What if user satisfaction is all that matters for ranking on Google?

What if user satisfaction was the only thing that mattered to Google search? I'm not saying it's the only thing, but I actually think it's the most important thing that we can optimize for. And if I'm right, if I can convince you by the end of this video, it means that much of what we do today in the name of SEO is probably not moving the needle. My name is Dr. Marie Haynes. I've been studying Google search for 18 years now. Kind of wild. used to be a veterinarian and then got really interested in just learning how Google works. I wrote a book recently that talks about the AI systems behind search and today I want to tell you the latest thinking that I have and again it's that I think that satisfying users is by far the best thing we can be working towards. You've probably heard of Google's quality raers guidelines. I started studying this document back around 2017 or so. They aren't Google's algorithm, but they basically tell us where they want the algorithm to go. And this document is actually used to train real people who help train Google's AI systems in search. I'm going to get to that a little bit more in a minute. Google also has this document on creating helpful content. And I've talked about this for many years. I have it highlighted with things that I share with my clients. And again, this is not Google's algorithms, but it tells us what people like. And Google wants to reward people with things that they find satisfying. So, here's my theory. Google first uses traditional ranking systems. Those initial rankings are then reranked by AI ranking systems, and those AI ranking systems are trained and fine-tuned based on two things. One is the ratings from the quality raers and the second is the actual actions of searchers. Let me see if I can convince you of this by the end of this video because if I can, I think that it will change how you do SEO and how you approach trying to get recommended by Google. Now, I want to say that I don't want to discount the amazing work that is done by my friends in technical SEO. There are a lot of things that are important in keeping our website sound. If Google can't find your content, if Google can't crawl your content, then you're not going to rank. So technical SEO is is very important. But in most cases, I don't think that it's technical SEO that moves sites ahead above their competitors. Danny Sullivan from Google recently said, "Don't turn your content into bite-sized chunks." He said, "Uh, one of the things I keep seeing over and over in some of the advice and guidance that people are trying to figure out what we do with the language models and to turn your content into bite-sized chunks because that's what LLM's like, things that are really bite-sized." Now, people who understand how vector search works, vector search is the way that AI search essentially works. I'm kind of making it far too simple, but really vector search works by embedding the query or variations of the query or I think the query plus related intents maybe the fan out of the query into a uh an embedding space and then in that vector space content that is nearby that query will do well. Now, if you're trying to rank in AI overviews, then uh one of the things that tends to work well is to understand how vector search works and to write your content in chunks. And what I mean by that is uh I might say, all right, the AI overview wants to answer these five questions, and so I'm going to make sure that I have a chunk of content that that concisely answers the question in a very small chunk. And that's good advice. It is something that it's good to look good to language models. The thing is that if you overdo it with this, it doesn't look good to people. People don't want to read when they go to a blog post, they're not going there to read this chunk, this chunk, this chunk. They're going there to read something that helps inform them, that is interesting, that's original, that's not just repeating what's in the AI overview. So, how a lot of SEOs that really understand how AI search works are like, "Danny must be wrong." Well, let me share with you why I think when Google says, "Just create great content." And and I can understand why people would say, "Well, no, that's not the case." Because even back in the days of Matt Cuts, Google has been saying, "Oh, all you need to do is create great content." But what was winning is people who actually knew SEO, people who were building links, people who were doing incredible keyword research and doing good jobs with internal linking and all of these things that we know really really matters. And over the time, over the last several years, especially since, well, Google started, they started using AI many years ago with spelling correction. And then in 2015, they used RankBrain, which um was kind of a start of using vector embeddings to to better understand which content was related to which uh concepts that were in a query. And up till current day, I mean, Google's made all sorts of advancements in how they use AI to determine which content is likely to be relevant and and helpful to a searcher. So, it kind of makes sense that like if we can write in a way that looks good to AI and if AI reanks Google's search results, then why wouldn't we do that? Well, let me share why. First, let's talk about the DOJ versus Google trial. This trial was primarily about monopolization. They basically said that Google has a monopoly over all of the other search engines. Nobody has a chance to compete against them. And the primary thing that they were worried about was that Google had such an advantage because of user side data that nobody could ever compete. Bing has a tiny fraction of the searches that Google gets and they don't have nearly as much user side data. And so much of this trial was uh showing that Google's user side data was actually the key to their success. And and the judgment from the trial basically said that Google had to share their user side data with competitors. Google filed an appeal just recently to this and said we're asking to pause the implementation of specific remedies that would force us to share search data and provide syndication uh services to rivals. So, they don't want to share their search data. We dug into this a little bit more. I recently wrote an article. If you go to my site and find the blog, you'll see an article with all of these things that Liz Reed, head of search, said in regards to why they don't want to share this information. This is from her declaration that was just recently written and she said that Google is supposed to share two things. user side data used to build, create or operate the Glue statistical model. Now, Glue is a form of Nav Boost. I've I've written a lot on Nav Boost in the past. Glue looks at recent queries. There's a fantastic discussion by a gentleman named Douglas Ord. I'm going to talk a little bit more about Glue in a second. The second piece of user side data that Google did not want to share is the most interesting. User side data used to train, build or operate the rankinbed models. Google doesn't tell us a whole lot about rankinbed. I've been doing my best to learn as much as I can. It is really fascinating. But first, let's talk about glue. In the DOJ versus Google trial, there's a testimony by an individual named Douglas Ord. And he is this is actually it's worth reading the whole testimony. It is absolutely fascinating. Talks so much about user side data and how Google uses it. Douglas or is an expert in information retrieval and he's testifying against the things that Google said where Google was like no no user side data isn't that important and he goes on to show that yes it is important and he talks about the system called glue which aggregates the signals that they talked about before. So, Glue looks at clicks, hovers, scrolling, which is and these are all things not necessarily on pages, but in Google search results. So, every single search that you do is actually stored by Google, the query that you searched and the actions that you took and whether you hovered over certain images, whether you clicked on certain pages, whether you went back, you clicked on a page and went back to the search results, that's all stored by Google. and a system named Glue puts that all together. And glue is essentially a form of nav boost. It doesn't really matter whether we differentiate glue and nav boost. My understanding is that glue is for uh it makes Google better at understanding recent queries that are happening. So here's the part that I think is most interesting. Douglas or says, "Yeah, there's this system called instant glue." And he tells this scenario of people searching for pictures. If I search for nice pictures, I might be looking for something to improve my Photoshop or my slides or I might be looking for some some nice art. And uh that's what he says. You know, you you might need something to decorate your PowerPoint slides. And then Google tries to give you nice pictures. Then he says, "But if there's a terrorist attack in Nice, France," so clearly spelled the same way, NICE, and you ask for NIC pictures, which is spelled the exactly same way. The day before the terrorist attack, you probably wanted nice pictures. Now you want nice pictures and Google can tell because everybody's clicking on these pictures from Nice or paying attention to the pictures from Nice. So this is a simple example of how Google uses user data. This is not for every single query. This is supposed to be for fresh changing things that if something new happens in the world, it's not okay for Google to gather weeks of data and then two weeks down the road start showing us what people really wanted. Rather, they need to change very very quickly. So this is part of what the glue system is. What's more important to me or what's more interesting to me is to look at the second type of user side data that Google does not want to share and that is the user side data that goes to train the deep learning model called rank inbed BERT. I had not heard of this before the DOJ versus Google trial came out and there is some good information although Google doesn't write a whole lot about this proprietary system that they have. This comes from the Pandu Nyak testimony. Again, something that is highly worth seeking out and reading. It tells us a bunch about how search works. So, first they start talking about the traditional ranking signals. So, they say on receiving a query, uh, Google uses the index to retrieve pages matching the query. So, there's another place where there's a diagram. Let me find that diagram for you. This again is from the documentation from the trial where Google says that all right somebody searches a query there are trillions of possible results for that query. You know that Google's mission is to organize the world's information and make it accessible and useful. So they take the entirety of the world's information and they sort it down into uh eventually just several hundred sites. So this is talked about in the trial. They say Google uses the index to retrieve pages that are matching the query. And uh there are many pages that match a query. A simple match could yield as many as 10 million results. And and that's that's you know often the case. And here's the important part. Google uses what Pandu Nyak called core algorithms to bring that down to a set of about 200 documents. And Pandun Nyak said several hundred. Yes, several hundred. And that's the core ranking algorithms. I'm telling you that most of what we do in the name of SEO is to prepare and to look good to Google's core ranking algorithms. Those algorithms give the documents an initial ranking score. And then here's where it gets really interesting. Once Google has the smaller set of documents, then deep learning is used to adjust the score. Now, Google doesn't share a lot about their deep learning systems and it's something that I've been learning a lot about deep learning and this is why I want to share this with you because it's these deep learning systems that user side data goes on to train along with something else which we'll talk about in a second. All right. So, one of those deep learning systems is the rank embed system. Embed means that there is a vector space a query is getting embedded. Somebody asked me what type of tokenization does rank embed use? It's not written anywhere, but Rank Embed is a BERT model. You might have heard of BERT uh years ago, I think it was 2018 that Google wrote a blog post about how Burp Burp BERT was helping them find uh to understand searches even better than before. And at the time, I didn't they knew they told us that BERT was a language model, but I didn't know what a language model was in 2018. And BERT is fascinating. The actual way that they trained BERT is by showing it pairs of questions and answers and they'd mask out part of the answer and the model had to try to get the right answer and they took these question and answer pairs from two sources. One is Wikipedia uh where you know they would ask a question they would find the answer as it existed in Wikipedia and see can BERT actually predict it and if BERT predicted it wrong then the model would learn how to change its weights how to just slightly improve each time on uh getting it right. The other place where Bert was trained was on something called Books Corpus. And this is fascinating because Books Corpus is a bunch of uh science fiction and also romance novels, which is just wild that Bert learned to understand English by studying romance novels. I guess romance novels are good at describing things. I don't know, but uh I found that fascinating. Now, that's Bert. There's a special version of BERT called Rankinbed BERT and that is trained um not on Wikipedia and romance novels and science fiction uh but rather on actual ranking queries and so my understanding uh I can't prove this 100% is that Google can look at a person searched for this and we predict that this result is what they're going to click on or maybe these you know five results listed in order of probability that you're going going to click on them and then what the person clicks on actually helps to confirm whether their prediction was correct or not. So let's go back to these court documents here. I think they mean rankbed BERT here, not bot. I think that's a a little transcription error. Rankmbbed BERT is one of the deep learning systems that also does retrieval. So it's helping Google find results that are potentially related to the query. But um Rank and Bed Bert is just one of the deep learning systems that Google uses. They say in the trial that Google started using deep learning in 2015 and that's when Google told us about RankBrain. There's an article on search engine land written in 2016 by Danny Sullivan long before he worked for Google that tells us about RankBrain. It's worth seeking out this article and reading it because it really tells us a lot and I think at the time we had no concept. I didn't even know that machine learning was synonymous with AI when I first read this. But at the very end of this article, he says that Google tells us that if we want to learn more about RankBrain that we should read their blog post that they wrote on something called wordtovec. This blog post is about learning the meaning behind words and understanding how words are converted into vectors into numbers is fascinating. I won't go into too much detail here, but let me tell you in case you haven't heard it, this thing that just blew my mind when I first heard it that if you so the way that they create these systems and it's advanced far more from uh when they first developed the I think wordtovec was written in 2013. So uh it's come a long way since then. But you can convert words into numbers and then as you learn more about the words and words that are surrounding them. So if I, you know, am talking about a king and king is often seen in conjunction with the word castle, it makes those two words closer together in the vector space. And once you've seen tons and tons of training data, then you kind of get a sense of which words are close to each other. And it turns out that if you do enough training, you can actually do math with these vectors. If you take the vector for king and subtract the vector for man and add the vector for woman, guess where you end up at queen or very close to it. So vectors are fascinating and this is the whole idea uh that started off what has now turned into rank embed BERT. There's a lot more that we can learn on vectors and it's really fascinating to learn it because once you start learning this stuff, if you're an SEO, your mind starts spinning and going, "Oh my gosh, I know how search works and now I can rank for it." I'm going to tell you why this is a dangerous tactic to do. Let's go back to the trial. So, Pandu Nyak talks about these different uh deep learning systems and they all contribute in some way to the final ranking score. The three main deep learning systems that Google uses are RankBrain, Deep Rank, and Rank Embed BERT. I want to learn more about Deep Rank. I feel like I haven't really uh learned enough about that yet, but there's not very much that's written on it. Uh but for today, we're going to talk about mostly rankbrain and rank embed BERT. And these deep learning models are trained in part on click and query data. So, what that means, I'm pretty sure that the way that I mentioned that they train BERT where they say, "All right, here's a question. Uh, here's an answer with some words masked out and can you predict those words?" And if the prediction is wrong, then the way that these neural networks work is that they just know how to go back and slightly adjust some of the weights in their system, some of the numbers in their calculations, and how much weight they give to each of those numbers so that they get just a little bit better at potentially predicting those results. Well, I think the way that RankMed BERT works is based on what people click on. So uh the system predicts it's going to be your result and if people actually click on another result then uh that trains the system that oh whatever we did to make that prediction uh you know maybe wasn't completely accurate let's adjust our our thinking and it's not uh onetoone so I think of this analogy that if I go out of my house and I walk down my driveway and I go to cross the road and I almost get hit by a car Well, then I've learned some things. Uh, I've learned that I need to look both ways, that I need to be a bit more cautious, but not just for my street, for every street that I go on to. And this is kind of how I look at these models. So, when Google was asked, is click-through rate a ranking signal? Well, clicks are used, but it's not a direct ranking signal. It's not like, oh, if I can get more people to click directly on my result, I will immediately rank better. You can get some temporary uh results by trying to manipulate click-through rate. And you can do more harm than good by teaching the systems that oh, you predicted that we're going to rank be the best result for searchers, but actually searcher actions showed that they went back to the search results and got satisfied by another result. But here's the part that is most interesting to me in the Pandu Nyak testimony. Let's start with RankBrain. Rank brain. So Google's AI brain for ranking looks at the top 20 to 30 documents and may adjust their initial score. How wild is that? So we know that Google's core ranking systems gather the initial, you know, from trillions of results down to a few hundred results and then rankbrain or AI systems rerank those results. And this is why there's a big push right now to understand things like cosine similarity, vector search, nearest neighbor search, things like that because it only makes sense that if AI systems are re-ranking uh the search results, then we want to look good to those AI systems. I do think that this part is really interesting, not as relevant to our discussion, but I have some thoughts on this. RankBrain is expensive to run and uh they say RankBrain is too expensive to run on hundreds or thousands of results. That is correct. Um and uh and so I actually think so this was before some of Google's latest advancements. They've had more advancements in being able to put more in the context window for AI for this is for language models. It only makes sense to me that they can do more and more. Gemini 3 flash let them do a lot more and faster. I would imagine they're using that technology in search as well. So, we don't know if today the same holds true that Google's traditional systems get us down to a few hundred results and then RankBrain ranks the top two to three pages. It might be more. Uh but that's kind of interesting to know. So, let's talk a little bit more about how these AI systems are trained because a new AI system entered the mix starting in August of 2022. This was Google's helpful content system, which has been very controversial for a number of reasons. When Google announced the helpful content system, before they announced it, they reached out to a few of us in the community. So to me, Glenn Gabe, Lily Ray, and a few others. And Danny Sullivan had a conversation with each of us that essentially said that the way that ranking works is going to change. He didn't give us any inside information. I think they were just trying to get a sense of how SEOs would understand this change. And it turns out that really none of us understood it. I studied the helpful content system for years. And it took me a long time to really grasp what was happening here. But one of the things that I wanted to point out is when they announced this, they told us it was a machine learning model. Again, at this point, 2022 was before chat GBT had come out. So, most of us really weren't paying a lot of attention to AI at that point. Now, what is a machine learning model? I did a couple months after this, I did an interview, you can find it on my YouTube channel, with Alan Kent from Google, and he gave the greatest description of what machine learning is. I'm going to try to paraphrase him. It might not be exactly what he said, but he said basically you show a system examples of here's what's good, here's what we want to reward, and here's what we don't want to reward. And then the system learns what characteristics to look at and how much weight to give each of those characteristics so that it gets better at predicting what's good. So, the helpful content system, I think, took Google a step further in that they weren't just looking at what people clicked on and whether they went back to the search results because that's really noisy. The nav boost or glue signals really, I mean, I could click on something and then the phone could ring and I could go run away and leave that page open. That doesn't necessarily mean that I've been satisfied. the helpful content system. Google was learning to use AI to predict whether uh a page was likely to be helpful. My understanding today of the helpful content system, if you were impacted by it, it primarily means that you were really good at SEO and that it was SEO that was helping you look good to the machines, but maybe not as good to actual searchers, that searchers were preferring other content. It's also possible that you did absolutely nothing wrong, but you had the characteristics that were the same characteristics of sites that were meant to be impacted by the helpful content system. I think that that happens in some cases. I think that most sites that I've viewed that have been impacted. Uh you can see that they were really good at SEO. Um so I'm not going to dwell on the helpful content system, but it's important to know that Google started learning how people interacted. Now, did they use user signals for this? I don't know. I think it kind of makes sense that it did. I also think it's interesting though that the quality raers were told that the reason why they're doing their ratings is to improve search engines by providing examples of helpful and unhelpful results for different searchers. So whether that's a part of the helpful content system or just uh the rank embed system or whatever Google's deep learning systems, we know that Google learns by looking at results that are helpful and results that are unhelpful and then trying to make algorithms that do a good job at predicting the helpful ones. Then we'll go forward and look at the March 2024 core update with this update. So the helpful content system had been running for almost two years and Google said we're bringing what we learned from that work. So in 2022 we began tuning our ranking systems to reduce unhelpful unoriginal content on search and keep it at very low levels. We're bringing what we learned from that system from that work into the March 2024 core update. The March core update was a significant one. In fact, when Google announced it, they said it marks an evolution in how they identify the helpfulness of content. So, we went from right, Google uses clicks in the search results to kind of approximate whether pages have been found helpful to now they used uh so then they used AI systems to kind of classify the pages that were overtly unhelpful and then they integrated all of the learning into the March 2024 core update. I think it's fascinating how uh a lot of these core updates fall on the heels of Google making new advancements with Gemini. And I think that as they improve their AI capabilities, their uh infrastructure, their chips and everything that's involved in making the calculations for uh machine learning, then they can do more and more. They can look at more signals. They said there's no longer one signal or one system to determine helpfulness, but rather multiple uh signals are are used. The March 2024 cup uh core update made it so that uh Google was much much better at predicting which results were going to be helpful for searchers. Now, that doesn't tell us a whole lot about user signals. So, let's go back to Douglas Ord's testimony because this is the most fascinating part. In his testimony, he talked about frozen versus unfrozen Google. He used this slide. You can see this little ice cube where Google is frozen. And uh and then this is uh what he actually called retrained Google. Kind of interesting, right? This is talking about a score called IS-4 at five. This score is the quality raider score. We think that is stands for information satisfaction, but I haven't been able to confirm that. So how the quality raers work is Google puts two sets of results in front of them. Let's say a Google engineer has this idea to make search faster or to reduce the amount of spam that's in search and they introduce that into a new algorithm. Before they put that algorithm live for people to use, what they'll do is they'll show the quality raers the frozen results. So here are the results as Google search currently exists and the new results and the quality raers look at every single result. The at five on uh this score rather than just is score means that they're looking at the top five results including SER features. The latest version of the quality raider guidelines actually gave examples of AI overviews being included in this analysis as well. So, if you've noticed with the latest core update, I'm going to come out with some stuff soon on the December core update. If you've noticed that there were changes uh in um you know, maybe how frequently Google's showing the people also ask or SER feature or maybe Google's showing more popular products or maybe Google's showing more or fewer AI overviews. This is all based on the systems predicting what it is that the searcher is going to find helpful. Are they likely to click on a maps listing or are they more likely to want an organic result? This is all determined by AI. So the quality raers look at each of these results and they rate them as you know in this example on the right here the top three were healthy good uh highquality sites and then these two uh didn't get such a good rating. So these ratings of the quality raiders it's not like you know if this particular site was marked by some quality raiders as lowquality it's not like that site gets penalized rather the system learns that all right whatever happened in order for these lowquality sites to be uh impacted we need to adjust our thinking so that that doesn't happen so much it's not going to be perfect and I mean imagine writing an algorithm that's perfect in every situ situation, it's going to be very very difficult. So, the system is constantly learning. So, I think it's worth reading some of the testimony going back to the Douglas or testimony because it talks about how this quality writer system works, but also how it has some issues and the issues are what makes it necessary for Google to use user data. So, uh it they talk about the frozen version of Google was evaluated by quality raiders. These are people that are hired by Google through a contractor and they're trained to perform the rating. It's actually quite difficult to become a quality raider. You have to pass an exam. I've spoken to several people who have uh who are quality raers and quite a few people don't pass the exam to when it comes to understanding quality. They give a quality score to each of the items that appear on their page. So, they're looking at every single page. You probably if you look in your Google Analytics data, you can probably see uh I think they come from Raiders Guide or uh there's a couple of places, but uh they've probably been on your website, but that doesn't mean that there's anything wrong with your website. It just means that it appeared on one of these searches that they were uh told to do. And so they they don't care whether uh these results appeared like first on the page or fifth on the page. They're mostly just wanting to know are they good or not? And they use the information that they've been trained on in the Raider guidelines to help make that uh decision. So they give an example here of uh you know, oh all of the results here are good, so they all get green stars. Now when Google retrains the system, so this is when the AI system gets retrained based on quality raider scores, based on, as we'll see in a second, user data. Then the frozen system, that's the ice cube with the little experiment symbol in it. Uh it might be that some different information is shown on the search engine result page that they're assessing. So the new system, so I said, what if a Google engineer wants to try a new type of um algorithm or something? I think most often what the raiders are rating is not the introduction of a human written new algorithm, but rather the new algorithm that's been suggested by the AI systems. So, uh, the question is, um, as they do this, is it worse or better? Is it worse than it would have been with more user side data? Let's get to the user side data. So, Douglas or talks about two issues with the way that these ratings are done. The first is that humans aren't really good at assigning a score to something. He talks about how uh in diving competitions, some humans are going to hold up a different card than others and we're not always going to be the same in our judgments of quality. But he said, but they are good at making comparisons. So you can say that this particular diver was better than this diver or in the case of the quality raers, this set of results was better than this set of results. But then the other thing is uh that they don't know everything. So, uh, this is why they do live experiments with live users because live users can show you things that human quality raers aren't good at measuring. And this isn't something that Google is shy about. They tell us that uh I'm not sure what year this is in, but recently they did uh nearly 17,000 live traffic experiments. So, they show uh one set of results to a very tiny portion of actual searchers and a different set. the different set would be the retrain set to uh to other searchers and again they're looking at what people clicked on and Douglas or talks about this where he says the best approach that we have in information retrieval where we're actually trying to understand what's going on in the user's mind is to look at this from multiple perspectives and um let's see here he says the amazing thing about search engines is that they also have all these users all this behavior from which we can generate this insight and This is this user behavior is the thing that Google doesn't want to give away to their competitors. Their competitors being Bing, being open AAI, uh I I think there are several others as well. He says the experiments that are run with human users run with thousands or maybe even millions of human users because any one user can behave eclectically. I mean, the phone rings. Oh, I just said that. The phone rings and all of a sudden, you know, it seems like they're on a page for a long time, but really they weren't. It's the user side data that gives Google a perspective that would be hard to generate otherwise. And here's the reason why they do live experiments because equality raers don't know everything. Uh Douglas or says, "So if I type in a query and I ask you what I meant from the query and there's only a couple of words, you might guess wrong what it is that I was looking for." and he talks about how he has a PhD in electrical engineering and he could type in a query that how are the quality raiders supposed to know whether it's a good result if they also don't have a PhD the raiders may not understand the highly technical queries also raiders might not understand queries that come from children they might be looking for something completely different and Google wants to provide good results to everybody no matter what your age is so the point of this is that Google uses these live experiments ments to see what real users are actually clicking on. Now, one of the questions I have is whether they actually use Chrome data. Google doesn't tell us a whole lot about how data from Chrome is used, and it really wasn't brought up a lot in the trial, but I did think that these two things in these documents were really interesting. The two exhibits in the trial suggest that popularity is based on Chrome visit data. Popularity used to be synonymous with links. But if I asked you, if I gave you a bunch of web pages and I said here are the ones people have linked to and here are the ones where we can see Chrome data to show that people are using them. Like what is the better indication is uh you know if they can see how people are using them and we don't know exactly what they see. The former appears to be a type of user interaction data albeit from Chrome visits. So here's what I think. I I initially thought that every time for every query Google was able to see were people submitting your forms, were people actually making your recipe, were people buying your products. I don't know that they do that for every site. It might just be for the live experiments. So, we don't know. But we know that what Google's trying to do, their ultimate goal is to show people results uh that that they're going to find helpful. And then the systems are fine-tuned based on what people actually did find helpful. Which brings me to my last point. I wrote an article a while back on this, but I want to hammer home the point because as we start to realize that AI systems are primarily the thing that are driving rankings, then it really is tempting to optimize to look good for those AI systems. And here's what I think. I think that Google gives a prediction for every website, for every query. So let's say you have learned everything about vector search. You've learned about cosign similarity, about nearest neighbor search, about any of the ways that Google any of the new innovations that Google's made in terms of using AI to determine whether content is a good match for a query and uh you do an excellent job so that the AI systems think you look amazing. That's going to initially increase your chances of ranking well. I don't know if you've noticed, but the vast majority of the time when somebody is bragging about how they used AI to scale up content creation, you'll see these charts from hrefs. I saw Glenn Gabe called it the AI mountain where traffic goes up and then it comes down again. And uh the reason is that those systems initially you might look good. You might have some initial boost by looking good to AI but then you're retraining the systems based on user data whether it's user interaction data. Um and it might not even be user interaction data on your site just like I learned to look both ways when I went down to my road. Uh Google might learn the characteristics that belong to pages that people tend to like. And so the whole system is incredibly complicated, but what we need to know is it's built to reward the pages that people actually did find helpful. And so what I would encourage you to do is to understand how AI search works. I do think that it's a really good thing um instead of optimizing intensely for vector search, understand the intent of your audience and make sure that you have things in your content that meets that intent. It doesn't have to be in discrete chunks. I do think that it's important to use good headings that help people skim and find uh oh this is the information that I was looking for. Um but ultimately your goal is to provide people with the best result. And I think that it is really important to look at these questions that Google gives us in their documentation on creating helpful content. These aren't a checklist. It's not like Google is going through your page and saying ah this page has original information. I mean maybe they are um but what they are doing is saying this is what people like and I think the most important of this is does the content provide substantial value when compared to other pages in the search results and I would make that your main goal is to look at your pages to look at what Google is showing searchers and to figure out why did Google think that that was more valuable sometimes it's things that you cannot fix if Google is showing all brickandmortar stores and you are just an online shop. It might be that that's something that the algorithms have weighed heavily or that's something that people uh people's actions have trained the algorithms to reward. Um but then sometimes you can look at pages and go wow they have unique imagery or a good thing to look at is they cover this aspect of a topic. Liz Reed, the head of search, recently gave an interview with the Wall Street Journal where she talked about um how AI overviews are answering people's questions and when people click on websites from within the AI overview, that what they're clicking on is pages that go into greater depth, that pages that are not just rephrasing what's currently in the AI overview. And I know for a lot of sites that are trying to optimize for AI overviews, you're essentially just trying to write the same text that's in the AI overview, I don't think that's the way to go long term. So I would say write for searchers, create great content as as trit as that sounds. And really focus in your metrics at Google Analytics where it talks about your user engagement signals. Are people scrolling down your pages? Are they spending time with your content? Are you truly satisfying users? and work to continually improve on that. If you've enjoyed this, I'd love for you to join me in my community. You can get there by community.mmariehanes.com. And I share the latest in everything that's happening with AI and search. And if you're super serious about becoming a leader as AI changes our world, join the paid area. We've got loads of fantastic discussions. We just had a really great event where we learned how to use anti-gravity to make a Google search console tracker. I would love to have you in the community.