In this episode, Bill Hunt and Bill Scully explore how hreflang, geo-targeting, and international SEO perform in a digital landscape influenced by AI-driven search results.
They discuss why correctly implemented hreflang tags can still result in pages from the “wrong” country appearing in search, and why this behavior reflects how modern search systems prioritize relevance, authority, and content structure over geographic signals.
The conversation provides actionable insights for global brands, website owners, and digital marketers looking to optimize international websites, improve localization strategies, maintain content consistency across regions, and protect brand visibility in a rapidly evolving search environment.
Key Takeaways
- Hreflang still works in classic search, but AI does not use it for serving decisions.
- AI selects sources for accuracy, not geography or purchase intent.
- Query fan-out increases the chance of non-local pages being cited.
- Chunk-level retrieval rewards dense technical content over storefront pages.
- Language and country versions are often treated as interchangeable.
- Source diversity rules can surface multiple markets from the same brand.
- This is cross-market cannibalization, not a hreflang failure.
- Brands risk traffic, revenue, and control of regional messaging.
- Semantic parity across markets is now a defensive requirement.
- Forced geo-redirects create more problems than they solve.
Listen now on your favorite platform:
Youtube: https://youtu.be/Z5MaMv0hv-g
Spotify: https://open.spotify.com/episode/4nCyA2X4r57jHVeKufttun?si=lQXNlD0pSmylMd0fyzIGVg
Apple Podcast: https://podcasts.apple.com/us/podcast/hreflang-in-the-ai-era/id1755789311?i=1000746872229
Show Note Links
Why Global Search Misalignment Is An Engineering Feature And A Business Bug
Free International Web Mastery GSC Question Report
Transcript
Bill Scully (00:00.0)
Welcome to episode 41 of the International Web Effectiveness Podcast. I’m Bill Scully, here with Bill Hunt, and together we focus on strategies and techniques to improve your web presence and manage your global websites.
Today, we’re talking about hreflang in the AI era. But before we dive in, make sure you subscribe so you never miss an episode. If you enjoy the show, tell someone and leave us a quick review.
In our last episode, we discussed website migrations and the challenges that come with AI and large language models. That conversation really got me thinking about things I hadn’t considered regarding LLMs.
If you haven’t listened yet, I recommend going back—it ties directly into some of the topics we’ll cover today.
Bill, I noticed you responded to a LinkedIn post where someone claimed they had correctly implemented hreflang, but the AI grounding links were going to the wrong countries. Many people commented they were seeing similar errors.
From there, you did what you usually do—you dug deep, geeked out, and found some answers. So, why don’t you walk us through what’s happening?
Bill Hunt (01:32.6)
Absolutely. I definitely geeked out, because hreflang is one of my passions. I’ve been working on it for years, and the last thing I want to hear is that it’s not working—especially when people insist that it is.
I started by looking at it broadly, and then I reached out to the person who posted it. They shared details that let me dig deeper, and I found some fascinating insights.
First, hreflang still works in traditional search. But there are new realities when you look at AI-driven environments. This ties into some of our previous discussions about Google’s disparate systems.
To unpack the issue, here’s what happened: the original poster, Lily Ray, shared that the AI overview for her query was pretty accurate. But the grounding references—the links on the side—were sending users to Norway, the UAE, and a different page in the US.
She confirmed that hreflang was implemented correctly on her site. And that’s where the problem begins. When you move into the AI overview, the dynamics change completely.
We’ve all discussed SEO frustrations with the loss of the 10 blue links. Now, Google picks content it believes best answers the question. And that’s the core issue—the mechanical challenge behind what’s happening.
In my article for Search Engine Journal, I posed the question: Is this working the way it’s supposed to from an AI perspective, or is it broken?
Bill Hunt (03:57.163)
The short answer is that AI is working exactly as it’s designed to. The AI component is focused on fact-checking and providing the most accurate response.
AI overviews take your prompt or query and try to answer it authoritatively. One key thing to understand is query fan-out. Mike King, for example, is doing a webinar tomorrow on this topic.
When I analyzed the actual query, I could see what was happening. The first part is that AI seeks the right answer—an authoritative response. The second is something called semantic overlap.
Unless you work extensively with international sites, you might not immediately grasp how AI handles multilingual content. AI models use chunk-level retrieval, pulling content in pieces.
If you ask Gemini about this, it will tell you it isn’t “multilingual” per se—they call it language-agnostic. Essentially, what it does is take a piece of content and convert it into an entity.
Bill Hunt (05:54.88)
If you’re searching for a product, a place, or anything that can be boiled down into an entity, that entity is collapsed into a semantic unit—a numerical representation.
AI doesn’t actually “see” language. This holds true across 44 languages. That entity—whether it’s a home, a catcher’s mitt, or a bicycle—has the same numeric equivalence at a semantic level, even if there are 40 different words for “bicycle.”
Bill Scully (06:41.547)
That makes sense when you think about the model. It doesn’t treat a bicycle in Portugal differently from a bicycle in the US.
Bill Hunt (06:52.105)
Exactly.
Bill Hunt (06:58.015)
A bicycle might have 30 different language terms, but they all collapse into the same concept: information about a bicycle.
If I were explaining this with examples, I’d stick with “bicycle.” That’s essentially the first thing the AI does.
Next, I looked at where it gets its source information. That’s the second piece I discovered.
I took the query from the AI Overview and ran it through Mike King’s query fan-out tool. Using my Gemini ID, it returned a query fan-out.
In the basic AI Overview, it produced 13 results—13 fan-outs or augmentations.
Bill Scully (08:03.509)
Very cool.
Bill Hunt (08:05.695)
For anyone who hasn’t used one of these fan-out tools, especially Mike King’s, the first row is the essence of what’s happening.
It rewrites the prompt. Honestly, it makes you feel a bit inferior as a human because of how well it rewrites it.
Bill Scully (08:25.355)
The way you search, right? Because it knows your intent.
Bill Hunt (08:34.16)
Exactly. You see it and think, “Why didn’t I phrase it that way?”
That’s step one: the AI takes what you wrote and infers what you really meant. There’s even a note that says it rewrote the query for context and clarity.
Step two is intent inference. Is this informational? Commercial? Navigational? It categorizes the query accordingly.
Reading through the 13 fan-outs, you can see that after rewriting, it recognized this as an informational query. The next 11 were refinements aimed at retrieving precise information.
Bill Scully (09:29.749)
Quick question—were those refinements testing topical knowledge, follow-up questions, or just the same question phrased differently?
Bill Hunt (09:52.872)
It was a mix. Let me explain. First, it rewrote the query. Then it created something more explicit, because the original query had a specific audience and use case—it was about a product used for a particular purpose.
Bill Scully (10:15.691)
Mm-hmm.
Bill Hunt (10:23.986)
The first refinement came back as: “How does this product help with that use?” The inferred intent was to understand practical application and benefits.
The next three refinements drilled further into that application. Then the second or third level focused on the audience itself, followed by three more queries specific to that user group.
After that, it pulled back to a higher level, looking at the function in a more macro context.
Bill Scully (11:05.088)
Wow.
Bill Hunt (11:16.325)
Then there were two queries comparing a higher-order model, like a premium version, with an everyday model. That’s where the system started pulling in results from Norway. I’ll explain why in a moment.
Another query focused on setup for maximum efficacy, and yet another reformulated the function mechanics. That one felt a little odd.
When I ran the same prompt in AI mode, there were 37 refinements. AI mode drills much deeper. Some focused on this function, this audience, and technical nuances. One even touched on legality.
There was a lot of detailed, cryptic information in the mix.
Bill Hunt (12:41.926)
Hopefully, people are following along—there was so much insight. I took all of this—the query fan-out CSV, the URLs, even the vector embeddings—and fed it into AI. I asked, “Please tell me why it picked Norway as a grounding.”
What was really cool, and I think it’s in the show notes, is that AI explained it using implicit intent versus local ranking. Several subqueries—rows one, four, eight, nine, and ten—were categorized as implicit. AI was looking for definitions, legality, and mechanics.
When it searched for related information, it prioritized informational density. Then it introduced the concept of “best-first answer logic.” Out of the 13 queries, it said that in rows one and four, the subsequent query was looking for a specific chunk. It found that chunk in Norway—in the Norwegian version.
Because it was a specific chunk, AI didn’t need to check for an English-language equivalent. It found the first-best answer in Norway and assigned that block a high confidence score for the subquery: “How does this product type help with a specific need?”
I had the client exit. Subquery number four was the explicit point of reference. Norway had content satisfying that prompt, and AI converted it into “computer speak” to fulfill the query.
Bill Hunt (15:12.101)
Then there were role-specific searches where no single market dominated, but there was enough content in Norway to satisfy the model. Again, it gave a nod to Norway, pointing to rows one, four, eight, nine, and ten in one instance.
Bill Scully (15:33.483)
Yeah.
Bill Hunt (15:42.055)
In another instance, it pointed to rows one, three, and seven, highlighting the exact points on the page. Then it considered functionality deemed better in this use case, which warranted a higher-level, premium product.
Norway’s content was skewed toward the premium product—they used the trade name instead of the general market-friendly version. AI honed in on that, giving another boost to Norway, while noting similar findings in the UAE.
Bill Scully (16:24.853)
Huh. Wow.
Bill Scully (16:35.659)
And this is primarily evaluating the vector embeddings you fed it?
Bill Hunt (16:44.25)
Yes, vector embeddings. It looked at the pages I isolated, the vector embeddings, and what it crawled and parsed on each page.
It could even tell me which paragraph it pulled the insights from. The weightings were heavier on Norway, and when I checked, it all aligned.
Bill Scully (16:54.687)
Right, right.
Bill Hunt (17:16.708)
We don’t know exactly when the system—Google or Gemini—visited these pages. But when I checked the last true update in the header status, the Norway and UAE pages were newer than the US page.
It wasn’t that the system rebuilt the page live; the last actual update was more recent. The content in Norway and UAE had been refreshed more recently than in the US.
It didn’t provide a specific date, but the content differences were clear—it mirrored the US content but with unique variations.
Bill Scully (18:00.128)
My God. Yeah. Yeah, yeah, yeah.
Bill Hunt (18:14.937)
How my brain works led me down five rat holes. Let me take a breath.
What we determined was that we could absolutely identify why Norway was used. Another insight Gemini pulled out—something Claude, ChatGPT, or DeepSeek didn’t—was a diversity filter.
Because Gemini was criticized for being too white or too American-centric, it specifically highlighted diversity. I actually removed it from my notes because it felt too geeky, but it’s in my bigger article.
It explicitly says that if US and Norway were identical, Google could have used content from another market for grounding, due to a diversity standard.
Bill Scully (19:48.395)
Okay. But basically, there are no local result signals coming from these AI Overviews.
Bill Hunt (19:58.903)
Right. That was the second thing I found. Then I checked the hreflang tags, thinking maybe that mattered.
The system said, “Sorry, nowhere in the processing cycle do we consider geo signals.”
Bill Scully (20:15.595)
I don’t know if you’ve made this public, but you shared with me what you think is the process order of Google’s normal ranking systems.
Do you have an idea of how the AI Overview processing works?
Bill Hunt (20:45.75)
It’s all about trust—whoever, wherever, whatever provides the best answer. I couldn’t find a case where the system would second-guess itself.
I researched across five different tools and tried every angle. I tried to debunk it, play devil’s advocate, or counter-argue the AI. Nothing. No scenario where commercial convenience, logistical factors, or market location influenced the results.
The AI calls this “logistical weight,” meaning it considers practical factors for users. But in reality, none of that influenced the answers.
Bill Hunt (22:46.142)
There was a term in my original article, something like “geographic weight,” but we changed it later. I’ll have to check the article.
Bill Scully (22:51.179)
I’ll find it while you keep talking.
Bill Hunt (22:54.656)
Yeah, the word you suggested was more accurate. Anyway, to circle back, every single process within AI is about getting the right answer and presenting a complete answer.
Remember, the links weren’t showing where to buy anything, because there was no purchase intent in this prompt. When I added, “Where do I buy this?” the dynamic completely changed.
Not only did the answers shift, but it also gave me locations to buy from. Norway was still on the list because AI thought I could purchase from there, but the results prioritized my commercial intent.
In the query fan-out, it even accounted for my location. It asked, “Are there places to buy this from the US?” That was new, because the original query didn’t include any purchase intent.
Every aspect pointed to relevant cross-language retrieval. It even explained vectors in high-dimensional space and how information compression treats entities as the same, regardless of language.
Bill Hunt (24:50.000)
To your point about traditional search processes—this came from a Google roundtable. They discussed “duping,” and someone named Alan or Adam mentioned that Google has disparate teams, each optimizing performance independently.
These teams don’t always communicate. What he emphasized, and what John has reiterated, is that in traditional search, it’s all about the serving layer.
The better page could theoretically be the Norwegian page. It’s unlikely because it’s in Norwegian, but a UK or Australian page could be topically stronger. That ties back to one of your earlier terms.
Bill Hunt (25:23.819)
It could be the better result, but because my IP is in the US and my language preference is US English, all the signals say I’m in the US. The serving layer would deliver a US version of the page, even if it’s topically inferior.
Bill Scully (25:48.149)
Mm-hmm.
Bill Hunt (25:49.813)
Exactly. Especially if there’s any hint of commercial intent.
Bill Scully (25:58.634)
Right, right.
Bill Hunt (25:59.989)
Right. AI doesn’t have that bias—the “goodness of our heart” factor doesn’t exist.
Bill Scully (26:06.429)
Yeah. The term I suggested—geographic leakage—you said it’s more like geographic indifference.
Bill Hunt (26:19.357)
Exactly. Full credit to you for that term. The system truly doesn’t care where the information came from or where it’s grounding a link. Location doesn’t matter.
Bill Scully (26:39.903)
Yeah. Information can come from anywhere—pictures, text, any source. For LLMs, entities are entities. They have geographic indifference. Language, page location, or tags don’t matter. They just process information to generate the right answer.
Bill Hunt (27:15.745)
Exactly—the best possible answer. That’s where the frustration comes in.
I’ve mentioned this before. In a past episode, Motoko wrote about it. People were upset that Google, especially outside the US, was automatically translating content.
In this case, the Norwegian reference was translated to English for the grounding link. The UAE reference was converted from Arabic to English.
This gave two mirrored English-language references, even though the URLs pointed to Norway and the UAE.
Bill Scully (28:15.627)
I didn’t realize that. You said earlier they were showing three links to three different pages.
Bill Hunt (28:24.34)
Right. Two were identical translations—the first to Norway, the second to the UAE. The third link was a different shopping page for the US for that company.
Bill Scully (28:37.791)
What wasn’t clear to me was the link—the grounding links you mentioned before.
Yeah, right, right. So I didn’t realize Google had translated that. If you clicked on it, did it actually translate the page?
Bill Hunt (28:48.852)
The grounding links appear on the right rail—it’s what we would consider links.
Bill Hunt (29:03.008)
Here’s what happened when you went to the page. It didn’t open in a translation window, which I expected. Maybe that depends on the site.
When you got there, it popped up a gate saying, “It appears you came from the US. Do you want the US content?” This brand gated it and asked if you wanted to go to the US version or stay on the local site.
Bill Scully (29:12.629)
Okay.
Bill Hunt (29:32.997)
Same thing on the Norwegian and UAE sites. The gating prevented confusion, but the right-rail links were in English, not Norwegian. I didn’t realize I was on the Norwegian site until I navigated there, which was interesting.
Bill Scully (29:34.731)
Perfectly done. Yeah.
Bill Hunt (30:01.353)
The brand was forcing me through door A or door B. It would have been interesting if Google overlaid Google Translate—maybe some sort of proxy could do that.
Initially, I thought this was a Google marketing feed or a shopping feed gone wild. It looked like an extraction from the shopping feed.
Because the users were gated, it treated me as an American entering the site. This raises the question: Does Norway “lose” because it got the snippet? It could potentially show Norway received a citation over the US, which might affect US team reporting if they use citation tools.
In the old days, SEOs didn’t care which market ranked as long as there was a ranking. As traffic recovery became important, market-specific rankings mattered more. That’s why I built H.R.F. Lang Builder—it became a massive problem.
Bill Hunt (30:27.177)
I want to come back to the workflow in traditional Google. There’s the serving layer, but also the crawl layer with hreflang.
From an LLM standpoint, we haven’t found evidence or tests showing that an LLM ingests metadata. By metadata, I mean raw HTML tags—not structured schema like price.
So for both model training and retrieval-augmented generation (RAG), the LLM environment doesn’t ingest hreflang or raw HTML metadata.
Bill Scully (31:54.281)
Really? Wow.
Bill Hunt (32:24.71)
I did hundreds of searches and consulted smart people. I couldn’t find a scenario where an LLM would ingest HTML—whether for testing or any other reason.
If it’s not ingesting HTML, then on-page hreflang tags would never be seen in the LLM environment. Would we agree?
Bill Hunt (32:58.717)
Right. If we use Claude instead of ChatGPT, because I want to control the context…
Bill Scully (33:02.747)
Cool. So what would happen if you have stuff in JSON?
Bill Hunt (33:09.733)
It’s probably not going to be seen. We’ve seen interesting developments with Screaming Frog—they now have a report showing how much content, how many words, are in JavaScript.
Bill Scully (33:29.771)
Oh wow. I haven’t heard of that.
Bill Hunt (33:29.885)
Most LLMs aren’t ingesting content in JavaScript. If information like schema tags or hreflang is served via JSON in JavaScript, it’s likely invisible to the model.
Bill Scully (33:37.259)
Right, right.
Bill Hunt (33:58.949)
To my knowledge, it’s not seen. Maybe someone has evidence otherwise. But if it’s in JavaScript, LLMs are probably missing it.
Bill Scully (34:05.878)
Do you think Google is using schema as part of content selection in the AI Overview?
Bill Hunt (34:16.7)
It could be a cross-validation layer. Let’s step back: indexing, crawling, rendering, deduplication—we suspect hreflang plays a role there. It may help explain why there are multiple English or Spanish versions.
Then there’s the serving layer. If schema is ingested during indexing and stored, could Google’s Gemini AI reference it from its crawling systems? Possibly.
Bill Scully (35:13.068)
If I’m an engineer, I’m saying no.
Bill Hunt (35:15.929)
It could be…
Bill Hunt (35:20.634)
Because as a superior AI, your understanding of content is far better than mere mortal crawling systems.
Bill Scully (35:32.652)
Who cares? The purpose of AI Overview is to provide an answer, not navigation.
Bill Hunt (35:41.488)
Right. Could Google layer in semantic or schema validation at the entity or Knowledge Graph level? Potentially.
If I have an organization with a name and address pulled into a Knowledge Graph, AI could use that as a validation source. That’s what brands have done via Google Business Profile.
When you upload addresses, phone numbers, and other facts, it could cross-check with schema. If the schema is accurate, it could be used; if it’s poor quality, it may be ignored.
Bill Hunt (37:25.945)
I looked at 27 businesses in one vertical. Only two had reasonably complete and accurate schema. So overall, the validity and usefulness of schema across industries is limited.
Some brands might do better if they submit through a product feed or merchant feed. With UGC data, the model can “train”—or more accurately, ingest it quickly to improve answers.
I like to think Google, as the absorber of all this data, might have a reconciliatory database to validate schema, but that’s speculative.
Bill Scully (38:29.45)
Yeah.
Bill Hunt (38:48.426)
And then, if we think about this AI processing—during the Gemini hunting, gathering, validating process to answer a query—there were 13 sub-queries. That’s what Mike King’s tool extracted in the JSON file he exported.
Those 13 pieces determined this was on the left rail—a really correct, detailed answer to the question, with a couple of references.
We all got derailed by the two references on the right, which came from Norway and the UAE.
What was on the left rail was solid. That answered the user’s information need. They probably didn’t need to query further unless they said, “I want to buy that.”
Bill Scully (39:46.262)
Right, right, right.
Bill Hunt (39:46.542)
Right. So all of this was moot because the answer on the left was enough to fulfill the question—100% yes.
When I added purchase intent, I got better commercial answers.
Bill Scully (40:16.577)
Yeah.
Bill Hunt (40:16.641)
Long story short: nowhere in this processing was language, purchasability, or geographical correctness factored in.
That’s why I said if Google were to do that, maybe—but I couldn’t find any evidence. I spent hours trying. There was no secondary check for market specificity.
Nothing needed to know my market to give the first answer.
Bill Scully (41:04.394)
Right, right. I actually forgot—I wrote an article last week. Google just released a new standard for e-commerce sites called Universal Commerce Protocol. They developed this specifically for agentic search.
Today, you can go into some browsers and say, “Buy me this, buy me that.” They developed this protocol with all the big sites. I’m guessing this will become part of their AI Overviews because 90% of the time you need that is when you’re about to purchase.
Bill Hunt (41:56.995)
Right. And if you look at the questions and fields, that overlays neatly with a lot of schema fields.
This is actually a business opportunity I just sent to someone—how easy would it be to take what we’re ingesting and validating for schema and create that UCP?
It’s UCP, as you said; I said UGC. So all these U’s, C’s, P’s, and F’s—yet another acronym—but it ties back to Google’s knowledge hopper.
If that data is coming in, and we have schema, do we even need schema on our website? Maybe not—we can just feed it.
It’s very commerce-skewed. As Google noted in their overview, availability, pricing, and nuances are key. Brands could tune this.
They even mentioned gamification—upsells, cross-sells, accessories. Merchants could enable different prices depending on these variables.
And this is already reaching Congress—a point someone made about Google gaining control over commercial personalization.
Bill Scully (43:29.674)
Okay.
Bill Hunt (43:56.664)
Stack the deck against people. But really, all this is doing is trying to remove ambiguity, improve clarity, and ensure specificity at the product level.
Again, this was an information-level query—location, product, and purchase ability never came into the equation.
When I ran the same query with purchase intent, it skewed not only the answer but also the grounding results. The intent of the query makes a huge difference.
Two things I want to get to—we aimed to keep this to 30 minutes because it’s a little geeky. Fundamental problem number one: the traditional serving functionality of hreflang does not fit into this process.
Bill Scully (44:32.275)
Wow.
Bill Hunt (44:56.075)
It actually causes a problem because it doesn’t work in this context.
We’ve talked a lot about geo specification. I think we did an episode about the ambiguity of, say, Nike in Australia.
The only thing different from all the other languages was the currency—Australian dollars, not U.S. dollars. Nothing indicated this content was truly Australian. It used a U.S. shoe size chart instead of an Australian one. Those signals get crowded out as metadata noise.
Another issue is vector collapsing across markets. In an AI overview, there’s an erosion of brand control.
We have a Norwegian site meant for Norwegians, in Norwegian. But it could be presented to anyone as a purchase path. This leakage creates user friction and material impact—zero clicks, in some cases.
Now, we’re not even sending people into a market because of semantic collapse. The entity is treated the same everywhere.
In this example, when I clicked the Norwegian page, I got a prompt: “Do you want to go into door A or door B?”
On a commercial query, this could create pricing ambiguity, VAT ambiguity, or trust degradation. It impacts both the brand and Google. Norway had the higher-level product, whereas the US had the consumer-level product.
So what does this mean for business? AI gave me a term: fact density. Can we determine if content from a different market is more extractable?
Think back to our previous episode on migrations. If we launch a new site, content, or product in a secondary market, that secondary-market data could become the primary data presented worldwide.
We also need to ask: is there a source of truth? In my test, the AI solution basically said the content in Norway is fresher than the content in the U.S.
What can a business do? We came up with the idea of a “market of truth.” In a decentralized environment, some may resist, but do we create a market of truth? Is it the U.S.? The U.K.?
If you’re a German company, you’d skew to Germany. If you’re a U.S. company, the mothership is the U.S.
Bill Hunt (45:30.000)
So is there a market of truth? We have to present all options in that market of truth. Maybe this supports why some clients have a global site: we show high-level products, low-level products, global products, and let markets filter.
Bill Scully (45:45.000)
So catchall, yeah, default.
Bill Hunt (46:00.000)
The second thing we need to revisit is a non-local visit protocol. Many companies haven’t updated this for years.
If someone from the U.S. comes to Norway, what do we do?
Do we use a door B model, a suggested route model? Door A usually sends people based on IP and language preference—so a U.S. visitor gets redirected to the U.S., which may not show the AI-represented answer.
Or do we do a door B model, with a banner suggesting a route?
Bill Scully (46:45.000)
I like that. I think most sites don’t make it obvious where you are. As SEOs and digital marketers, the first thing we look at is the URL. Most people would never know what region they’re in.
Sometimes the region selector is in the footer, and people never scroll that far.
Bill Hunt (47:15.000)
Exactly. And that’s another thing—you’ve moved that switcher, whether language or market. Some companies, like IBM, have gone to a language-centric approach: one English version in English.
But as we’ve seen with collapsing across all languages, even a language-centric approach doesn’t fully solve the problem. If you’re selling actual products, you need a market-centric approach. Price, availability, and other details require it.
This requires more thought—beyond what we can cover in a 50-minute podcast. It warrants a follow-up where we say, “Okay, we’ve identified the problem.”
I’m with you—if we pop something up, we need to make sure it doesn’t block the LLM or engine from accessing the source of truth. Users should get the best UX experience. One approach you’ve advocated for is a bailout.
I’ve rarely seen anyone chart: okay, 5,000 people came to choice A or B—how many chose A? How many chose B? Did that reflect positively or negatively in the checkout queue?
If they were routed back to the U.S. and the product wasn’t available, we shot ourselves in the foot.
Bill Scully (48:30.000)
Mm-hmm.
Bill Hunt (48:45.000)
A feedback loop is key. If 90% of users routed to their true market never enter an e-commerce path, maybe we rethink the routing.
Where did we send them? Homepage or product page? With our AI Now system, we know the path they came in on and the intent they followed.
Bill Scully (49:00.000)
Yeah.
Bill Hunt (49:05.000)
Maybe we can use the intelligent agent to route them to the same page in a different market, instead of dumping them on a homepage—which we know fails.
This is probably beyond our remit, but experience is worth discussing—especially in reports where people can analyze bailouts, option A vs. B, and see the downstream impact.
Bill Scully (49:40.000)
All right. If a brand did one thing differently this quarter to help with localization in LLMs, what would it be?
Bill Hunt (49:55.000)
Number one: check for citation cannibalization. Use the latest AI Overview tools—they show where you were cited or referenced, and they list the citation.
If you ran a U.S. report and the citations are non-U.S., that’s a flag.
This is like old-school SEO: SEOs would see rankings, but then when you looked closer, 60% of URLs ranking for Australia were actually U.S. or UK. They’d say, “I don’t care, I’m ranking,” and we’d say, “Yeah, but it’s not helping you.”
Bill Scully (50:10.000)
Right. All for the homepage.
Bill Hunt (50:15.000)
Exactly. So one action is to filter citations and references by market. See how many point to another market’s content. That’s my takeaway.
Bill Scully (50:25.000)
Okay, that’s great. I can also share a Looker report showing GA traffic to pages. That might give insight without using the bigger tools—like seeing where traffic is going today, or where it isn’t.
Bill Hunt (50:35.000)
Right. Or it could be going somewhere else—classic cannibalization.
Bill Scully (50:45.000)
Okay.
Well, we hope you’re finding value in our conversations. If you enjoy the show, please rate and review the podcast. We want to hear from you—your feedback helps us plan future episodes.
Follow International Webmastery on LinkedIn, Facebook, and X. You can also follow Bill and me there, and on Substack.
For episodes, links, and transcripts, visit internationalwebmastery.com/podcast. Thanks, Bill.
Bill Hunt (50:55.000)
Thanks.