Garrett Sussman: My goodness. Dawn Anderson, ladies and gentlemen. Dawn is the managing director at Bertey and international SEO consultant helping enterprise brands like GSK, Skyscanner, and Madeline scale organic growth. She’s a lecturer at Manchester Met and a search awards judge with deep expertise in site migrations, audits, and cross channel strategy. Dawn shared her SEO insights across the UK, US, and Australia. We’ve gone super international SEO week, and is always ready for a good whiskey or gin. I’m so excited. She is the coolest person and so fun to talk to. Presenting, From Villains to Heroes. Please welcome Dawn Anderson.

Dawn Anderson: Hey, everybody. Thank you for staying. And I dispute the thing about whiskey and gin. I barely drank in February, and I did dry January. You go, ignore that bit. It’s definitely the lie. Okay. So thank you for having me. I appreciate the invitation. So what I’m going to talk about is some of you have heard a little bit about it today. At the heart of web search is the field of information retrieval. It’s an academic field. It’s the field that Sergey Brin and Larry Page studied whilst they were building BackRub, their search engine back in the day. It’s been going for decades and decades and decades longer than the web. And I became interested in it when I was taking a Master of Science in Digital Strategy. Back in 2015, I started to go and look at papers around crawling. I was a little bit obsessed around crawling and indexing. And I just became a bit obsessed by it then. So I’m going to be talking about this field and tell you about some of the things it’s going through at the moment. 

I cannot stress enough that how important this field is. Obviously, it’s everything. It’s everything behind search. It’s everything behind any website that has a search field, any search engine and so forth. But it’s going through some challenges at the moment. So what am I gonna be covering today? I’m gonna be talking about the outdated perception of SEO by the IR field as an academic or research industry. Why the search results in my view have gotten so bad. How the IR field is trying to resolve this. And how SEO is stepping up in this brave new world of search and IR to some extent. That’s me. Been doing SEO for a long time. And I used to speak at a lot of conferences, not so much just recently. Kind of went into my cage a little bit. Cage, hutch, whatever, cave. And then I had a nice invitation from Mike. So here I am. I decided to come out of my cave. And obviously, if anybody does know me, they’ll know I’ve got two fabulous Pomeranians. Whenever you travel and you’re a dog lover, I find myself talking to any random dog that I see on the street. I went running this morning in Central Park. I was stopping to speak to all the dogs, and I am into running as well. I ran a marathon on Sunday. So there you go. Thank you. 

Right. So let’s get back to it. Even before the day the term adversarial information retrieval was coined by a very, very distinguished engineer from Google, Andre Broder. You may not think you know who he is, but you will know his work. He was at the time at Alta Vista back in the day, and he coined this term adversarial information retrieval. Kind of talking about how SEOs are the enemy. Now, Andrea Broder is very famous for something called the taxonomy of the web. Now you’ll have heard of this because you know, everybody knows in SEO, the notion of informational, transactional and navigational queries. Well, he’s the guy behind that. And many other things, as I say, he’s a distinguished engineer, he’s been there for decades. But aka, spam fighting. Now also, if you saw Andre Ricardo this morning, who I’ve known also since 2016 seventeen where I met him at an information retrieval conference and he came up to me and said, who are you? I’m an SEO. He said, so you’re guessing then? It’s like But anyway, yeah, they are actually a very, very welcoming community. So as I say, aka spam fighting is really what adversarial information retrieval means. But really when it comes down to it, it’s kind of SEO bashing. It’s kind of been the norm, hasn’t it, for a long time. SEO bashing is kind of a bit the norm really for search and for the IR field and so on and so forth. Not as people individually. As I said, they’re very welcoming, but we’re not particularly popular amongst them. We’re kind of the enemy.

You know, many, many years ago, we talked about the Google dance and it was written up about the Google dance. Was like people were coined as those sneaky internet entrepreneurs. How dare they? How very dare they want to actually make money off the internet. Yeah. Even though we’re providing all this content and helping people to provide content which feeds search and so forth, how very dare they actually want something back from it. I remember reading a blog post on Matt Cutts’s personal blog that talked about how his engineers were actually really scared of all the SEOs. They were really petrified of like talking to us too much in case we ask them questions. Yeah. And I think even though we love Gary and John and Danny and so forth, there’s almost like that little bit of a you know, they’re kind of a bit scared of saying too much to us. And Martin New York, who’s a huge search engineer, still with Google, has been a research engineer in the team and generally across lots and lots of research teams. He’s huge in the IR field. He said back in 2006, the conflicting goals of search and content providers are adversarial, and that’s where it came from. It’s like, hey, these are the enemy. We do not share common goals. We want the truth. They wanna make money, you know? And I think to some extent, that kind of view, I think it’s moved on a bit, but I still think there’s a little bit of that about, you know, IR field generally and SEOs. It’s a bit outdated in my view. This was a track that they actually had that was about adversarial web search, the enemy. And it ran for a few years. It was literally looking at the very many ways in which SEOs were doing things to take advantage and inflate search results. And obviously, we have this ridiculous thing that after all of these years, we’re still seen as like comedic characters with various colored hats. You know, I think we’ve been likened a few times to like people sat in their underpants in the basement. I’m sure I’ve seen that in a few different write ups. I don’t literally I personally don’t sit in the basement in my underpants, but there you go, each their own. 

Yeah, I think it’s an outdated view. We’ve moved on. There are many, many professional people in the SEO world. We’re really not just the same old spammers. And actually, despite spamming search supposedly being an endless battle, having said that, the vast majority of it now, I would say is pretty much they’ve picked up the patterns. When you’ve got machine learning going on and you look at enough paid blog posts, you kind of know what a paid blog post looks like. When you’ve looked at –  there’s papers for instance in the IR field, one in particular is pretty famous, it’s quite old now, it’s called Do Not Crawl in the Dust. And that was like learning how to swerve really crappy content and thin content and so forth and learn kind of build it into crawling schedules and so forth. So there are a lot of these elements of machine learning now in search that means that they do avoid the vast majority spam automatically. And like I say, is Danny just very recently saying, even though 50% of what we find from the open web is spam, 99% of it actually doesn’t get into web search anymore. So in a way, that’s kind of a big huge win for them. But as I say, much of it now will be done by machine learning. Still that said, Google is supposedly getting worse. I was at an IR conference back last year in Glasgow. And as I said, it’s an academic There are always academic conferences. Very hard to actually get into. Lots and lots of papers get submitted. It’s an academic research conference and they have reviewers and so forth. So I was surprised to see an actual talk about SEO in there as such. But it was basically this German institute that looked and they’ve done quite a longitudinal study about SEO and web search and how Google was actually still ranking SEO optimized pages. And it actually made them national press and many, many It made Search Engine Land and Search Roundtable. Actually, academically, it only had 32 citations. So it wasn’t massively respected in that field, but it did make the national press. And it’s this study. And actually, I was sat there watching the guy doing the delivery of this paper, and he showed this website and actually, I would say that the vast majority of the pages that he showed were things that we’ve come to see as commonplace for ranking for reviews. So I would say it probably wasn’t that spammy. Was kind of, you know, there’s a lot of this best of stuff and so forth. Not great, but I wouldn’t call it spam. And I would say probably, those researchers, no offense, but I didn’t think they really knew SEO. They would talk about links out of the page, whereas most SEOs are kind of talking about links into a page, for instance. So it kind of felt as though, again, you know, it was we’re kind of getting dragged into this as SEOs rather than it being about the search engines and such.

And obviously, it’s gone on to become part of wider, the wider mainstream. We’re ruining the web. We’re hated. The alligator party and the content goblins and so forth. And there’s a veil of suspicion that completely surrounds us. And this is a paper just done just not long ago actually, saying that actually SEOs, we kind of keep everything to ourselves and the vast majority of the time nobody knows what we do. So he’s bashing us, bashing us again, SEO bashing. And but actually, it all seems to take a turn for the worse around 2023. It was the dawn of Generative Information Retrieval. And that’s actually a term that’s been coined in the IR field. It’s going through a huge but reluctant transition, I would say. They don’t necessarily want to change, but they’re kind of being forced to. And obviously, the dawn of Generative Information Retrieval is not like all the other shifts in Google. It’s – you cannot underestimate how huge this is. It’s bigger – and actually this is quite interesting. Funny enough, somebody shared a deck that Jeff Dean had done years ago actually, talked about the many, many times in which Google literally rebuilds their index and nobody notices. They literally just keep re-engineering everything and just kind of doing a cotton shot on the index, just re-engineering this whole thing as it gets bigger and bigger. And you know, it just nobody nobody notices whatsoever. But it’s bigger than that. It’s bigger than Google just rebuilding everything from scratch over and over. 

It’s a total paradigm shift to Generative Information Retrieval, a change to fundamentals of information retrieval as an industry. And it goes back really to the days of library science, to indexing, back to post war times actually. So huge. See, what happened is over the many many years, search has really just had this the same fundamental system. The search is undertaken, results are presented and prior to search engines, this was predominantly in libraries or organisational systems. But this is huge. In classic IR, it’s always been very modular, bits and pieces all connected together. Over time, it’s evolved somewhat. This was actually a paper that was shared by a huge IR researcher, massively respected. He was doing a workshop last year on generative IR and looking at the way that it’s changing in the industry, and showed the transition of the years of the way search engines have evolved. And ultimately shows that over time, it has become quite modular with lots and lots of bits and pieces that take different efficiencies, move at different paces and so forth, run possibly by different teams. This was shared by Martin Bjork from Google last year’s or the year before SIG IR 2023 in a keynote. And he’s talking about this is the typical system. Yeah. Obviously, you have your front door, which is the search interface. The offline, which is where the index is stored, and then obviously everything in the middle. But machine learning has been part of Google search for years already. We have re-ranking and late stage ranking where a recall fetches everything and then there’s an element of personalization, localization and so forth that goes into the machine learned heavy lifting part, which is quite expensive. That’s called re-ranking and late stage cascading ranking. And obviously, mentioned crawl scheduling and so forth, importance prediction, again, learn patterns through machine learning. Pair wise, point wise and list wise rankings, learning to rank processes and index pruning, they have now dynamic index pruning. Even the other week, I think it was Martin Split was talking about how if you don’t have interaction with web pages, pages just get de-indexed and so forth. And that will be part of dynamic pruning because you can’t just keep making things grow more and more and more.

But of course, there are evolutions as well to semantic IR. And this again, shared by Martin Bjork, search is becoming more and more of a hybrid system, not just necessarily your classic IR, but more building in the semantic elements to it. And as I said, it’s been really a traditional pipeline. Index, retrieve, rank. That’s been the pipeline for decades, decades and decades and decades. But now, the Generative IR vision is completely different. What it means is this I mean, it’s quite hard to explain really. I’ll show you on the next slide. A sequence to sequence model is trained to directly map a query to its relevant document identifiers. Basically, it means pulling everything in together without going off into an index and finding the various parts. Just literally generating everything on the fly. It’s huge. It’s a paradigm shift, as I said, that consolidates all information within a corpus in a single model. So it looks like this, yeah. This again was shared during SIGIR Generative IR Workshop. SIGIR is, as I say, a huge IR conference, one of the largest annually that they hold. But ultimately means that the whole of search is changing. Cannot stress how big this is. May not be fully today, but it’s happening right now. It’s literally happening all the while. It’s literally going to change the whole of just the whole model, everything. And to kind of reiterate here and reinforce it, this was only shared by John Mueller a couple of weeks ago here in New York. Talks about how AI could fit in here. Kind of looks very much like I mean, that middle bit, that’s it. That is the whole of search ultimately. As I said, the rest of it is the front door, the search interface, the index is stored elsewhere, and potentially, it’s the whole system. It literally is everything. 

So yeah, that’s it. Literally, it’s just gonna change everything entirely. And whilst this change is underway, what’s happening, ultimately everything’s on fire. It’s like just up in flames. The search results have been kind of, you know, a bit not that great. But instead of SEO fashion, it kind of feels like we should ask if information retrieval as a whole is getting worse because this is bigger than Google. This is bigger than This is like their whole industry, bigger than Google. Just everybody is kind of getting on board with having to change things. So it’s not you as an SEO. It’s not us. It’s them. Yeah. It’s them. It’s them that’s making Everything’s kind of a bit of a shit show at the moment because of them, not us. 

Yeah. So why the rush to Generative Information Retrieval? Well, one might describe this as the information retrieval’s existential crisis. Is at the end for classic information retrieval. So as in SEO, where we have like people who are specialists in local SEO or technical SEO or digital PR, link building, whatever. They also have in the IR field, people who are also specialists in in various areas of research and in academia and industry, natural language processing, recommender systems. Obviously, Discover is a recommender system and Spotify, Netflix and so on, which is also part of IR is recommender systems. Classic IR, ethics and fairness and responsible AI. Ricardo is an absolute mean, Ricardo is an absolute expert in everything in IR. But it’s an area he’s focusing on, ethics and responsible AI. Medical IR, huge because of the unbalanced data that is in that industry, and obviously the very great deal of good that it can bring, and the knowledge graphs as well. But the NLP crew, when Bert came around a few years ago now, they actually were blazing hot. They have all these leaderboards, I was competing with each other one on, they were literally on fire. They took IR forward, you know, thirty years in five years. The same old fashioned industry had been the same for decades. And all of a sudden, VirTra came, literally rocketed everything forward. And now all of a sudden, this massive like overflow of like huge amounts of different models came through. We know some of these, the various BERTs and the GPTs and all the stuff that OpenAI did and the bings and so forth. Yeah, absolutely huge, just one after the other. Just absolute cascade.

And it could be in the classic IR mind, Well, could it be that we don’t really need classic IR anymore? Can search rely just on NLP and large language models and so forth? And then if you remember, I wrote this a few years ago for Search Engine Land. They started talking about how actually search engines no longer need humans quality raters, they’re going to use large language models instead. The guy who was one of the researchers, Charles Clark, he does – he’s an IR researcher professor in Canada. He did a talk at SIGIR a couple of years ago and it was incredibly unpopular in the IR community. He was basically saying, hey, we don’t need humans anymore to do evaluation and to be quality rated in search, large language models can do it. 

So basically, everything seems to be going against classic IR. Is it dead? But actually, turns out that NLP and language models are not actually that good on their own, unchecked. They’re hallucinating, as we know. They are not faithful, they’re not faithful to results, they are not factual, unfortunately. So, know, for an example here, you know, there was I don’t know whether you’ve come across this, but there was a really quite famous example of a lawyer that literally cited lots and lots of cases that he’d come across, past, present and so forth, went to court, turned out none of them were true. None of them had actually happened. So there is you know, it’s just pretty disastrous really. You know, we all remember the cheese not sticking to pizza. And I believe that the classic IR and knowledge graph crew were probably chuckling their heads off of what was happening as these were all being rolled out. Yay, classical IR is not dead.

Welcome to the world of SEO. I mean, we’re constantly having to dispute that SEO is dead. Hey, we’re getting the gang gang back together. I was gonna put on something from the Blues Brothers, but then was like, will anybody actually know what the Blues Brothers is? Because maybe I’m too old and you know, yeah, you know. Does anybody actually know what the Blues Brothers is? Okay, yeah. Fine. Okay, well, now I feel like I’ve failed. Okay. So and actually, this is the major reason why obviously they’ve had to rush because ChatGPT and our large language models were literally threatening search. That was it. They’ve had to like save their whole skin as an industry. Also started to get philosophical and Mike touched on this the other day. Back to Andre Broader, here he comes again. Whenever Andre Broader starts talking about concepts, search changes, literally, that’s it. He brought out this paper 2023, talking about the notion of Delphic costs, saying ultimately, search needs to change because instead of us sending people to go off and search through websites, there is this cost to the searcher called a Delphic cost about the search for knowledge. Really, it’s an expense to the searcher because they’re wasting their time. Why we send them to results that are not that great? It’s effort and so forth. And that is the disadvantage we have against ChatGPT and large language models and so forth. So we have to basically bring the knowledge to them. Hence, the AI overviews and the generative IR. And that’s gonna get bigger and bigger and bigger. Ultimately, search is not free to the user. There’s a notion of effort, especially with these next generations coming through. This is Broadus point. So bring the knowledge to the user, reduce the friction. Instead of sending the user to research themselves, hey, we do the work for them and bring it bring it to them in AI overviews or whatever, whatever method. 

And basically, and we’ve seen this. Only a few weeks ago here at Search Live New York, I was like, hey, I recognize that. Making search effortless, Broader paper. Yeah, making search more effortless than ever. You are gonna see this be talked about more and more and more. This is gonna be justification over and over again for AI Overviews. So goal is our goal is to make search effortless by synthesizing multiple related queries into one clear personalized response. That’s it. Yeah, regardless of how bad it is at the minute. And also there has been push back in their own space. It’s not just us saying that these are not great experiences. Their own the IR space itself has been pushed back against generative IR, generally speaking for a while. There was this paper that Google submitted, rethinking search making domain experts out of dilettantes. This was the second paper that they submitted. The first one talked about, you know, we’re gonna make domain experts out instead of librarians. Well, the IR field was very offended because in actual fact, they thought that they were insulting librarians and obviously the origins of search is in library science. There was real push back. People were saying, hey, it’s like people actually do love to search. There’s been a real huge push back. And Charles Clark, again, finally came back and said, by the way humans, there’s still hope for humans now as quality raters rather than just AI. As I said, Shar and Bender, they were the ones who were like, hey, people love to search and also there’s this toxicity in in generative IR. There’s been a huge whole in their industry. So also Google has been playing catch up. This was only on Wired a couple of weeks ago saying that they’ve raised, they’ve lowered the bar. They’ve they’ve had to like take a they’ve had to do a lot of cut and shut jobs basically. They’ve lowered their standards. In the meantime, minimum viable product. And as we’ve seen, it’s not robust. As soon as you start to extend it to out of distribution, because there is this notion of closed book and open book in generative IR, we start seeing things like this, the rocker day and the health benefits of running with scissors. And because the reason is generative AI is not based on facts, it’s based on patterns. That’s it. You know, you can’t kinda make facts. It’s just regurgitating patterns over and over again. I’ve got a lot of slides, I’ve got to get through these.

Okay, so we know, we saw this the other week as well. AI search is wrong more often than it’s right at the moment. It is going to get better. It’s like a dodgy photocopier. Keeps like, and it’s not just me saying, well it isn’t me saying this. I saw this though, I just put the picture of the copy on there. It’s the National Cyber Security Service that literally said, it’s like a photocopy that just gets worse and worse and worse the more you repeat it. So it’s that downward cycle. And it’s kind of the new Guess SEO. Guess SEO. Yeah. This Joy Hawkins talked about the terrible local results that generative AI produces. It’s not robust with news. Bloomberg has to keep repeating like summaries and putting out apologies because it’s just regurgitating nonsense. The problem is we are back to disambiguating that probability in search. But this time, instead of it being featured snippets like it was a few years ago, it’s now for generative IR, where search engines basically assign equal probabilities to outcomes when they are judged to be possible. i.e., when two options are available, which one’s the likeliest? Because you’re just picking up on patents. Probability determination. 

This one, is from years ago, from a deck I did, well, five years ago, six years ago maybe. Talk about why feature snippets used to pull pictures from one thing and words from another. It was literally this notion of equipossibility where it’s like, well, we’re not quite sure, so we’ll pull that picture and we’ll pull these words. I mean, was the best one I remember. Teenage Mutant Ninja Turtles becomes a work of art. Michelangelo, there you go. That was years ago. But we’re kind of back to that again and that’s what’s happening. The problem is none of this works where the stability of a reliable index is stable memory. We’ve seen ChatGPT talk about memory as it grows now. They’re gonna try and build a memory of the individual user and so forth. And obviously that relies on consistency. So how are they looking to resolve it at the minute? Well, they’re doing a huge amount of ferocious research. They have their own like field now. Generative IR is literally just the second year of this. They’re saying, look, is this just a trend or is it as some claim a paradigm shift for IR? It is paradigm shift. I think they’ve kind of agreed on that now. They talked about the hero of generative AI. Yeah. So it’s a brave new world. Years ago, 2017, I went to the European Summer School for information retrieval. It was all about all the things we know and love, crawling, indexing and so forth. This last summer I went again. It was literally everything that we know from IR, from search, sorry, web search was all shopped into one session. Forty two things you didn’t know about IR. The rest of it was all about machine learning, AI, recommender systems, totally different. So it’s changed. And as I said, everything now is around the notion of Gen IR.

And we have two parts to it, open book and closed book. Closed book is basically where the model just literally pulls from what it has built into the model tends to be out of date and so forth. And now, it’s extending to open book, which is being augmented with other knowledge. And hence, we have RAG. Hence we have the classic IR field is now back in the game. It’s back in the game. The gang is back together. They have a job again. Yeah? Grounding. We have to like reinforce what large language models are saying because they just make nonsense up. So this is why grounding and RAG are really feels where the classic IR people can come back in. For connecting the AI model to verifiable source of information and reducing hallucination. And again, Google is starting to talk about this a little bit and I did notice that on LinkedIn, they started to talk about they’re gonna be doing another search summit where they’re gonna go deep and stuff. So maybe they’re gonna start telling us more and more and more about this stuff at some of their their search summits. So they’re touching on this. A lot of it, as I said, is being tested in Gemini and so forth. That’s kind of almost like word the grounding is being reinforced. And obviously, we know that in some cases, grounding is not required where it’s really obvious, but there is a temperature apparently with these things. Sometimes it’s needed, so we’ll have to adjust our strategies accordingly. Pro tip, if you do block block Google extended, you can’t appear in anything to do with Gemini. But at least and also now, as we enter this field of retrieval augmented generation where classic IR and LLMs are working together, it’s again like the old Bert, where Bert started with just one and then it became absolutely tons of tons of Berts and lots and lots of models. They have twelve different types of RAG that are going around. I particularly like the golden retriever RAG. Great for me if you’re a dog lover. 

Now, this other approach is also being taken. Knowledge augmented generation, cash augmented generation. This one stores knowledge in the cash, not the knowledge augmented generation pulls in knowledge graphs together. So it’s all lots of different approaches trying to like make this a lot more efficient and so forth. This, Andrea shared this and I spotted it. This is something that Microsoft are doing, who are huge, absolutely huge engine, huge and fantastic engineers. This is an approach that’s being taken which is much more efficient than some of the other retrieval augmented generation approaches, has really good legs on this. And it’s around the attention here. The attention is really what drives a lot of the large language models. This is much more efficient due to its linear approach rather than this notion of quadratic dependency which makes it too expensive.

And this was again something by Google. This is called the Exalta framework. This is where they’re trying to understand what happens in news. Why news hallucinating as badly as it is? This is a paper again by Martin a joke. And much of this, this is my point really, much of this requires assistance from SEOs. We still have a big job to play here. We have to assist with equity probability disambiguation. We have to recommend different approaches to schema and building knowledge graphs and so forth. Whether you agree with LLM’s text or not, we have to weigh up the pros and cons. We have to monitor for Bing indexing because you can’t get into ChatGPT if you’re not indexing Bing. That’s fact. Yeah. There’s lots and lots of things we need to do still as an industry. All these, yeah. We have to be big advocates for extreme consistency because the bottom line is consistency will be king or queen or whatever with this new generative IR field. Google are telling us, you know, we need further context. Well, who’s gonna recommend that further context? Well, it’s us. It’s us as the SEO field, the SEO world, because of imputation where there’s too many gaps basically in the in the data.

And that’s not going to change anytime soon. So basically, things like clustering goes a long way to adding extra content in unsupervised learning. Who’s going to recommend the clustering? Building the keyword clusters and the content clustering. SEO heroes, that’s who. We’re the bridge. We’re the bridge between developers, between data scientists, between the companies, the CA, CO’s, the marketers. We hold everything together. So we still have a huge part to play in all of this. So yeah, know, we need you’ve got to give us content. It’s a two way street there. How dare how very dare we want actually to be part of this? Well, we do. So we may be a necessary evil, but we are necessary. So sending traffic to us is definitely something you’ve got to do. We are the SEOs in the middle, helping them to get through their existential crisis as they muddle through. And you know, I love Fabrice, Canal and Bing, but you know, it’s true. They need our help. We’re not villains, we’re heroes.

And as Lily said in her great deck that she did on Brighton the other week, I thought that’s amazing. That’s exactly the slide I won. SEOs are the internet army. That’s what we are. Yeah. So we’re kind of their annoying sidekick that they just cannot shake off. Rather than is SEO dead, we should always ask is information retrieval dead? Well, no it’s not, says Martin the jock. There’s plenty still for all of us to do, and that’s the point. 

And my last two thoughts on this, I’m seeing a lot of like, hey, is it GEO, is it SEO, Is it AI? Whatever seo? Who cares? You know, who cares? I mean, bottom line is please let’s not waste our time on things that are kind of pointless discussions. We don’t want to end up in a subfolder versus subdirectory discussion for the next decade and let get left behind. Whatever. Keep calm and SEO on. So thank you for your time and for hanging around. Thank you.