The Week AI Search Grew Up: A Third AEO, Google's AI Contribution Pilot, and the NYT Accuracy Study

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

Four stories this week, all pointing in the same direction: AI search is no longer experimental. A Google Cloud AI director just published a new optimization framework, Google is quietly piloting AI visibility reporting inside Search Console, the NYT dropped hard numbers on AI Overviews accuracy, and Ask Maps is reshaping local search around recommendations.

Cassie walks through each story, connects it back to the FSA Framework, and closes on the throughline she keeps coming back to: when infrastructure solidifies, behavior follows.

In this episode, you'll learn:

• Why Agentic Engine Optimization is a third optimization layer (not a replacement for SEO or GEO) and the 500-token rule for your highest-value pages
• What Google's AI Contribution Report pilot signals about where publisher reporting is headed—and why it matters even without screenshots yet
• Why "91% accurate" still means hundreds of thousands of wrong AI Overview answers every minute
• The ungrounding problem Gemini 3 made worse, and what it exposes about AI Overview manipulation on high-intent commercial queries
• How Lily Ray's fake-expert listicle test maps onto the Authority pillar of FSA
• What changes in local SEO when Ask Maps stops listing businesses and starts recommending them
• Why reviews, Google Business Profile specificity, and agency reporting all need to evolve with this shift

Whether you're a marketer, a founder, or an in-house SEO, this episode connects the week's biggest headlines to a clear next step for your visibility strategy.

Resources:

The FSA Framework on HubSpot

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, welcome back to Found in AI. I'm Cassie Clark, a fractional content strategist and AI search optimization expert, and the host of the show where we talk about GEO, AEO, and everything happening in AI Search so that we don't get lost in this new wave of user search behavior. Today is Thursday, April 16th. Last week I told you that I usually record the news updates on Wednesdays. Yesterday got it away from me. It was a long day on the road. So this one is coming at you fresh this morning and also a little later than normal. I'm sorry, I'm so sorry, but that's okay, we'll just listen to it later. This week's news kind of made me glad I waited though, because there's a lot to get into it. We have four stories to cover, and all four of them in different ways are telling us the same thing. The infrastructure around AI search is hardening. We're moving out of the is this even a thing phase and into well, here's how you measure it, here's how you optimize for it, here's how it goes wrong kind of phase. Here's what we're covering this week. First up, a director of engineering at Google Cloud AI published a new framework this week called Agentic Engine Optimization. Yes, that is now the third AEO. We'll talk about why this one matters here in a minute. Second, Google is also piloting an AI contribution report inside of Search Console. If you listened to the episode a few weeks ago where I geeked out over Bing's AI performance report, this is the Google answer to that, and it's gonna be a big deal once it rolls out. Three, the New York Times commissioned a story or study, not a story, on how often AI overviews are just wrong. It was mentioned on the headlines podcast a couple days ago. I usually listen to that while I'm getting ready in the mornings, and I thought, ah, that is super important. The numbers are worth talking about for sure, and so is the response from Google. And finally, Google's Ask Maps is changing how local search works. This one matters a lot if you have clients in home services, hospitality, or anything with a physical footprint. Okay, let's get into it. Okay, so story number one. This one dropped yesterday, April 15th. So again, I'm glad I didn't record yesterday because I would have missed it, and this one is a big one. Search Engine Lynn covered the story, Danny Goodwin wrote the piece on it. We need to talk about it. Addie S. Mani, who is director of engineering at Google Cloud AI, published a new framework that he's calling Agentic Engine Optimization or AEO. Now I want to take a second here because this is the third AEO we have floating around now. We have answer engine optimization, that's the one that most marketers mean when they say AEO. We have artificial intelligence optimization, which some people are using interchangeably with GEO, and then now we have agentic engine optimization, which is optimizing for AI agents specifically. Now listen, I know I am not super, super thrilled about this one either. Talking about the name, the framework is actually really useful. So if we can get past the name mess for a second and then really look at what he's saying, like that's where we need to focus. I, for one, will be super, super glad when the names are solid and no one is confused about what's what. Anyway, us Mani's argument is this AI agents don't consume content the way that humans do, which is what we've been saying for AI search optimization. But he adds, they don't even consume it the way that current AI search engines do. That's doing the heavy lifting here. An agent isn't opening a tab, it's not scrolling, it's not clicking around, it's not reading your intro. An agent is fetching, parsing, and acting all in one request, which means most of the engagement metrics we've been measuring for 20 years are irrelevant to that kind of visit. Now, here's a constraint that I want us to all pay attention to. Tokens. Every AI engine has a context window. It has a cap on how much content it can actually hold in memory while it's working. In the early days of Chat GPT, you probably notice this, you'd have a conversation, sometimes it would lose stuff. Same thing with your web pages. So if your webpage is too long or it's buried under navigation or scripts and sidebars when an agent visits, it will do one of three things. It will either truncate the content, skip your page, or just make up something. It will just hallucinate. Which, if you're a brand, is kind of terrifying because the agent might confidently misrepresent what you actually offer. So Osmani's recommendations are roughly these. Put that answer in there early within the first 500 tokens if you can. I like to think of tokens as like five first 500 words. Maybe it might be a little bit differently, but first little bit, put your answer there. Structure your content so that the agents and the engines can just lift it without leaving something important behind. That same advice applies here, just with the harder ceiling, because the agent literally cannot see past its context window. This is very consistent with what I'm teaching inside the FSA framework. For anyone new to the show, you can also read the full breakdown of the FSA framework over on the HubSpot Marketing blog. But TLDR version, keep pages compact and focused and skip those long preambles. I know we love a good opener, but the agents do not. They are busy. They are busy and just want the information to get out of there. In his framework, Osmani also gets into the more technical stuff. He recommends serving like a clean markdown versions alongside your regular HTML pages because Markdown strips out all the scripts and navigation noise that makes parsing expensive. He does reference the LLL let me try again, it's early. LLMs.text for structured documentation, and he also mentions the skill.md files and the agents.md files, which are essentially machine readable entry points that tell an agent what your content is and how to use it. Now, there's been some argument around these files, so it's really interesting to see this being called out. I also want to flag something else that's important. The search engine land piece does a really good job of pointing this out too. Google's John Mueller has publicly said that Google does not recommend separate markdown pages, as Google does not use llms.txt as a ranking signal. So Osmoney's AEO framework is not about Google search rankings at all. It's about making your content usable inside agent workflows like Claude or ChatGPT agents or inside whatever tool is fetching your page to complete a task for someone. Think about all of these projects that you have built inside of Cowork or Claude. If you have it, go look at a web page and then come back so you can do something with it. That's what we're talking about here. I think this story is really a signal that there's a third optimization layer emerging. Not instead of SEO, not instead of GEO, but in addition to it. Agents are gonna be a bigger and bigger part of how work gets done. And if your content is structurally hostile to them, you're gonna get skipped for whoever made theirs easier to use. Now, here's what I would actually do about this right now, like today, sometime this week, sometime in the near future. Audit your highest commercial value pages, just the ones that actually convert. See if those pages have the core answer right there at the top within the first couple hundred words. Not after your mission statement, not after a video explanation, but right there at the very top. Tighten them. If your comparison pages are 4,000 words of something like, I don't know, we're on a mission to revolutionize. That's an agent problem. Be very clear, very specific, get right to the point. I don't think I would run off and go build those lm.txt files or the agent NG files right away on every page. Um, the standards here are still moving, and what matters right now is just structural clarity of your existing content. That's I mean, that's my opinion here. I would also like keep an eye on it, but I don't think we need to spend so much time building it right now. Definitely a future task, not a right now thing. Alright, so story number two. This is a fun one because it's call back to the February episode where I spend half the show nerding out over Bing's AI performance dashboard. If you remember, I'm like, this is a big deal. It's the first one, it's the first search engine that said, hey, yeah, this is a thing, here's what we're doing about it. Well, on April 13th, Barry Schwartz over at Search Engine Roundtable, spotted something new. Google appears to be testing a feature inside of Search Console called AI Contribution Report. There are no screenshots yet, no announcements. This is just people being nosing and reading their documentation. But we do have references to, and I quote, AI contribution pilot inside of Google support documentation. John Mueller dropped a subtle hint back in February that something like this might be coming, and now it looks like it's actually being tested. Now, before anyone gets too excited, let me set some expectations here. We don't know exactly what is going into this report. We don't even know when it's gonna roll out to everyone. We don't know if it's gonna be as detailed as Bing's version, but I think it's interesting and it's worth paying attention to or at least putting on your radar for near future. Now, remember when I talked about Bing's AI performance tool, that dashboard tracks total citations and AI answers, which URLs get cited, grounding queries, and page level citation activity. It kind of turned AI visibility from guesswork into something measurable. And I said at the time, when the infrastructure solidifies, behavior follows. So when brands see what's working, they start optimizing for it. If Google ships out an AI contribution report that's even half as detailed as Bing's, this is going to be a big deal because Google already handles the majority of search volume. AI overviews are appearing on somewhere between 25 and 48% of queries, depending on whose study you believe. If Google publishers, or if sorry, if Google gives publishers native reporting on how their content shows up inside of those features, that means a couple things. First, we're going to finally be able to separate AI overview visibility from traditional Blue Link visibility, which you currently cannot do neatly inside of Search Console's performance report. Two, we'd be able to track which pages are most often used as source material for AI generated answers, which tells us something real about which parts of our site have the most authority in Google's eyes and how to fix the sites that are the page how to fix the parts of our site. That's where I'm going. I don't know if I've not had enough coffee. Anyway, how to fix those parts that are maybe a little bit weak. Three, it also gives us a feedback loop, which is the part that's been missing. Up until now, AI search optimization has been publish, test, cross your fingers, document the results, do it again the next time. A dashboard turns that into something repeatable and it makes it a little bit easier to see where there's movement instead of just doing like I do and writing it all down and then putting it in a spreadsheet and then making your graph. It does it for you. So my read on this is that Google is still testing all of this, and that's more important than the details because Google has been very slow to give publishers anything around AI-generated search. First, it was trust us, normal SEO works for AI overviews. If you've been hanging around on LinkedIn or anywhere I've been active, you know I've been saying that is only half true, there's more to this. Then we got some kind of notification, I covered it on another episode. Here are some controls through Google extended, but don't opt out because you'll lose visibility. Which was them saying the quiet part out loud. And then now it seems like okay, here's how you can actually see what we're doing with your content. Let's move it in the right direction. And they've been really, really slow at getting this one done. I'm gonna keep watching this if and when screenshots appear or the pilot expands. I will do a follow-up, I'll post it on the blog, on LinkedIn, YouTube, all the places. But for now, if you have Search Console access, keep your eye on the interface. You might already have a version of this and not know it. I checked mine this morning, I do not, but I don't know who has access to the pilot. It is not me. Alright, story number three. This one had the whole SEO world talking last week, and I want to come back to it because I don't think the takeaway landed the way that it should. The New York Times commissioned a study from a startup, an AI startup. I'm gonna butcher the name. Oumi? Omi? I'm not real sure. Startup names are kind of if you don't hear them pronounced, I don't I don't say them correctly. Super sorry to this startup. Anyway, they ran about 4300 Google searches through a benchmark called Simple QA, which is a standard industry test for how accurately an LLM answers short fact-seeking questions. They ran the test twice, um, once in October when AI overviews were powered by Gemini 2, once in February after Google moved to Gemini 3. The results were that Gemini 2 was accurate 85% of the time. Gemini 3 pulled that number up to 91%. Now, I want to be fair, 91% is a real improvement. But when you look at the numbers and you do the math, it kind of gets interesting. First, Google handles around 5 trillion searches per year. 9% of 5 trillion is tens of millions of incorrect answers every hour. Hundreds of thousands every minute. That is a volume problem, and even at 9% error rate at a search engine scale, that is a misinformation fire hose. That is not good. Second, and this is the part that I care about most as a content strategist, more than half of the answers that were technically accurate were ungrounded. That means that the AI overview gave the right answer, but the sources that it cited did not fully support the answer. So the answer was right, but the citation was wrong. I have noticed this happening sometimes. Now, this ungrounding problem actually apparently got worse when Google moved from Gemini 2 to Gemini 3. It did not get better. Google pushed back, obviously. The spokesperson said that the study had, quote, serious holes, and that the simple QA benchmark doesn't reflect real search behavior, which is fair. That is a fair point, somewhat. Simple QA is designed for short fact questions, not messy commercial queries. But Lily Ray made a point in response that I think is more important than either side of the accuracy debate. Lily essentially ran a stress test. She spun up an AI-generated article titled something like The Top 10 Experts at Knowing What Google Really Wants in 2026, and she published it. And within a day, Google was citing the article inside AI overviews. Her point, which I agree with, is that the manipulation problem is not limited to these weird edge case queries. These are high-intent commercial search terms, the ones where the stakes for buyers are real. And those are the ones that get gained the most heavily gained. So let me let me translate what all of this means inside of the FSA framework because I think this really puts it into perspective and what we should actually do about it. Authority, so the A in the FSA, is not just about being mentioned everywhere. It's about being mentioned consistently and credibly, emphasis on credibly, across sources that can be verified. So when Google AI Overview pulls from a blog post that just declared itself an expert, that kind of weakens, I mean, is weakness in the retrieval, not strength and authority. So if you are a real expert, your defense isn't really to outpublish the manipulators. I'd be curious if that works, but also that's a lot of time. Your defense is to build entity density. So consistent mentions across trusted third-party sources, those podcasts, the industry roundtables, industry publications, the whole distribution side of FSA that I keep pushing people to take seriously. This is also why grounding matters. Structure your content so that the claims that you make can be traced back to sources. Cite your data, reference the frameworks, name the tools and standards. Because when an AI engine is deciding whether your content is groundable, whether it can be confidently used as a source, clarity and traceability are going to matter here. Now, the takeaway that I want you to hold on to from this story is not that AI overviews are broken, don't bother. It's closer to this. The current retrieval systems can be manipulated. We've talked about that a couple of times over the last few weeks, but the brands that win long term are the ones that don't need to manipulate. You build real authority consistently, and short-lived tactics get filtered out, like they always do. Okay, last story, and this one is shorter but super super important if you work with local or service-based businesses. Search Engine Land ran a piece on Tuesday, April 14th from Rich Singer. He's tested Google's Ask Maps feature, which launched more broadly back in March. He tested it across a range of local services like plumbers, electricians, HVAC companies, starting from a simple category searches and then working up to more nuanced conversational prompts. Here's what he found. At the simple end, so a plumber near me, Ask Maps behaves pretty much like Google Maps always has. It'll list nearby businesses and it'll rank them like normal. But as the query got more nuanced, so as users added context, uncertainty, or decision-making language, Ask Maps started doing something different. It stopped listing and started recommending. It then narrowed the field. It framed businesses around qualities like responsiveness, specialization, honesty, repair first thinking. In some prompts, it even offered guidance before recommending businesses. So in other words, when someone asks, who's a plumber that won't try to upsell me on a replacement I don't need? Ask Maps isn't giving them 10 options anymore, and then given the burden of research to the user. It's interpreting the question and recommending two or three businesses that match the intent behind it. This matters because local search is getting the same treatment the rest of Search has gotten over the last 18 months. It's moving from retrieval to recommendation. Businesses that rank number one for plumber is not necessarily the business that gets recommended when someone goes in there and asks for an honest plumber who won't upcharge. Those are two different optimization problems. I I don't know if problems is the right word, but that's two different optimization tactics here. So what do we do? Well, reviews matter more, not less. The qualitative language in reviews is how the model figures out what you're good at. If every review says fast response, that starts to show up when someone asks for a responsive business. Your Google business profile description also matters. You need to be very specific about your specialization. Full service HVAC is weaker than specialists in older homes focused on repair before replacement. That second one is a recommendation hook and it is very specific. So go check what Ask Maps actually says about you if you are one of these businesses. Ask a few real prompts that your customers might use. If you don't like the framing, that's a content and review problem to work on and it can be fixed. For agencies with local clients, this is also worth building into your monthly reporting. Not just rankings and map pack position. What is Ask Map saying about this business when someone asks in a natural language? That is where we're working from here on absolutely everything we do for content strategy or marketing, marketing in general. That is what we're doing. So let's put all of this together because there's a real through line here this week, and it's mostly Google-based. And Google Engineering Director is publishing frameworks for agent consumption. Google is piloting native AI visibility reporting. Settings are putting hard numbers on AI accuracy and ungrounding, particularly in AI overviews, and local AI search is moving from listings to recommendations. Every one of these is infrastructure. Every one of them is an AI search ecosystem growing up in real time. Which means, and I'll say this. The way that I said it back in February. When infrastructure solidifies, behavior follows. We are at the point where AI search visibility is measurable, optimizable, and audible in ways that simply wasn't a year ago, or heck, even seven months ago. The brands that treat this as an opportunity are going to pull away from the brands that treat it as noise. If you want help figuring out where your brand actually stands across these surfaces, or if any of this made you realize you don't have a baseline you can measure, that's exactly what an AI visibility audit is for. Head over to CassieClark Marketing.com and you can get started there. But the links are in the show notes as always. I would love it and I would love you forever if you left a review. It genuinely helps other marketers and founders find the show. Alright, I will see you on Tuesday. Until then, stay visible.