Fable, Bing's New Reporting, and a Study That Should Worry Every B2B Marketer

📬 You like this podcast? You’ll love the newsletter.
Join the weekly 3-2-1 on AI search + marketing: subscribe

The US government just ordered Anthropic to pull Fable 5 and Mythos 5 — for all users, not just foreign nationals. A model available to hundreds of millions of people went dark mid-afternoon on a Tuesday with no advance notice. Not because it stopped working. Because of a government directive citing a potential jailbreak.

That's the story I didn't expect to be covering this week. But it's also the clearest argument yet for why model-specific optimization is a strategy built on sand.

In this episode:

• What actually happened with Fable 5 and Mythos 5 — and why this matters more as a strategic signal than a product update
• Why building your AI visibility strategy around any specific model puts you at risk you can't control
• Bing Webmaster Tools' four new AI reporting features — Intents, Topics, Citation Share, and Compare — and how to actually use them
• How to map the new Bing reporting to the FSA Framework (Freshness, Structure, Authority) so the data tells you something actionable
• Why AI visibility still has no click-through data, and what we're measuring instead
• A new study of 20,000 ChatGPT responses found that 80% of product recommendations change when search is enabled — and why that number should reframe how you think about content freshness

If you're listening to this and thinking I need someone to lead this for me, that's what I do.

I'm an AI search visibility consultant and a fractional content strategist for startups and enterprise brands. If that sounds like the kind of help you're looking for, email me at cassie@cassieclarkmarketing.com. 

Or request your 7-Day AI Search Visibility Audit: https://cassieclarkmarketing.com/ai-search-visibility-audit/

Let’s connect:

LinkedIn → Cassie Clark | AI Search Visibility Consultant
Website → https://cassieclarkmarketing.com

Transcript

SPEAKER_00 0:00

Hey, welcome back to Found in AI. I'm Cassie Clark, a fractional content strategist, an AI search optimization expert, and the host of the show where we talk about what's actually happening in AI Search, GEO, and AEO, so you're not left scrambling when the ground shifts beneath your strategy. And this week, the ground shifted once again. We have three things to get into today. First, there's a quick note on Claude Fable, that's Anthropic's newest model, and why you probably can't access it, and why that matters more as a strategic signal than a product update. Then Bing Webmaster Tools just dropped a meaningful expansion to its AI performance reporting. We're gonna break down what's actually useful there. And then finally, there's a study that ran 20,000 chat GPT responses and found that 80% of product recommendations change the moment that you turn the search feature on, which is a number that should make every marketer stop and think about this. Alright, let's get into it. Alright, the first story is Fable 5 and Mythos 5. Now I want to be really upfront here. I'm not gonna do a deep analysis of the story. It's still developing. I'd rather just give you the facts and speculate, but here is what happened. On June 12th, Anthropic published a statement saying that the US government issued an export control directive to suspend all access to Fable 5 and Mythos 5, not just for foreign users but for all customers. Now, this directive arrived at 5.21 p.m. Eastern Time and required Anthropic to disable both models immediately to ensure compliance. Now, the stated reason, according to their announcement, is that the government believes it has become aware of a method to jailbreak Fable 5. Anthropic reviewed what was shared with them, disagrees with the severity assessment, and is complying under protest while working to restore access. Their position is essentially that the jailbreak method demonstrated is narrow, non-universal, and the same kind of capability already available from other publicly deployed models, including GPT 5.5. Anthropic was pretty pointed in their statement. They said, and I'm paraphrasing that if this is the standard that pulling a commercial model over a narrow potential jailbreak was applied consistently across the industry, it would essentially halt all new frontier model development. Like I said, I am not going to analyze this one beyond the facts, it's entirely too early for that. The legal and policy dimensions are real and they're complicated. I will follow up on that when there's more clarity. But I will say this. If you needed a concrete reason to stop building your AI strategy around specific models, this is it. A model that was available to hundreds of millions of people can be pulled within hours. Not because it didn't work, not because of a product issue, but because of a government directive that arrived mid-afternoon with no advanced notice. That's not a criticism of anthropic, it's just the reality of operating in this space right now. The AI landscape is unsteady at best. Not in a things are changing fast kind of way, but more in a model that you are building around can disappear before dinner kind of way. And I know that sounds alarming, but here's the takeaway. Your strategy cannot depend on any single model. These things are just updating and changing way too fast. So the question isn't which model that you're optimizing for, the question is whether your content is structured so that any capable AI system can read it, extract from it, and trust it. Whether your brain has enough presence across the web that you're not invisible when one system goes offline and then another one fills the gap. That's the FSA framework doing exactly what it's supposed to do: freshness, structure, and authority. None of those things disappear because Fable is temporarily unavailable. And to the authority that you've built doesn't need a reset, and structured content doesn't stop being retrievable. The fundamentals don't get government ordered offline, so just keep building those things. Okay, so moving on to Bing. Bing Webmaster Tools just previewed a meaningful expansion to their AI performance reporting. This started rolling out globally on June 16th, and if you've been following the show, you'll remember that I covered the original AI performance report back in February, which was the first time any major search engine built native infrastructure specifically for tracking AI-driven citation activity. Google's AI reporting and search console did not arrive until June, and Bing beat them there by months. Like, I'm gonna keep harping on that because I just I think it's kind of funny. But now Bing is adding four new features on top of that foundation. They've added Intense, Topics, Citation Share, and Compare. Let's walk through each one of those. So Intense it classifies the grounding queries in your AI performance report into categories. Those categories are informational, commercial, navigational, research, creation, local, and then a few others. A grounding query, if you're new to that term, is the phrase the AI model used when it went to retrieve your content. So intents tell you why your content was pulled. Are people getting cited for informational lookups? Or is it more commercial intent, like those comparison queries, purchase stage questions? This matters because it tells you whether your content is showing up in the right part of that buyer journey. Topics, it groups those grounding queries into thematic clusters. So instead of seeing a hundred individual queries, you start seeing those patterns. Maybe it's solar energy or B2B content strategy as a cluster rather than 20 slightly different versions of the same question. Microsoft's point here is that AI systems actually reason across themes, not keywords, and those topics reflect how the model is thinking about your content, not just the words that triggered a citation. Then we have citation share, which shows what percentage of citations your site received for a specific grounding query out of all the citations shown for that query. So if 10 sites are being cited for something like best CRM for small teams, and you're showing up in 40% of those citations, that's your citation share. It's directional, it is not traffic, doesn't tell you about competitors by name, but it does tell you whether you're concentrated in one URL or building broader authority across your site. Microsoft was very specific. This is an observational metric, not a ranking system. The citation ecosystem is dynamic, it's constantly changing, and those citation patterns they shift based on freshness signals, model updates, user behavior changes, which is exactly why we have the compare feature. This is the one that I actually find the most useful. It lets you overlay a previous time period onto your current reporting so you can see whether your citation activity is growing, shrinking, or shifting across topics and intents over time. Before this, AI visibility felt like publish, prompt test, guess, write it down, come back to it later, do it again. Now there's a feedback loop in these tools, and those feedback loops are how strategies actually mature. If you're using the FSA framework, here's how those features map. Intents and topics validate your authority pillar. You can see whether the model associates your brand with the topics and use cases you're actually targeting. Citation Share tracks how that authority is distributed across your site versus concentrated in one piece and compares your freshness signal validator. If you made content updates and your citation activity went up in the subsequent period, that is useful data to have on hand. Do keep this one thing in mind though, Bing still does not have click-through data, but neither does Google's version. So we're still measuring proximity and frequency, not conversion. But the infrastructure is getting more sophisticated quickly, so I'm sure it's coming in the near future. Okay, last one, and this is a piece of data that I want to leave you with because it had me think in when I read it. Visibility Labs ran a study of 20,000 Chat GPT responses. They tested 1,000 product recommendation prompts. They ran each one 10 times with ChatGPT search enabled and 10 times with it disabled. The finding is that 80.2% of product recommendations changed when search was turned on. Only about 19.8% of products recommended without search also appeared when search was running. That's less than a 1 in 5 overlap. Here's where it gets even more interesting. For the products that ChatGPT recommended every single time when search was off, so the most consistently cited products, only 15.8% of those were carried over when search was enabled. So the products that ChatGPT was most confident about without the internet were the least likely to survive contact with real-time web data. The study also found that products mentioned more often in the sources ChatGPT actually cited tended to show up more in recommendations. There's a 0.4 Pearson correlation there, which isn't casual but it's directional. Appearing in the sources that an AI system cites seems to matter. Now, the study is about product recommendations, so it's mostly directly irrelevant for e-commerce and brands with a product to push, but the implications are broader. What this tells you is that ChatGPT without search is drawing from that training data. The training data is just that base level knowledge that ChatGPT or any of these AI models where we talk about it already has on hand and can draw from. Just kind of think about like a pool of stuff they already know. But that has a cutoff date, which reflects the web as it existed at some point in the past, and which may not accurately represent what's happening in your category or your industry right at this very moment. So the moment that you add live web access, everything reshuffles based on what's actually on the web today. That connects directly back to freshness in the FSA framework. If your content isn't being updated, if your brand presence isn't active, if the pages search engines are indexing are months or years old, then when ChatGPT with search goes looking for the best recommendation in your space, your content may not be what it finds. So what does all of this mean? It means that you can't treat AI visibility as a one-time optimization. It's not a build a great page, get cited forever kind of thing. It's closer to a publishing problem, which means it requires a content strategy, not just a technical fix. Although that does factor into it at some parts. If you're running an AI visibility audit and you see strong citations without search, but don't know whether those hold when search is enabled, the study is a reminder that those are two very different questions with potentially very different answers, and we need to be looking at both of those. Okay, so those are your three updates. There is one clear through line here. The space moves fast, and the strategies that hold up are the ones built on fundamentals, not chasing specific models or specific features. Master the fundamentals, that's freshest, structure and authority, and then understand how AI systems are retrieving your content, not just whether they are. And stay current because a recommendation that existed in training data six-ish months ago may not exist at all once that model goes online. Hey, I am Cassie Clark. If you want to go deeper on any of this, whether that's an AI visibility audit, working through what FSA looks like for your specific content program, or figuring out where your brand actually stands across these AI engines, you can find me at CassieClarkmarketing.com. There's more information in the show notes. That's it for this week. Stay visible. I will see you on Tuesday.