius hat here: we’ve been working on a course.
Bill Hunt (36:13)
For site migrations, where we go through similar steps no matter what your migration definition is, and then unique things. I think there are twelve unique scenarios where there’s some different action—like an acquisition, a divestiture, a joint venture, all those.
Getting back to what you were saying: that’s what I would do. I would use schema to create the parent-child relationship, start adapting their entity lineage, go to all these public databases (which LLMs do get their data from), and make sure the parent-child lineage is updated there. Then, circling back to our kumbaya moment, make sure we’re adapting this for Google, because many of these LLMs actually source their information from Google.
I think we can. Again, what I’m basically saying is we need to think about this a little differently. It’s the same, but different. Just like we’re checking rankings post-migration, as we said, we should be checking AI/LLM references to see when the new content piece gets picked up—especially if it’s a materially different change.
Entity-wise, does the entity change? Maybe you move. I mean, look at all the businesses because of this California tax proposal. Google founders said, “Hey, we’re moving all of our businesses to Texas.” So a lot of companies have moved. Have they updated these data sets?
The last thing I want to talk about—again, this is an interesting point—is that migration failure may be harder to detect.
Bill Scully (37:53)
Yeah.
Bill Hunt (38:11)
In this new environment, because of the zero-click impact. We would know if a migration failed within days. The big one is when all the pages disappeared because some knucklehead failed to remove the robots block. The second is when developers lied and said they did 100% redirects one-to-one, then rolled it back the day before to no redirects or partial redirects.
Bill Scully (38:19)
Yeah. Yep.
Bill Hunt (38:40)
We’ve got seven hops. All those things gave us a clue because we could see, within a week, typically in Search Console, the nosedive in traffic. I think it’s going to be harder to detect now because we don’t have as strong of a traffic signal as we had. Rankings may still look fine, but what’s the impact on the AI overviews?
And like we said, is there a way we can push the goodness in and see that? I don’t know. Anything you’re thinking about why we may not see an impact is much because of AI?
Bill Scully (39:24)
Well, I think it’s going to be hard to have an immediate impact, right? So you migrate, and you’re still competing with an existing model. That model doesn’t get updated regularly.
I don’t focus so much on that except for the errors. And I think that if you do a good job creating content—semantically for the user as a typical SEO—I think you get the benefits of being included in the AI. It’s… I mean, it’s certainly not going to be days before you see changes. It’d be interesting to see what happens. Like, do you drop out, and then does it take a while to come back as new?
Bill Hunt (40:24)
Interesting test. I think maybe it’s the depth of material change—that’s the sort of legal mindset: “material change.”
Bill Scully (40:35)
So, if you just want another example—like, what if, in this case, the site’s going to be completely different because it’s going to be tacked on and integrated—but what if you were changing your company name? Like, what if it was a domain change?
It’d be really interesting to see what happens in the AI/LLM reviews. And I don’t pay for a tool to track that, so I don’t track that—but it would be really interesting to take maybe some key pages and see what’s working before the migration, and what’s working after. Yeah, that would be a really good experiment. Because right now, all I’m doing is worried about traffic.
Bill Hunt (41:32)
And I think, if we step back and think about some of this, one of the things we’re talking about is—you nailed it with this whole point of inclusion.
Let’s walk through the scenarios, right? The one thing we wanted to do, we hoped that whatever we did on the new site—let’s just stay on your piece. We’ve absorbed a company, we’ve acquired them. All their assets now become our assets, which means we probably had a domain change. Their legacy domain has now been redirected and absorbed into the existing company.
Can both traditional Google and new Google/LLMs understand this? Did we lose any content that was driving AI overviews? This is my bugaboo with acquisitions. I’ve ranted about this: I’ve seen how many brilliant companies were devastated. Every company I’ve ever worked with where 10,000 pages of content became one page as a feature, as a tool, or as a solution on their mega site. So they want to know, “Why did the leads dry up?” Well, all the content we had to drive leads is gone.
Bill Scully (42:55)
And there’s a reason they were acquired, you know.
Bill Hunt (43:01)
Who thunk it, right? And that’s a whole other issue. We embedded a process—definitely at IBM—and integrated this into the acquisition workstream. I’ve worked with some private equity firms when they’ve done it recently to try to preserve this.
But that’s number one: What did we buy? I think it’s a whole different scenario with an acquisition, because the acquirers are giddy about the acquisition. My rant about brand marketing companies: when they buy a brand—say, Brooks Brothers—they’re so giddy about acquiring the legacy brand that they carve it up into markets and sell it to the highest bidder. They make their money off $5 million or $10 million a pop per market. They really don’t care what happens to it.
I think this is what happens in acquisitions. We’re so giddy about the hunt, winning the acquisition, and integrating it, that we don’t think about preserving legacy value. One thing I loved about Michael Dell: post-migration, post-acquisition, he always made sure that all of that knowledge was dumped into site search and into the website to preserve the equity behind it. He’s the only senior executive I ever met who really cared about that—the whole legacy of why we bought the company.
If we pivot it to, say, replatforming from Marketo to Shopify, nothing changes except the engine. Obviously, we’ve got URL changes, template changes, per se—but the same page is probably there, and the content is there for the most part.
Bill Hunt (44:55)
You know, that didn’t materially change. Now we’ve got potential crawling issues, which is sort of a moot point with LLMs, especially on the RAG side of things. They can still fetch it if the domains stay the same, if they know you’ve made it. But when we have a significant change in content—like when we go from 30 pages to… I’m seeing a lot of people now talking about compressing content.
I think there was one example, maybe from Wix or somebody, where they went from 3,000 pages to 30 pages in a particular vertical, and their traffic tripled. I think people are going to start doing more of that. But what did we lose? And I think, to your point, if we’re not monitoring what’s showing up in AI overviews, we may not know what content contributed to it.
In traditional search, we could run that report we built where we could get 30,000 rows of data out of the Search Console, look at the words on the page, and say, “Wow, okay, that page is gone.” And so all the content associated with that page either: A) migrated elsewhere, or B) is now gone. We could see why we lost that traffic. In AI overviews, we don’t have the same luxury—unless we pay a small fortune to write prompts to see that in AI tools.
Bill Scully (45:05)
That’s the point. I mean, you would have to have thousands of prompts, which would cost tens of thousands of dollars. You could, I’m just thinking off the top of my head, pull from GA4 the referrers and see what pages are driving traffic from LLMs.
Bill Hunt (45:27)
So let’s noodle on this. Maybe we do an episode on how we might be able to check this without breaking the bank. My other one is: I’m all about, what are we known for? What can’t we lose? It’s always interesting when you ask people this. My whole site migration model—I built the tool and, literally, when I was shutting it down because I was out of the migration world, I could not give it away.
Nobody wanted that rigor of even doing a data-driven “what must we save” scenario. They’re like, “I don’t want to… who’s gonna collect that data? What nut job is actually going to look at 14 variables to weigh pages?”
Bill Scully (46:05)
Pay me for the task, and I won’t tell you what happened.
Bill Hunt (46:09)
Exactly, right. You know, just migrate it, or we’re going to pick and choose based on our intuition. But the idea was: maybe we can allocate 10 precious prompts to something material. What is the biggest thing we don’t want to lose? What is our claim to fame?
Bill Scully (46:31)
Put a different spin on it. Don’t even… well, I guess you would have to worry about it for the migration, but that’s a great question. What do you want to be known for on the web?
Bill Hunt (46:46)
And what are you known for? Is it accurate? Is it still there post-migration, post-change event? We’ve been trying to figure out a word because it’s more than just a migration in the truest sense. But, anyway, if we move into a wrap-up, I think we’ve given some great tips for folks. One is, like you said, go into GA, do that filter to see what—
Bill Scully (46:47)
What are… yeah, right, right. Right.
Bill Hunt (47:08)
Traffic is coming to you from LLMs today. Let’s make sure you didn’t lose that content in the migration event. Two, let’s brainstorm—what are the 10 things that would be devastating if we weren’t known for them on the web? Make sure that you’re still known for them, that they exist somewhere, that they’re still visible.
God forbid we forget about the existing, non-AI version of Google, which is still driving, for many people, 90% of their traffic. All the things we used to think about with migrations are still true; I just think putting in the AI-era flavor may bump it up on people’s priority list. I’m actually pulling data now for one team that was given two days to do a migration because they’re losing FY 2025 funds. They’re going to migrate as much as humanly possible in the next two weeks, and they came to me and said, “Bill, can you create your magical model of what we must migrate and what we need to make sure we transfer? And can you limit it to 100 pages?” Very interesting how people approach this.
Any other tips from your experience? I love this acquisition-migration challenge, and I think we should do a dedicated episode, but anything else off the top of your head?
Bill Scully (47:02)
I guess I would ask: how could a brand protect itself from being replaced in the AI answers? Like you said, you want to make sure that you’re not cutting back on all of that great content that might not be driving traffic but is showing up in AI overviews.
We need to rethink this: fixing rankings is temporary, but fixing narratives in AI is going to be permanent. Rankings change, narratives probably don’t—unless you’ve screwed it up.
Bill Hunt (47:10)
Right. Hopefully, they can come quickly. I really want to test this idea of doing prompts to trigger a RAG visit to see if that materially changes what it knows about you—especially around product launches. That’d be an interesting test.
Cool. I think that’s all we’ve got. We’ve beaten this to death, but we’ve already got a couple of ideas for other episodes. I guess this is our time to plug people: tell us what you want to hear about, what problems you’re having, and we can hash through them and help you navigate this uncharted territory.
Bill Scully (47:18)
Yeah, we’ve talked to some people recently and gotten great feedback. If you’re enjoying the show, we would really appreciate it if you could take a moment to rate and review the podcast on your favorite platform. Bill and I are on LinkedIn, Twitter, and Substack. If you like these topics, please connect with us.
For all episodes and transcripts, visit internationalwebmastery.com/podcast. As a final thought, I’m going to create a link to a Looker Studio report that will show you all of your LLM traffic, which you can download and run yourself.
Bill Hunt (47:22)
Cool. Yep, see you next time.
Bill Scully (47:25)
Thanks, Bill. All right, bye.