AI is changing the way we interact with information, but what happens when it gets things wrong? In this thought-provoking episode, Bill Scully and Bill Hunt unpack the growing phenomenon of AI overview hallucinations—when search engines or chatbots confidently deliver false or misleading content.
From search visibility to brand reputation, they explore how these AI-generated inaccuracies can quietly damage businesses, erode user trust, and spark larger misinformation issues. With real examples and strategic advice, this episode equips business owners, marketers, and content creators with the insight needed to safeguard their digital credibility in an AI-driven world.
Key Takeaways
- Learn how AI hallucinations can lead to misinformation and confusion.
- Trust in AI-generated content is declining among users.
- Businesses must work to verify the accuracy of AI responses.
- AI Overview and Chat errors can be brand damaging.
- Companies should monitor results and 404 errors linked to from AI outputs.
- Proactive answer management of online information is necessary.
Catch the Full Episode
Youtube: https://youtu.be/wgCF7F0FuIU
Spotify: https://open.spotify.com/episode/4jzrwjsxgcikz70id2YwW9?si=XW4vvpEfTZuB7O-lv16ryA
Apple Podcast: https://podcasts.apple.com/us/podcast/ai-hallucinations-in-search-why-businesses-need-answers/id1755789311?i=1000702883484
Show Note Links
Barry Schwartz Study: AI Search Engines Are Confidently Wrong Too Often
https://www.seroundtable.com/ai-search-engines-wrong-39038.html
Implementing your Answers Management Strategy by Bill Hunt:
https://www.back-azimuth.com/implementing-your-answers-management-strategy/
Soundbites from
AI Search Has A Citation Problem By Klaudia Jaźwińska and Aisvarya Chandrasekar
https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
AI Link Hallucinations: Are they harming your brand? by Dixon Jones
https://waikay.io/link-hallucinations/
Transcript
Bill Scully (00:01)
Welcome to episode 31 of the International Web Effectiveness Podcast, where Bill Hunt and I—Bill Scully—focus on the strategies and techniques that improve your web presence and help manage global websites.
In episode 30, we talked about our trip to the Google Search Central event. In this episode, we’re focusing on AI overviews, hallucinations, and how they’re generating URLs that lead to 404 pages.
Hey Bill, how’s it going?
Bill Hunt (01:35)
Good, good. You?
Bill Scully (01:36)
Good. There’s been a lot of discussion recently around how AI search results are getting things wrong. I saw a post from Barry Schwartz on Search Engine Roundtable about a new study showing that AI-powered search engines often respond with confidence—but the answers are just flat-out wrong.
And you mentioned you had an article you wanted to talk about?
Bill Hunt (02:10)
Yeah. There are two articles, actually—one might be the same one you’re referring to.
One was from the Columbia Journalism Review. They looked at where citations in AI-generated answers are coming from—or if they’re coming from anywhere at all. They found that most of the time, there’s no clear source. Sometimes, the citation didn’t exist or wasn’t tied to the actual origin of the information.
So it’s not just a matter of transparency. It’s also about trust, credit, and the potential revenue loss for publishers. In some cases, these systems were attributing headlines or stories to the wrong people—sometimes even to people who had never published anything at all.
This ties back to one of the topics discussed at Search Central Live, where people raised concerns about AI giving out wrong phone numbers, company addresses, and all kinds of inaccurate business details.
Now, from Google’s perspective, you’re supposed to be able to validate that info through your Google Business Profile. But clearly, something’s off.
The third piece I came across was from Dixon Jones. He wrote about this on his new platform, YK. He pointed out a major issue for search professionals: not only are AI systems hallucinating answers, but they’re also fabricating URLs—links that don’t exist at all.
He gave an example involving IKEA. The AI cited a chairs page and linked to it, but the URL went to a 404 error. And it wasn’t a case of the page being removed—it had never existed in the first place.
I don’t remember the exact number of errors they found, but it was significant. When citations are included, they often reference sources that simply aren’t real.
Dixon had this great line: “You can’t patch hallucinations out of LLMs.” In other words, it’s not something you can just fix—it’s baked into how these models work.
So whether it’s a citation that’s missing, inaccurate, or totally made up—those are serious problems. And they lead to bigger issues for businesses. We’re not just talking about a wrong phone number. I get tons of spam—maybe a thousand emails a week—from people who think I’m an executive at JP Hunt Trucking Company.
They’re offering me $600,000 loans out of nowhere. I have no idea where that number came from.
Another example we were discussing before the show involved an Australian guy trying to travel to Chile. He asked ChatGPT whether, as an Australian citizen, he needed a visa. The AI told him no—he could stay for 90 days.
So he shows up at the airport, and they say, “You’re not flying without a visa.” Apparently, Chile didn’t offer an e-visa either, so that guy wasn’t going anywhere that day—and that’s not a cheap flight to miss.
Now, whose fault is that? Is it on the AI for giving the wrong answer, or on him for not double-checking the info at the source?
Ironically, 90% of the responses to his post were people saying, “Should’ve Googled it, bro.” And when people did Google it, they found that the AI’s answer was actually based on the U.S. State Department website—so it was responding as if he were an American, not Australian, despite his very clear query.
That’s a problem. These systems are confidently wrong, and the consequences can be very real.
Bill Scully (06:46)
Yeah.
Bill Hunt (06:56)
And it can cause real harm and damage. We just want to make sure people are aware of this. I think anyone providing information should be vetted through additional sources. But people tend to assume that Google’s response is the trusted voice.
I’ve been thinking—maybe we need something like authenticated results. Kind of like how, for pharmaceutical information, you’re required to have a pharmaceutical license. In some cases, you even need to submit the license.
In China, for example, you have to put your company registration code on your homepage to prove that you’re a legitimate business. Anyway, I think it’s a big issue.
Let’s get into some of your thoughts—where do you see problems? And then let’s talk about what companies can do.
Bill Scully (07:56)
Sure. I hadn’t really considered broken links before, but Dixon did some research and put together a pie chart that breaks it all down. According to that, Gemini had 41% broken links. ChatGPT 3.5 had 38%.
ChatGPT 4.0 was much better, with only 13% broken links. Claude was the best—only 8%.
If you’re in a hurry and doing research, and you see what looks like a citation, you assume the information is correct. But how many people are relying on this and getting it wrong? That’s pretty scary.
I’ve had a couple of experiences myself. For example, in our last podcast, I was trying to get info on the most recent Search Central event in New York City using Gemini. I asked something like, “At the most recent Search Central event in New York City…” and it came back saying there was no such event.
Bill Hunt (09:30)
Yeah.
Bill Scully (09:32)
It’s been about a week since that happened. But then I tried Grok, and Grok had solid information coming through—clearly using better sources.
Bill Hunt (09:49)
It should. I think Lily Ray posted on LinkedIn—either today or yesterday—saying that her recap of the event was already showing up in some engines the next day. For others, it showed up on day two. I don’t remember her exact breakdown, but she was monitoring where and how it appeared.
She wrote it, published it, and then worked her magic to get it picked up. Then she tracked how often and where it showed up.
That’s interesting because it ties back to when these LLMs are trained. When was the data updated? When did they process it?
Matoka wrote an article about capital shifting into emerging markets. Then I heard Scott Galloway on a podcast saying he’d moved 60% of his portfolio out of U.S. stocks—into Europe, and increasingly into Latin America. I did something similar, especially with Europe, expecting more spending there.
So I asked various chat engines about these trends. Most said they don’t do well with financial forecasting. One even said its last update was in October 2024—so no current data.
Like you, I tried reframing my prompt to see what it could pull. When I fed in articles from several financial sources, I got some directional insights. But again, the issue is timeliness.
That brings us to a key takeaway: what do you see when you search?
Step 101 for businesses is to use these tools and actually look at the SERP. See what comes back. We’ve said this in our live events too—look at the search results, see what’s being shown.
In our discussion on answer engine VP, we asked, does your company need one to manage your answer program? Start by looking at what’s being returned when people search for you. Then dig into where that information came from.
Use tools like YK, Dixon’s BrandRank (brandrank.ai), Ahrefs, or SEMrush. These tools can show you what keywords and phrases you’re showing up for.
In a previous episode, we talked about asking questions like:
– Who’s the procurement manager for this company?
– Who’s the CEO?
– What are the refund terms?
Often, this kind of info doesn’t even come back in results. And honestly, I’ve never seen anyone include those as part of their keyword tracking.
So maybe it’s time to think differently:
What questions do people ask about your business? Create a whole separate set of keywords around that.
The point is—what’s out there about your business? Is it accurate? Is it damaging? And then go from there.
Bill Scully (13:44)
That goes back to the core issue—who owns your answers?
Even if you made a list and ran it through your SEO tools, you’d get rankings, sure—but not the actual answers.
You might be ranked number one in AIO, but you still wouldn’t know what that answer says.
These LLMs are making things up. And there have been so many examples coming out in just the past couple of weeks.
It’s like everyone’s jumping on the bandwagon. The pendulum has swung. First, it was the greatest thing since sliced bread—now it stinks.
Though to be fair, it can still be good.
Bill Hunt (14:27)
Right—now everyone wants to bash it.
Bill Scully (14:39)
I’m going to change my thought here. I almost don’t trust AIOs anymore—I always feel like I need to verify what they generate. I was going to say it’s good, but what I’m actually thinking is that they’re better for having conversations, like using chat tools.
I can’t imagine how many people are just copying and pasting that content to build pages. It could actually start to retrain what the internet thinks is consensus—especially if people are generating hundreds of similar pages with inaccurate information.
Bill Hunt (15:13)
Yeah, blame me.
[Laughs]
Yeah.
I mean, I can’t even tell you how many projects or requests I’ve had over the years from companies—especially when it comes to things like… what’s that feature on the right side? The Knowledge Graph, where it pulls in a phone number—sometimes the support number—and companies will say, “That’s not our support number. That’s our main number.”
Bill Scully (15:39)
Knowledge Graph.
Bill Hunt (15:52)
Exactly. And they don’t want support calls going to that number because it routes to a different system. So they ask us to fix it. But where did Google even get that number? Usually from some random directory—or someone who carelessly added it to the schema markup without using the proper support schema.
Then you’ve got the opposite situation. I’ve seen companies that hired someone to build their schema, and that person just arbitrarily added a support number. That number then shows up in the Knowledge Graph, and suddenly the company is panicking: “Get that number out of there!” Their support calls spike 5,000%, costing them a fortune. They want that number buried—not promoted.
So all of this matters. And if we’re not reviewing this stuff—like we used to with rich snippets—it becomes a problem.
In Data Prism, we used to pull back all the data to see:
A) Is there a rich snippet for this question?
B) Do we have it?
If we did, that would prompt us to take a closer look—whether scraping the data or reviewing it manually—and then compare it to what we want to be shown. If you’ve got an answer management database, like we’ve been promoting, you can cross-reference it. But that does require extracting data or having someone look at it directly.
Again, it’s all about identifying what’s most impactful for your business. You know, 80/20 rule.
And that reminds me—in the article we wrote on answer management, we gave an example. You and I were talking about those CO2 monitors. The one I ended up buying wasn’t even listed when I asked, “What’s the best one?” That’s when I realized the AI was actually scripting logic to define what “best” means.
Bill Scully (17:50)
Yeah.
Bill Hunt (18:16)
If you read that article, you’ll see—my device didn’t make the list. So I asked, “Why not?” The AI responded that it didn’t find many sources referencing the one I bought. Then it walked through about 11 reasons why it wasn’t included—things like certifications it lacked, etc.
Each point made me think, “Okay, I need to promote this product better, make sure it’s represented.” Ironically, in the end, it was the best option for me. But the content that did rank was from pages I’d describe as overly spammed. It was like reading a recipe: “This is why Grandma liked it, and this is why she used this particular brand of brown butter.”
Every imaginable angle was baked into this enormous, 200-page blog post about CO2 monitors. And that’s what the AI used—because it had all the “signals” it looks for.
So yeah, companies really need to pay attention to what’s being surfaced.
Bill Scully (19:15)
Hmm.
Bill Hunt (19:44)
You know, try to source where that information is coming from. Are you the source? If so, you can obviously correct your own content. It’s not like you can just say, “Hey, let’s tweak this page,” and it’s done. You have to wait for those cycles. So this is something you’ve got to act on very quickly.
Bill Scully (20:01)
One thing I haven’t done myself yet—but it came up in Search Central—is the feedback feature. There’s actually a thumbs-up and thumbs-down option, and you can provide feedback for inaccurate answers.
Bill Hunt (20:17)
Exactly. When you click the thumbs-down, it opens a feedback window. I don’t think we made this point clearly enough in our previous podcast, but what Danny said was really important. He mentioned that when he sends feedback—because he’s in social media—it goes to the right people.
The reason that feedback is so valuable is because every thumbs-down, along with your comments, goes into a database. So for certain answers in certain contexts—like you said, it’s like everything—if there’s an increase in negative feedback on something, it flags a review. That might not mean they’ll fix a specific entry, but they will assess it in aggregate.
As I mentioned in the article we wrote, it’s important to give feedback early and often when something’s off. Especially with Google’s AI-generated answers—if something’s wrong, say so. Repeated feedback like, “This is wrong,” increases the chance that it gets flagged for review.
And remember, the key difference between Google, Bing, and others is that Google has its own search engine to draw from.
One thing we didn’t even get to—that was my second rant—is schema. They mentioned three things about schema. First, they use it. Bing confirmed they use it in Gemini. Google is definitely using it too.
What was interesting is this: they said they use schema because, A, they have the infrastructure to extract it, and B, they have the infrastructure to use it. And the more sites that implement schema, the more valuable it becomes to those processing it.
If there’s a schema element reported on in Search Console—if it’s in the left-hand rail—they’re using it. Like he said, “If we’re telling you in our help docs and in Search Console that we use it, we use it.”
He gave a great example related to shopping. They often get criticized when product prices are wrong. But if your site provides schema that wraps the price in a clear format and specifies the currency, they’ll use that over something they scraped.
So again, do your part.
Bill Scully (22:51)
Yeah.
Yep. Yep.
Bill Hunt (23:18)
Make sure the information out there is correct—at least within your own platform.
Bill Scully (23:25)
Yeah. I kind of think of these things like a lawyer. I assume the schema comes back as “reg,” right?
Bill Hunt (23:36)
What do you mean, “red”?
Bill Scully (23:38)
Reg… like regulatory?
Bill Hunt (23:41)
Ah, yeah.
Bill Scully (23:42)
The LLM scrapers aren’t actually scraping the schema.
Bill Hunt (23:52)
Correct. The LLM scrapers are not—and he didn’t confirm it specifically—it’s just implied.
Bill Scully (23:54)
Right. I figured as much. It’s code, and it would be expensive to extract schema from the code.
Bill Hunt (24:01)
Exactly. Let’s play both sides of this. In the article, I addressed this specifically. One of my big takeaways from that event was this: schema is computationally more efficient.
Bill Scully (24:03)
Okay. Yeah.
Bill Hunt (24:19)
He said—and I’ll write a separate article on this because I’m passionate about it—Google uses schema. Google uses schema both in traditional search and in Gemini. They use it because it’s computationally efficient and more accurate than trying to infer things on their own.
Google’s got two parts to the equation: the search infrastructure and schema data. If you understand how search fundamentally works, every topic is tokenized with variables. From a developer’s perspective—especially a database developer—you’ve got an entity and a validation layer. If that’s tied into a retrieval-augmented generation (RAG) system, it can bind and validate data within an LLM.
That works for Google and Bing. It doesn’t work for other platforms.
So you nailed it. For models like ChatGPT or Perplexity, it’s not computationally efficient to extract schema. They’re already optimized for deduction and inference. So for them, scraping schema is counterproductive.
This is where our industry drives me crazy—someone says something, and suddenly three “white hat” purists run with it. They say, “LLMs don’t and never will look at schema.” That may be true for some, like ChatGPT. But that doesn’t mean no one does.
And if you’re getting any business value from Google or Bing, you should be using schema.
People often ask me, “Should we use schema since LLMs don’t use it?” My response is: “Do you get value from Google?” If yes, and if it’s in the left rail of Search Console, use that schema.
FAQ schema, for instance—what better way is there to validate an answer? Think about it from Google’s point of view: there’s a question, a vetted answer, tied to a business. They’ll trust that over scraped guesses.
Bill Scully (26:59)
Yeah.
Bill Hunt (27:10)
Now I can compare that to everything else out there. You used the word earlier—”consensus.” Is there consensus that this is the right answer? Let me add another variable to the logic: is it coming from the source? If we’re talking about blue widgets from Acme, and Acme itself says it’s X, Y, and Z, that’s probably the most accurate information.
Bill Scully (27:18)
Yep.
Bill Hunt (27:40)
Especially if it’s not marked up with schema. Compare that to something I’ve pulled from Joe Bob’s blog or a random reviewer. From a Google or Bing perspective, authoritative sources add a lot of value. That’s why, in an answer management program, we recommend tagging and structuring this information properly.
We advocate creating modules or blocks—I think the CMS crowd calls them blocks and blobs—that can be injected into webpages. If we’ve built a Q&A or FAQ block as part of our answers program, we can insert that wherever relevant. Think about the kinds of things people are searching for: temperature ranges, links…
Bill Scully (28:36)
Pricing.
Bill Hunt (28:38)
Exactly—pricing. Just the other day I was looking for the weight or dimensions of something. I found 15 different answers. Eventually, I stopped being lazy and went straight to the company’s specification page.
So again, think about what’s being asked and how it’s being answered. You’re never going to cover everything—there are too many possibilities—but focus on what’s most important: price, sensitive information, etc.
For example, you probably shouldn’t list your procurement manager’s name—they’ll just get spammed. But your support number? Absolutely. Other key business details? Definitely. I really like the idea of identifying whether my product is in the consideration set and asking the important question: why isn’t it being considered?
And when things are wrong, find the source and get it corrected there if you can. If someone’s made up content about your business, they’re probably not going to change it—especially if they benefit from being seen as the source of truth.
But if you’re the better source of truth, you need to surface that. Let me give you an example: “Is there gluten in vodka?” Scientifically, no—it’s removed during distillation. But is there gluten in the wheat used to make it? Yes. And legal teams often won’t allow you to say it’s gluten-free because—who knows—gluten might waft in during bottling.
Bill Scully (30:40)
Yeah, one person in 100,000…
Bill Hunt (30:49)
…might react to gluten just because it was in the air. It’s like peanuts. If you’re 20 miles from a peanut packaging facility, you still have to say it may contain peanuts. I’m not criticizing that—just pointing out that legal has different constraints. Meanwhile, someone not bound by those rules might boldly claim “no gluten” and get the top ranking. Maybe it doesn’t break your business, but it could influence perception.
So apply the 80/20 rule: what are the key things people need to know? Do you provide them? And more importantly, how do we structure this as an argument? Bill, I think you touched on this: when people glance at a Google answer, there’s a reasonable expectation that it’s accurate.
I wonder if there have been legal challenges to that assumption—that these answers should be reliable. If people skim and get incorrect info about your company, how do you regain control? Should you publish more? Structure content better?
I’m always surprised in B2B: the information people want—like specs—is often buried deep on the site. It’s like the fifth tab on a product page. I had an epiphany working with a large manufacturer. We discovered the content users wanted most was the lowest priority internally. The creative team didn’t see its value.
But if you’re trying to improve this, the word we should focus on is trust. Like you said, you can’t just trust any answer out there. You have to verify it. But trust is the new currency. Now that everyone’s aware of hallucinations and errors in AI, the importance of trust is only going to grow.
Bill Scully (33:31)
It’s really interesting. Four years ago, if you ranked number one on Google, you were trusted as the authority—even if you were a mom-and-pop shop in Podunk, Iowa. But today? I bet most people in search don’t trust search. Most marketers don’t. I don’t even know what my wife thinks—but she probably shouldn’t trust it either.
Bill Hunt (34:05)
Have we been burned by it?
This goes back to something we talked about offline—the idea I’ve been trying to float: some things should require a level of authentication. My rant around this is about visas. Especially for expensive or complex visas, you suddenly see hundreds of companies offering to “help” with the process.
For example, take Japan. For the most part, you don’t need a visa to visit. So there’s absolutely no paid search for it—no paid vendors on the first page—because there’s no money to be made. There’s only a basic pre-entry form, like the one we have in the U.S., and just a couple of sites try to assist with that. But because there’s no visa involved, there’s no revenue opportunity—so no one’s buying search ads.
Now compare that to the new UK entry permit. Suddenly, there are four paid ads—each with imagery that looks very official, almost like the actual UK government.
Bill Scully (35:16)
Okay.
Bill Hunt (35:39)
We talked about this in a previous episode—no one is vetting these ads. Some of them look too much like official government websites. Someone should have flagged them. But they’re still there.
Bill Scully (35:52)
To Google’s credit, I saw something from Brody Clark—the ratings, not reviews. Google is now starting to display ratings differently based on what they perceive as the trust level. If they trust a source, they’ll show the star rating with the number of reviews.
Bill Hunt (36:15)
Like on OutSchool.
Bill Scully (36:23)
Exactly. If they don’t trust you, you’ll just see the star image—no review count beside it. Like you said, even Google’s trying to figure out how to validate and verify trustworthiness.
Bill Hunt (36:46)
Yeah, and it’s both simple and complicated. I remember when my son worked at TripAdvisor—they made public statements about how they detected fake reviews. I don’t know if they ever patented their methods, but they definitely had systems in place to identify inauthentic content based on patterns.
Google’s got serious brainpower. I’m sure they’ve developed systems for this too. And it all comes back to trust.
Think about it: when someone gets scammed, who do they blame? ChatGPT? The embassy? Google?
Take the phone number scams. I’ve had four or five people tell me—especially in Amsterdam—that they missed a connecting flight and searched for airline customer service numbers. They’d Google “KLM customer service,” click the number that came up on their phone, and it turned out to be a fake call center.
The scammer would say, “Sure, we can reschedule you. It’ll be $800.” And in the moment—under pressure—people would pay. They’d show up at the counter and find out: they’re not even booked.
So who do they blame? Google—for showing a scam number. And the airline—for not making their real contact info easy to find.
In my case, I have a special number on my elite status card, so I know where to call. But the average person? They’re relying on Google. And again, it all circles back to trust.
We need to think about this in our own programs. And before we wrap up—I know we were trying to keep this one short because I’ve gone long on this before—I want to bring up the second part of this conversation…
Bill Scully (39:20)
Hmm.
Bill Hunt (39:37)
I want to talk about something Dixon brought up regarding incorrect URLs. Now, we can’t do much about misattributions—if Bill Scully writes something and credits Bill Hunt, we can’t do much other than go to the source and ask for a correction. But take a look at this specific case Dixon pulled up: this hallucination where the AI gave an answer and linked to an IKEA page that was /chairs. When you clicked it, though, you got a 404 error.
This is a key point for web effectiveness. It’s something everyone should look out for: common 404s that should have a redirect in place. Here’s what I recommend: check for 404s coming from chatbots. You won’t get these errors from Google, because Google doesn’t differentiate between AI-generated answers or not. We don’t get the actual query, so we don’t know what caused the 404. But for all the other sources, you can check the referral URL that led to the 404 page.
For example, if a lot of people are hitting that “chairs” page at IKEA, maybe they should redirect to their armchairs page, or their category page for sofas and armchairs. We could discuss this logic more later, but IKEA has a ton of data to make these decisions. The key takeaway here is to check if people are being referred to the wrong page—maybe it’s an old article you wrote a few years ago that doesn’t exist anymore and didn’t get a proper redirect. If it’s still in the training data of a chatbot or another source, it’s still sending people to the 404. This is something I think we all need to look at carefully as AI tools become more common.
Bill Scully (41:59)
Yeah, so one thing I’ve done is set up a 404 error report from GA4 for all my accounts. I look at it monthly, and if I see any URLs with more than six or seven 404 hits in a month, I’ll add redirects. That’s my simple way of handling it.
Bill Hunt (42:24)
Perfect.
Bill Scully (42:29)
That’s the cheap way to handle it, but I know you have some great ideas for tackling this issue. We can talk about them in a future episode.
Bill Hunt (42:32)
Yeah, exactly. But this is a real problem—it’s affecting trust, confidence, and even your business. You could be sending people to the wrong page or giving credit to the wrong source. The key is focusing on the most critical areas—those high in the funnel, those that impact revenue most. We need to think about how we can integrate proper answer management here. Listen to the episode and check out some of the resources we’ve written about answer management, and see if they can help you with this issue. It’s gaining attention now, so if your boss asks about it, you should have a plan. If you need help or have questions about this, especially around answer management, feel free to reach out. Join our Facebook group, ask us questions, and let us know if you want us to dig deeper into any part of this.
Bill, any final thoughts or recommendations on next steps?
Bill Scully (43:57)
So, I’d say use generative AI as a starting point for writing, but don’t just copy and paste without verifying sources and cross-checking links. I love the idea of having a VP of Answers—someone who ensures that the right answers are being used. As we discussed today, it’s essential to make sure that the answers in AI overviews are accurate. Hallucinations are a big issue—people are encountering them more often. I saw a stat from Search Engine Land that said up to 28% of answers could be wrong, which is alarming.
Bill Hunt (44:51)
Very scary. Dixon’s data was insightful.
Bill Scully (44:51)
Yeah, just five months ago, I would have said maybe 3%. But now, the answers are so convincing—if they’re wrong, it feels like a lie because they’re so close to being right.
Bill Hunt (45:11)
Yeah, I like how you put it—it’s almost like a lie when they’re wrong.
Bill Scully (45:15)
Alright, I think that’s a wrap. What do you think? Does it sound like a wrap?
Bill Hunt (45:22)
Yeah, I think so.
Bill Scully (45:51)
We hope you find value in our conversations. If you’re enjoying the show, we’d really appreciate it if you’d take a moment to rate and review the podcast on your favorite platform. We also want to hear from you on social media. Please go to Facebook, LinkedIn, X, search for International Web Mastery, and tell us what you think, or what we can do in future episodes to help you.