Gemma 4, Google's Core Update, and Why Your Search Console Is About to Look Weird

📬 Love the podcast? You’ll love the newsletter.
 Get the weekly 3-2-1 on AI search + marketing: Subscribe

Google released Gemma 4 this week, and while the tech press is treating it as a developer story, Cassie argues it's actually one of the most important content strategy stories of the year. Open-weight models mean companies can download a frozen snapshot of Google's AI and deploy it into internal chatbots, customer support tools, and vertical search products that run for years — each one carrying a frozen version of how the model understood your brand on training day.

Cassie also unpacks a double whammy hitting Google Search Console right now: a long-running impression-logging bug dating back to May 2025 that's finally being corrected, and the March 2026 core update, which completed its 12-day rollout the morning of this recording. Together, those two events — plus the recent spam update — are about to make your data look weird for reasons that have nothing to do with your content strategy. The episode wraps with quick hits on Anthropic's Claude Mythos Preview and Project Glasswing, plus Microsoft's viral Copilot Terms of Service moment.

In This Episode:

• Why Gemma 4 is a content strategy story, not a developer story, and what "open-weight" actually means for how AI tools get built on top of Google's models
• How every open-weight model release creates a frozen snapshot of the internet that gets deployed into internal chatbots and enterprise tools for years after training ends
• Why enterprise software upgrade cycles mean a Gemma-powered chatbot built next month could still be shaping buyer perception of your brand in 2028
• Why Google Search Console has been over-reporting impression counts since May 2025, and what to do when your numbers drop over the next few weeks
• How to interpret Search Console data when three Google updates — the spam update, the March 2026 core update, and the impression bug fix — all hit in the same window
• Anthropic's Claude Mythos Preview and Project Glasswing: a quick look at the model Anthropic decided not to release
• Microsoft's "Copilot is for entertainment purposes only" ToS moment and what it says about building long-term strategy around any single engine

Let’s connect:

LinkedIn → Cassie Clark | Fractional Content Strategist
Website → https://cassieclarkmarketing.com

Download Freshness, Structure, Authority: The Framework for AI Search Visibility:

Amazon

P.S. Is your brand losing its "Answer Authority"?

Most series A/B and enterprise brands are being "nudged" out of AI search results because of entity gaps and "stale" content. I am opening a limited number of specialized audit slots to help you reclaim your Share of Voice using the FSA Framework (Freshness, Structure, Authority).

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

Transcript

SPEAKER_00 1:13

Hey marketers, welcome back to another edition of AI Search News. This week, Google handed the open source community a gift that's going to reshape how brands get discovered for the next three years. That could be a guesstimate, but we need to talk about what I mean here. Google is also about to show you a bunch of scary looking numbers in Search Console for two completely different reasons, maybe three, and I need you to not panic about either of them. Plus, we have some quick hits on Anthropic and Microsoft to wrap all of this up. Hey, I'm Cassie Clark, a fractional content strategist and an AI search optimization expert, and you're listening to Found and AI, the show that helps marketers and founders make sense of AI Search so you don't get left behind in this new wave of user search behavior. Today is April 9th. Let's get into it.0 license. The 31 billion parameter version ranks among the top open models globally. The smaller versions can run locally on consumer hardware, and because of the license, companies can use it fully commercially. That means they can fine-tune it, they can modify it, and they can deploy it into their own products. Now at first glance, this really does sound like developer news. We are marketers, why do we care about this one? Well, because there's a content strategy story hidden in disguise. Let me explain why I think that. So when Google or Anthropic or any of these AI companies release an open weight model like Gemma 4, they're releasing a frozen snapshot of everything that model learned during training data. Whatever the open web looked like the day that Google stopped training, well that's the version of the internet that lives inside those weights. Forever. Or at least until the next training run, which could be six months, a year, longer, who knows. Here's the part that I think really matters, and that we really need to pay attention to. Every company that downloads Demo 4 and builds something on top of it is deploying that specific snapshot into the wild. They could use it to build internal chatbots, customer support tools, procurement assistance, research agents, vertical search products, who knows, like the possibilities are inlets here. But each one of those tools is carrying a frozen version of how Google's model understood your brand on training day. Enterprise software doesn't get upgraded the way that consumer apps do. So a company that downloads Gemma 4 today and builds an internal chat bot on it might not or they might still be running the same model two years from now, not really even updating it as Gemma 4 get updates. So that means whatever snapshot of your brand that got baked into training time could be shaping buyer perception for years. Inside tools you will never see and never get analytics from. So here's how I want us to think about this one. Gemma 4 is a model launch story, but there's really more to it. It's a reminder that every piece of content you publish is training data. I will say this forever. Everything is training data. And it could be training data for a model that hasn't been built yet. So Google froze a version of the internet to make Gemma 4. Some company is going to download it in the next month, they're going to build a chatbot, they're going to run that chatbot for the next three years. Whatever that chatbot knows about your brand is whatever you put on the open web before the freeze. You don't get to go back and edit it. It is done, it is solid, is what we're using. Now I know what some of you are thinking. You're thinking, hey, but Cassie, companies can connect these models to live retrieval systems, right? There's RAG, there's web search APIs, there's connected knowledge bases. That's the freshest correction. And yes, that is absolutely correct. That is how a Gemma powered chatbot can cite something that you published yesterday. But retrieval only works if the model already knows your brand well enough to service you as a relevant source in the first place. The weights decide whether you're in the candidate pool. Retrieval decides which specific pages get pulled. So you need both layers working together. Entity strength in the training data gets you into consideration. Structured, fresh, extractable content on the open web gets you cited at retrieval time. This is where the FSA framework gets a second dimension. Freshness, structure, and authority are not just a live retrieval moment anymore. Authority, especially in context of the next freeze, is a thing that we need to be worried about. So every podcast that you guest on, every Reddit thread your brand gets mentioned in, every third-party citation, every expert quote that you put out there, that's not just helping you show up in chat GPT answers today, it's shaping how the next generation of models will understand who you are and what you stand for. So when people ask me, hey Cassie, is AI search optimization worth the effort? If the landscape is changing so quickly, this is my answer. The content that you put out this quarter is not just for this quarter. It's training data for models that will shape buyer decisions for years after you've forgotten you've published it. Build before the freeze. Everything is training data. Okay, so now that that one is out of the way, we need to talk about Search Console. Two things happen this week that are going to make your data look weird, and I want you to hear about both of them for me before you log in. Look at yourself and then just panic. So thing one, on April 3rd, Google announced they're fixing a long-running bug in Search Console that was inflating impression accounts. Now, when I say long running, I mean it the bug started on March, not March, May 13th, 2025, and has been over-reporting impressions for almost a full year. Here's what Google said about it. They said a logging error prevented Search Console from accurately reporting impressions from May 13, 2025 onward. The fix is rolling out over the next few weeks, and as a result, you're going to see a decrease in impressions in your performance report. Clicks were not affected, other metrics were not affected. This was strictly an impression logging error. So, if your impression numbers drop over the next few weeks, it is almost certainly not because you did something wrong. It's Google correcting nearly a full year of inflated data. The number you're seeing now is more accurate than the number you were seeing before. And it means that the benchmarks that some of you have been measuring yourself against since last May might have been slightly off the whole time. Now, thing two, this one happened literally today as I'm recording. I record these new update news updates on Wednesdays because this goes out on like 6 a.m. on Thursday. So technically today is Wednesday, but whatever. Anyway, Google confirms that the March 2026 core update has finished rolling out. It started on March 27th and wrapped up at 6 12 a.m. Pacific time on April 8th. That's about 12 days, which is faster than the December 2025 core update, which took 18 days, and it is ahead of the anticipated date, which was April 10th. Google called this one a regular update designed to better surface relevant, satisfying content for searchers from all types of sites. It's just a standard core update doing its standard core update thing, which is reassessing content quality across the web and then moving pages up and down accordingly. Now here's why I'm pairing both of these things together. Over the next couple of weeks, a lot of you are gonna open Google Search Console and you're gonna see your numbers are gonna probably look a little wonky. You might see your impressions drop, you might see those ranking shifts, you might see a click pattern change, and your first instinct is gonna be to try to figure out which thing caused which change. Was it the core update? Was it the bug fix? Did we get hit with the spam thing? Did our competitors get boosted? Should I rewrite everything? Those are the questions that are they're gonna run through your mind. Before you do anything, slow down. Trying to diagnose changes when multiple updates are overlapping. You cannot separate what happened. You cannot clearly separate those things. The core update rolled out from March 27th to April 8th. The March spam update wrapped up on March 25th, two days before the core update started, and the impression bug is still rolling out right now on top of all of it. That's three sources of data noise in the same window. So before we do anything else, here's what we're gonna do. First, screenshot your current Search Console data before the impression fix fully rolls out. You want a record of that pre-correction numbers and what they looked like so you can compare against corrected numbers later. Second, we're gonna follow Google's own advice on core updates. We're gonna wait at least a full week after April 8th before drawing any conclusions. Your baseline should be the weeks before March 27th. Compared against performance after the dust settles from all three of these updates. Now, third, the whole situation is a good reminder. It's a reminder that the measurement layer is just as fragile as the retrieval layer. Google Search Console is the tool that most of us rely on to understand how our content is performing. I look at mine nearly every day, but software. Software has bugs, and sometimes those bugs run for 11 months before anyone notices. And when a core update and a data correction and a spam update all hit within the same two-week window, whatever story your dashboard is telling you is partly real and it's partly noise, and it's impossible to tell which is which in real time. Which is why I keep telling clients: do not measure visibility with one tool. Cross-reference everything. Look at your direct traffic, look at your branded search volume, look at your citations and AI answers, your mentions and third-party sources, your pipeline specifically, and your lead quality data. The more data points you have, the less vulnerable you are when any single dashboard goes sideways, like Google Search Console, whether it's a bug or an update or both all at once. Now, a real quick um update on the core update itself. Core updates reassess content quality across the web. They are not targeted penalties, so if your rankings drop, it does not mean that you violate a policy. It just means Google systems are weighing quality signals slightly differently than they were last month. Some pages are moving up, some are moving down. The response to the core update is to never panic rewrite your content. The response is to keep building clear, well-structured, authoritative content that serves your reader, which is exactly the same thing that gets you cited in AI answers. SEO and AI Search are not separate disciplines. They overlap, they're rewarding the same fundamentals. Okay, so before you go, two quick hits that I want to mention because they're interesting, but they're not the main story. First, Anthropic announced a new model this week called Claude Nathos Preview. This one's fascinating because it's so capable at finding security vulnerabilities that Anthropic decided not to release it broadly. Instead, they launched something called Project Glasswing. I think I said that correctly. It's giving limited access to a small group of partners like Amazon, Apple, Google, Microsoft, and about 40 other critical infrastructure organizations for defensive cybersecurity work only. Those are the only ones that get access to this. Now, reports are saying that the model has already found thousands of zero-day vulnerabilities, including a 27, a 27-year-old bug in OpenBSD. Now, that is a wild story, but it's really a cybersecurity story more than a search story. So I'm gonna leave a deep dive to this to the infos InfoSec podcast. Like if you want to read more about it, Anthropic has a blog post up about it, but that it was interesting and I wanted to touch on it. Now, second, this one kind of made me laugh out loud a little, I'm not gonna lie. Microsoft's Copilot Terms of Service, they went viral this week because someone finally read the fine print. Like we all skip it, we know this, but someone finally read it and they noticed under a section labeled Important Disclosures and Warnings that the term of service literally says, Co-Pilot is for, and I quote, entertainment purposes only. Do not rely on co-pilot for important advice. End quote. This is the same copilot at Microsoft sales as a productivity tool for up to$30 per user per month per month. Microsoft's response was basically, oh, that's just legacy language from the chat from the big chat days, so we'll update it eventually. And as of a couple days ago, they still had not done that. Now, small point I want to make here, the engines themselves are still figuring out what they are. If Microsoft's own lawyers haven't caught up to how people actually use Copilot, what chance does a marketer have building a long-term strategy around a single engine? Don't do that. Don't do that. Build your strategy around principles that travel across every engine. So that freshes, the structure, the authority. Let those principles show up wherever the retrieval servants happens to be. Don't rely on just one single model because things are changing all the time. Okay, so that's this week. Jem Afor reminded us that content is training data, not campaign assets. Google's about to make your search console look kind of weird for three different reasons at once. So screenshot your data, don't panic, we'll just reference it later. Anthropic decided that some capabilities are far too powerful to ship to the general public. And Microsoft flyers, well, they think that Copilot is for entertainment. If this episode helped you make sense of the noise, hit subscribe and leave a review. It really does help the show reach more people. And if you're trying to figure out how your brand shows up in AI generated answers right now and at the next training freeze, head over to CassieClarkMarketing.com and book an AI search visibility audit. We'll look at exactly how the AI engines see you today and what to fix before the next generation of models bakes it into their layers. Alright, I will see you in the next episode. Until then, stay visible.