# Entities Are the Past: Search Is Going Multidimensional by Tom Anthony | MozCon 2023

## Метаданные

- **Канал:** Moz
- **YouTube:** https://www.youtube.com/watch?v=damDqVTjIec
- **Источник:** https://ekstraktznaniy.ru/video/34768

## Транскрипт

### Segment 1 (00:00 - 05:00) []

10 9 8 7 6 5 4 3 2 1 Year after year, as a child, all my wife wished for was a Barbie just a regular old Barbie doll. But her parents never got her one or they couldn't do I mean maybe they could have done on the black market but it wouldn't have been easy. You see my wife was born in a country that no longer exists the DDA the Deutsche Democratic Republic known to us in English informally as East Germany. My wife lived in East Berlin with her parents and that meant that sometimes they could pick up West German TV being broadcast from West Berlin. So my wife, she knew that Barbie dolls existed, but she couldn't have one. My wife describes this time in her life as shades of gray. Everywhere she went, everything she saw was some shade of gray. When she went shopping with her parents, there was one brand of tin tomatoes, one type of yogurt, shades of gray, one of this, one of that. The world felt very one-dimensional. And then on my wife's sixth birthday, quite abruptly, East Germany and West Germany were reunified into the Germany we know today. Incidentally, this is still celebrated as a national holiday in Germany where we live with our kids. So, the way that my kids perceive it is everybody in the country gets a day off of work just for mommy's birthday. My wife describes the fall of the Berlin Wall through her six-year-old eyes as an explosion of color. Suddenly, everywhere was colorful, colors she'd never seen before, colors she had never imagined. Now, when she went shopping with her parents, there were different brands of tin tomatoes, different types of yogurts, new fruit and vegetables she hadn't seen before. The world had seemed so one-dimensional, but now we're so colorful, so full of choice, it seemed multi-dimensional. So, why am I telling you this? What has this got to do with search? Am I even at the right conference? I think we're on the brink of a paradigm shift where we're going to enter a new era of search. I think we've been living in a world where large aspects of SEO have been on one-dimensional, but I think there's a change coming and that change is going to make the whole world of SEO far more multi-dimensional than it has ever been before. So today I'm going to be talking a lot about context. And when you talk about context within the terms of uh SEO, there's either implicit context or explicit context. So implicit context talks about the state of the searcher, their location, their time of day, their language, all of that stuff. And then explicit context basically talks about the search query. What is that user looking for in that specific moment? And combined, these are what Google and other search engines use to generate search results. Today I'm going to be focused entirely on explicit context because I think we're on the brink of an explosion in the amount of explicit context that Google wants us to put into search queries and in their ability to understand that context. So I'm going to talk you through four things. keywords and entities, the current era that we're in. Latent spaces. This is the technology that's going to bring about and precipitate this change and lead us into the next era. I'm going to talk about what that looks like. And then finally, I'm going to wrap up talking about now what? What does all of this mean for us as SEOs, so that when you get back to your desk on Wednesday or Thursday, depending on your flights and or hangovers, then you will have an understanding of what you can talk to your boss or your clients about what the impact the strategic impact of this what I've told you today is. Okay. So, let's start off by talking about keywords and entities. Hopefully you're all familiar with keywords, the concept of keywords as SEOs, head terms and uh longtail. The thing about keywords is they're very limited. They don't allow us to talk about structure or relationships. Which is why five or six

### Segment 2 (05:00 - 10:00) [5:00]

years ago as an industry we and Google started talking about entities, things not strings. And entities introduced two things. Entities introduced structure. So now we could say okay well this entity has these attributes and relationships with other entities and the world was much richer and much better and entities as SEOs we typically represent these as structure markup. There are different ways that entities might be uh represented in the knowledge graph etc but I'm going to focus on structure markup via schema simply because that's the most widespread most documented way that we as SEOs represent entities. But the thing about structure markup is it can only represent a very small portion of our actual content. Let's have a look at an example. This is jamesvillas. co. uk. It's a vacation villa or apartment rental site. So let's try to mark up just the information that we can see at the very top of the page here with structure markup. We can say, well, this is a hotel. It's not quite a hotel, but that's closest schema allows us to get. So let's roll with it. We can mark up the name. photo just fine. But the region, we can call it a postal address and kind of budget, but it's not perfect. The occupancy, we can't mark that up. It doesn't allow us. The number of bedrooms, we can put the total number of rooms. Okay, we can't do anything about the number of bathrooms. The private pool, we can call that an amenity feature with a description of a private pool, but it's not an actual entity. So on so forth. The rest of the stuff we can't really mark up. And so if you look at this, you see that structure markup is actually pretty poor in what it can represent about our web pages. Uh Jonah Alderson, who's a former Moscon speaker and big uh structure markup fan, if he was here, he'd tackle me off stage right now for just for saying this. But so you can see it's actually pretty poor. So point being is if you think about the relationship between entities and that search context, entities can represent very limited context. They can only represent what is predefined in schema. org. So let's have a look at modern search habits and how uh context is represented there. So last year I wanted to take my family on vacation and the search that I wanted to do on Google was this. Show me holiday itineraries for a family of five with a toddler friendly pool uh and near the beach and some restaurants in or near Europe. Lots and lots of explicit context. But what I actually did was search for this Holiday villas in Europe and various variations of it because the last 20 25 years of using Google and other search engines have taught us that they're expecting us to type in two to five keywords. Dr. Pete referenced this earlier on that sort of constrained by this search box. Furthermore, as an SEO, I know that if I type two sentences of text into Google, I'm going to get very poor results. So what actually happens is I do this search on Google and then I search through the results opening each of those links looking for the content that I'm looking for. This is basically the second phase of the same search. Right? And this is a well-known phenomenon. It's called postarch browsing. This is something that the search engines know about. They've written academic papers about this. This is a well understood phenomenon. So basically what's actually happening is we're doing half the search in Google. then manually searching for the rest of that context ourselves. And what often happens then is you refine that search on Google. You repeat this process a few times until you get the search results you're looking for. Can be a frustrating experience. The other point to note about this is that search interfaces are still keyword oriented. The box is small enough that it's expecting me to type in just a small collection of uh keywords. So very limited amount of explicit context goes into the search queries that we do in the world of Google. So what's going to change? I said there's going to be a change. Right now this is us. We're living in this one-dimensional world, shades of gray. But we want to be over here, right? Colorful with lots of options with 80s haircuts. So what is this? What is going to bring about this change? And that is latent spaces. Latent spaces are the building blocks of the next era of search. I'm convinced of it. So to explain latent spaces, I need to mention chat GBT. Again, I'm sorry. Carrie already mentioned like there's been too many British people up here talking about chat GBT. So when it comes to chat GBT, I'm much less excited about this because even though we've had Bing chat search and we've got Bard and all that stuff, uh my former friend and colleague, no, that's the wrong way around. My former colleague and current friend or Tom Kappa Cyrus just called him a bastard

### Segment 3 (10:00 - 15:00) [10:00]

so let's roll with that. Tom Kappa the bastard pointed out yesterday how little uptake chat GBT has actually had and how small it is in comparison to web search. So yes, one day we will end up with these Star Trek interfaces that are all empowering, but we're not there yet. The technology is not there, but also the search behavior is not there. This requires quite a significant change in search of behavior and we spent the last 25 years telling people to type two to five keywords. So, what I'm actually much more excited about this part, the GBT part. GBT stands for generative pre-trained transformer. It's just a type of large language model. When we talk about chat GBT, we're referring to OpenAI's product. They invented GBTs as a thing, but now GBTs are out there in the world. So, any company can be using GBTs to build whatever it is they want. And so my question is like well sorry let's say the promise of chat search the thing that got everybody excited when chat GBT came and when uh Bing was out there in the world was this idea of it understanding the context of being able to do those complex searches that I referenced in my vacation example and basically then refine those right in the search engine itself. self basically bringing that second phase of search into the search engine where it should belong. But the thing is this understanding of context this ability to refine everything comes from the GBT part of chat GBT not the chat part. So the question becomes can we get the same huge impact on search without the chat part. In order to do that we need to have a little look at how GPT works and how it learns. So there's two types of ways that machine learning algorithms work. The first is with supervised learning where you show the machine learning uh algorithm a training set. This is a picture of a cat and you label it cat. This is a picture of dog. You label it dog. And you do that lots and lots of times. And the machine learning algorithm learns that there's two buckets, the cat bucket and the dog bucket. And then you give it a new image without a label and it says, "Okay, that goes in the cat bucket. " Incidentally, if we go back to looking at the structure markup on our web pages, I'm fairly convinced that this was basically uh labeling for Google using our data as a training set for machine learning algorithms. Basically, they would download our web page with our structure markup which would tell them what entities are on that web page just like this is a cat in this picture and then they could learn to recognize those entities on other web pages. Okay, so the other type of machine learning um uses unsupervised learning and this is where you remove the labels. So you feed in lots and lots of data and basically using advanced pattern matching the algorithm learns to recogn learns the concept of dogs or cats or whatever it is. But here there's no labeling going on and these concepts develop naturally. And this is much more akin to how human children learn, right? They go out into the world and they look at they see cats and they see dogs and they see cars and they see trains and they start to develop the notion of what these things are before they ever know what they're actually called. But what's really important as well is there's space for ambiguity. They recognize, oh, that's a cat that looks like a dog. There's a tram, so is it a train or is it a car? So on so forth. And so there's much more nuance goes into this. It's not the cat bucket and the dog bucket. The other thing about how children learn is that they learn to recognize cats and dogs at the same time as they're learning to crawl and tie their shoes and recognize sounds and all of that stuff. And so this is much more akin to how GPT learns. It learns from natural language text and it can learn different shapes of data at the same time. And this change is massive. It's really hard to explain how big of an impact I think this is going to have on search in order to try to get the point across. Schema. org which is how we represent entities right now has approximately 1,400 types of entity it can uh represent. GPT4 has an estimated 100 billion neurons. So if we visualize that for a moment, this is schema. org. You might think that I've forgotten to put something on this slide, but if you look really carefully, there's a pixel just there. That's schema dog. Now, let's look at GBT4. This is two scale. I know it's too scale because I remember doing the math uh on my kitchen table, and my wife teased me. She said, "That doesn't look right. " Uh and my daughter didn't even look up from her book. She's like, "I agree with mommy. " I was like, "You didn't even look. Angela Merkel doesn't get the day off of work for your birthday, daddy. "

### Segment 4 (15:00 - 20:00) [15:00]

So point being is I know I'm not comparing apples to apples right here, but what I am comparing is the current way that we represent entities and the new way that Google is going to be representing entities and the magnitude of this size change is just absolutely phenomenal. It's going to have a huge impact. So we know how schema. org stores data. So how does GBT store data? And that is where latent spaces come in. Latent spaces are how deep learning algorithms organize the data that they learn. So let's read this. In deep learning, a latent space is a vector space where input data is encoded in a way the model learns automatically. I'm not sure that's cleared anything up. So let me try. So imagine you've got a piece of paper and you draw a picture on it. That's 2D. Right now we're all in this conference room. That's three-dimensional. And so in mathematics, there's nothing to stop you adding a fourth dimension, a fifth dimension, 100 dimensions, and all the mathematics and all the geometry you've learned in school still works in that world. And so deep learning algorithms can then put stuff into this n-dimensional space and organize their knowledge within it where every dimension represents a different type of relationship that they've identified. And this is really important because it allows the deep learning algorithm to develop concepts and new relationships that you might not otherwise recognize. Let's have a look at a concrete example of a latent space. So this is a simple two-dimensional latent space from a machine learning algorithm that learns to recognize the written digits from 0 through to 9. If we look at the top left, that's clearly a nine. Top right, that's clearly a zero. But if we zoom back out again, there is lots of ambiguity and nuance in here. Is this a three or an eight? Where do the sixes stop and the zeros start? And this is really important. We don't have these discrete buckets of knowledge, but instead we can learn concepts and relationships, ambiguity, nuance, and all of that stuff which a machine learning algorithms will do at scale. Another example of a latent space, one that you might have seen before, it's a few years ago, is word tovec, which basically looks at natural language text and learns the relationship between different words. And then it allows you to do geometry on those words. So you can do king minus man plus woman and it will give you the answer queen because in the latent space that n dimensional space is created it's learned that on this particular dimension king is here queen is here man is here woman is here and so there's a relationships there and so you can see that you get these emergent concepts and so once you can do that you can start inferring context what do I mean by that so imagine you've got a web page for that vacation holiday. The swimming pool surrounded by a lockable safety fence has a depth of 2 ft to 8 ft. And now imagine in a GPT powered world, we do a search for a toddler friendly pool. GPT using in the inferred context can understand there's a link here. It understands that toddlers are small people and so the depth of the pool is important. It understands that when you say toddler friendly in reference to a pool, friendly means safety instead of lockable safety fence is important. And so there's quite a lot of understanding and nuance that you need to understand to get from toddler friendly pool through to this. This would never be possible in a world of structured markup. There's too much nuance, too much ambiguity, too much complexity. So I'm going to skip over this. So the point being is latent spaces which is how GPT organizes knowledge can represent any attribute any relationship. They don't need predefined buckets like schema. org does and GPT models can learn any context at a scale far beyond what schema ever could do. So what does this mean for the next era? What does this look like? So I think we had the keyword era and I think we're coming to the end of the entity era. So we're going to call it the context era. And there's more evidence that Google are actually working on this. About a month ago, Google updated their terms and conditions to announce that they wanted the to be allowed to scrape all of our online data in order to train AI models, GPT models presumably. A couple of months before that, they gave uh an interview with the New York Times. This was directly in response to Bing launching their chat G uh G chat GBT powered interface. And in this interview, Google said a couple of interesting things. One quote was, "We're upgrading the existing search engine with AI features. " Well, Google have already had AI features for years and years. So in this where they're responding to uh Bing launching GBT stuff, they're almost certainly talking about we're upgrading our existing uh engine with GBT powered features. The

### Segment 5 (20:00 - 25:00) [20:00]

other interesting thing about this is they're upgrading their existing search engine. They're not talking about a new intelligent agent app or a chat interface. They want to bring this directly into the interface that we know because that's where all the searches are. We saw that from Dr. Pete. Tom Kappa. The other quote was the new search engine will attempt to anticipate users needs. And when I read this, this just struck me as they're talking about explicit context. They wanted to get more context into the search. So let's have a look very quickly at two impacts that GBT would have on search once Google started uh implementing it. It would imp it would impact how we search and how Google indexes. So let's have a look at how we search. Firstly, if I do a search on Amazon for Barbie doll, I get specific set of faceted uh filters that are specific to the type of search I've done. So doll type, material, doll occasion, etc. If I do a search for sneakers, I get facets that are explicit to that. And so this is basically the Amazon in this case enticing you to add more context to your search. So as I was preparing this and noodling this many months ago, I am predicted that Google would do something like this. add dynamic facets to any search we did to entice us to add more explicit context to our existing searches. And then a month or two ago, a couple of months ago, they rolled this out. We've seen these filter buttons elsewhere in Google's interface before, but never in part of as part of this the primary search engine. And here, what happens is if you click one of these buttons, it just extends your search. This is basically Google incentivizing us to add more explicit context to our searches exactly as we predicted they would do. And so I think we'll I think this version is a beta test type version like a prototype but I think they will maybe double down and do something more like this where you do several um you do a search and you can select several facets and it filters the search more and more. Generating these facets is entirely within the capabilities of a GPT powered uh system. What's other the other really interesting thing about this is that it's entirely compatible with Google's existing business model, right? They don't like Tom Kappo once again, I got to stop talking about him, um pointed out yesterday that they've done a very poor job of putting sponsored contact um content into SGES, whereas here it's very easy for them to add uh sponsored content because that's where they already sell ads. So point being is that for us to benefit to get this extra explicit context into search, it doesn't need to have a chatbased interface. It might be dynamic facets. It might be filter buttons. It might be an advanced autocomplete. I don't know what it will end up being, but I definitely think Google want us to start adding more context into our searches. And when we do that, the long tail as we know it, I think it explodes because suddenly you're going to have an explosion of what I'm calling the latent tail where Google asks you to press a few filter buttons and now you've got what is essentially a completely unique search that we'll never see again. So things like search volume, etc. don't really come into play. So yeah, this sort of dynamic context makes the longtail as we know it look small. Okay, so the other thing I mentioned that uh will be impacted by GBT is how Google indexes. So this is a slide from a presentation I did in 2013. It's quite hard to know what I was thinking 10 years ago, but I think what I was aiming for with this takeout box with structured markup on it is the idea that there was a lot of discussion at that point in time of Google might implement something a bit like um Bing's index now API where you could submit your structured markup for a page and they would basically update their index based solely off of the structured markup on your page. But when you look at this and you see how poor structure markup is at representing the nuance and complete picture of our content, I think that me from 10 years ago was entirely wrong about this. We've gone full circle with a GPT powered engine. I think we've basically gone full circle back to the point where structured markup has been superseded and natural language content on our sites is the best API because it allows us to represent the full richness of our pages. So now what do we do with all of

### Segment 6 (25:00 - 30:00) [25:00]

this? Let's quickly recap. So entities current ability to represent explicit context is very limited. Google's current understanding of But GPT models can learn any context. They don't need structure markup to do it. And latent spaces as the knowledge of those models can represent any context. They're not constrained to a specific shape of data. And then finally, we saw that context doesn't need to be chatbased. So I think that the primary thing that I'd want us to take away from this is just a awareness of that this change is coming. But there are a few specific things that we should have a look at. So the first is uh structured markup. I think I'm not suggesting that you remove structure markup on your sites, but I wouldn't be spending any significant amount of resource or time on structure markup projects because I think it'd be better spent in this world feeding that GPT algorithm with contextrich content. uh you're much better off going and discovering what types of context might users be searching for and putting content on your product pages on your category pages that feeds that algorithm and lets it learn. Incidentally, I think we might see a shift in traffic patterns where more traffic arrives at product pages rather than category pages simply because if you're adding a lot more explicit context to search, then the search results become more specific and so there's a natural tendency to drive traffic towards product pages rather than category pages. So, how do you come up with this contextrich content? Well, obviously chat GBT itself is a useful tool for this. So in natural language processing, there is the concept of entailment. So entailment is the proper term for what I talked about as inferring context. With entailment, you can ask uh a natural language model or a large language model, why might someone search for this? So, I've asked with this prompt, why might someone search for holiday villa with shallow pool? And it spits out a bunch of answers. And the top of that list is kids safe swimming pool holiday rentals. So, you can see it's doing that exact sort of inferring context that I mentioned earlier and predicted would happen. And so, lots of people I think are very against using chat GBT as a keyword research tool. I think lots of the keyword research tools will be implementing GBT powered tech anyway, but lots of people are against using it because it doesn't have volume data etc. But in a world of the latent tale where so many searches are entirely unique, keyword research becomes quite a different problem. Instead, what you need to do is recognize categories of context like kids say swimming pool etc. which current keywords research tools I think would struggle to do simply because they don't understand all the entities and relationships. So for yeah 20 bucks a month you can get access to GBT4 and do up to 12,000 queries. So I'd encourage you to do this as well as your current keyword research stuff. Eat becomes uh ever more important just like Dr. Pete. I don't know how to pronounce it with two e so stuck to one e here. that becomes more important because if you're going to train a GPT algorithm, then you want to do that primarily off of content that is trustworthy, authoritative, expertise on it, etc., etc. There's potential to have a new Google dance. So, at the moment, I think we've seen certain um Lily Ray has talked about like certain um changes to our site only happen when there's a core update. uh GPT algorithms take months to train at large scales. GPT4 I think was somewhere between four and seven months to train and so apply that to Google search rankings and you get this issue where there'll be lag but this is really good opportunity for anyone who wants to get rich context onto their page first. I'm CTO at Search Pilot. So, it'd be bad if I didn't mention Search Pilot, but basically we started Search Pilot 7 years ago on a hypothesis that search was going to become more and more complex. And so, SEO BT SEOA AB testing is in becomes more and more compelling the more and more complex search becomes itself. So I don't think I am exaggerating too much

### Segment 7 (30:00 - 30:00) [30:00]

if I suggest that this could be the biggest change in search since page rank. Consider with page rank there was no requirement to change the search interface. for searchers to change their behavior. And the same is true if you start integrating GBT technology into the back end of search. the technology already exists. I think I outlined the magnitude of difference between the old way of doing things with schema. org and what GBT4 is capable of when it starts representing entities etc. And I think that's also the magnitude of the opportunity that we have with this. This is a lot to fit into 30 minutes. So if you want to talk about this more, come and find me at the after party. Thank you so much.