 So, my name is Feroz Park and I'm going to be talking about title snippets and result previews. My general area of focus is core ranking in search, but the last few years I was asked to look into this area about result previews and there's a pretty good reason for it. If you think about it on a search page, one of the large tasks that a user has is to sort of find the content they want, find the page they want to get to. And in the world of Ten Blue Links, this was the only information they had literally was a title and a snippet and perhaps the URL. So part of what we did was we sort of rethought what would be sort of the important pillars for an effort around result previews. So primary to our goal of course is relevance to the user need. We certainly cannot have previews not be relevant to the user query. Our second goal would be to make sure that whatever preview we generated actually helped users make a choice. We wouldn't just generate previews, apps in mind of that. A third one would be we would want to illuminate the depth of content within websites. I personally believe that there's a ton of content out there written by wonderful people out there like yourself and it's important for users to understand that there is a lot of relevant information within sites and if they don't know that information is within that site, they're unlikely to visit it. So this is a large pillar. And the final one was this notion of diversity. There is a ton of different kind of content out there. It goes from imagery to videos to forums to lists to tables. I mean it's sort of boundless I think in terms of human imagination. And so we needed to catch up and we need to be able to express the diversity of the content that was out there in the ecosystem. And our sort of sense was if we did this then people would benefit, users would benefit and then consequently the ecosystem would benefit. So here's what we had to work with. We had to work with something called attribution which is the source of the site, the name of the site, the title itself for the URL and then this sort of abstract notion of a content preview with its character limits and size limits. So the next few slides I'm going to talk through some of the newer formats that we sort of have built out. These are all algorithmic driven by those original pillars. But as you'll see they're actually quite useful for users. We were putting a lot of effort. It's not that the previews themselves were just built without effort. Here's an example of us attempting to change the snippet in response to the query. The same algorithm that decides that a page is relevant for a query can be used to decide what part of the page should be relevant for the user if they were to choose it. So if you change the query you can expect your snippet to change and you can expect the user to read that to understand if that page indeed is relevant for them or not. But we can do more than that. So an example, one of the things we discovered I think it was around 2016, 2017 was that images really help users choose between different results. This latent desire, I want to see a particular type of image, I want to see a particular type of flower or I'm looking for a particular type of shoe. And the image suddenly becomes exceedingly relevant for them. So this was, we can algorithmically extract relevant images from the page and show them. The placements on the right, this is important because you want the image to be secondary to the main source of information which is the title in the snippet. Occasionally we can extract galleries. This is when the content of the page cannot be rendered by a single image. Rather the user would benefit if they saw a sense of the kind of content, the kind of imagery on the page and we make sure not to use all the images. There's limits on what fraction is used. And as a consequence we noticed that users actually would start visiting a greater diversity of sites. So this was sort of a very strong, in a strong sense this helped us understand that this can benefit the ecosystem. Following images the obvious thing to do is then look at videos and you can understand similarly that you're going to look for relevant videos. There could be multiple videos on the page but you're going to look for the relevant one. And you can sort of mutate your preview type depending on whether the video is the dominant content on the page and you'll see this notion in subsequent previews too or whether it is only supportive and therefore on the right. If you wanted to express the depth of content within sites then the sort of within that height limit of ours, the only information we have is this ability to sort of generate what are called site links. And these are links to pages within your site that we believe are relevant for the user. So if you're looking at a restaurant then clearly contact drinks, private events, finding tables. These are all links that can be very useful for users. These are all extracted algorithmically. Occasionally the site structure does help us and it drives traffic into your site. Sight link images was another change we added when we realized that if their page had image previews then it should be simple for us to again algorithmically decide what the site links could have images. And this actually helped in cases where the site links don't really help the user disambiguate the type of content that's on your site. We could go further. We went down this path of entity facts. This is when, if you can imagine, as in the bottom right, if you're, in this case I think Killen Jorné is a wrestler. And there are facts around this wrestler that are pertinent to the user. Entity facts are a way for us to sort of illuminate some of those facts so that people can then say, oh, this site is pretty legitimate. It does explain facts about this user and therefore it's deep and therefore I'm willing to go to this site. Similarly, forums. I think Allen had talked about structured markup to help us understand forums. But there are tons of forums, community sites. There are tons of forums out there that don't have structured markup on them. And it's a big disservice to users if we're not able to help people understand that this is a forum and help people understand what kind of contents out there. So algorithmically we can extract posts subject to our ability to do so. We choose the forum cluster. There's a set of related content on the site that might be related to the query for the user. We certainly see users engaging with this a lot. And we know that if you're going to provide us with markup, then we certainly want to use that. And so markup always takes priority over these algorithmic forms. We can do the same thing with tables and lists. So if your site has a dominant table, and we want to show the preview for that page, and clearly showing that table to some extent is a better thing for the user, because they can understand this is really what the content is on this page. And the structure and the position of that HTML does guide the preview. So in summary, relevance infuses everything. If these previews have to be relevant to the user, they have to be relevant for the query. Attribution helps the user choose the source. We make a lot of effort to ensure that the depth of your content is well expressed. And therefore, a user is sort of invited to visit your site. And you can see that both with site links and forums. And we make sure that we have a diversity of preview formats to sort of support the ecosystem. Thank you.