0:02Hi, I'm Julie. So I run product at Squiz. We build the platform that's behind a lot of university and government websites. So I spend a lot of my time thinking about what happens when the visitor is an agent instead of a person. So here are three MCP servers. With this one, someone made an 80s synthwave track in Ableton from only two prompts. This one lets Claude build. 3D scenes in Blender, Weird Dragons included. And this one covers off a university asking about the fees.

0:40Sorry, one second. I don't have any notes. Thank you. Right, so what is MCP? So MCP is a model context protocol. So it's an open standard for connecting AI apps with tools and data. And it works just like a menu. So the agent asks what you can do, and you hand back a list of tools, and it calls the one that it needs. So in this menu, you can see here two of the tools answer questions, and one of them books a table for you. So WebMCP, which is a new browser standard, currently in OriginTile and Chrome, brings the same idea to web pages. So an agent in the browser can call your site's tools instead of just guessing at what the buttons can do. This is Cloudflare's WebMCP. preview. So you go into agent readiness and you flip two switches and Cloudflare adds a script tag to every single page at the edge. Nothing changes at your origin.

1:29So these tool switches, these switches are tool packs. One reads the image provenance and the other one takes the tools from your own MCP server and hands it over to the visitor's browser agent which calls them using the visitor's own logged in session. So this is a really big step up from agents use your site today so without tools an agent reads your pages clicks all your buttons in the way that a person would hopefully which is slow and can break whenever the design changes agents can also get really lost if you've got too many buttons on your page and agents are going to do this whether you switch on this tool or not they don't need your permission or an MCP server and if a visitor is signed in the agent can use also use that session as well so tools are the best chance that you have to give the agent the right information to start with, even though the agent is still deciding what it does with that.

2:25So with tools, it can call your server and get a clean answer back. And because it runs in the visitor's own session, it can also do things with their account, like checking the status of an application without you building a separate login for agents. So if you've already built an MCP server, you get all of that without building anything new. This switch doesn't take anything out of it. So if you have built an mcp server it won't really do anything and but there's a few things that are kind of keeping this a little bit small right now so you do need that mcp server and web mcp is still in origin trial so most visitors browsers can't use it yet but both of those limits are shrinking so we all heard in the keynote this morning matthew prince said automated traffic has overtaken human traffic on cloudflare's network uh that eight months earlier than he even predicted so these agents that are running the internet now they don't have to wait for Chrome so Cloudflare's own agent browsers they've got browser run and kitesurf can already find and call webmcp tools and microsoft edge is also running its own webmcp trial as well the next limit is also going soon so Cloudflare AI search can already crawl a site you don't even need a sitemap and give you an mcp endpoint it.

3:44And Cloudflare has already said it's going to add ready -made tools that let browse agents search your site. So where all of these come together is now your site is going to be open to agents and you'll need to make sure you're controlling what is actually given back to them. And then where I think this is really going to end up is the visit is probably going to go away. So at the moment the agent still goes over and visits your site. If it can just pull the data back in, the visit is going away. So it's really, really important that the agents understand what your site can and cannot do. So what belongs behind slash MCP depends on what the agent wants to take from your site. And there's a few different answers depending on what type of organization you are. So for retailers, the agent is showing up as a buyer.

4:31So Cloudflare, as we saw again this morning, they're working with Visa, MasterCard, American Express. on verifying trusted shopping agents and launching wallets so that agents can actually pay. So behind MCP, if you're a retail org, you put search and you put your checkout. For publishers, the agent shows up as kind of a taker. So CloudPlayer's new defaults, they block the mixed-use AI crawlers on ad-supported pages, and they also have pay-per-use now, which pays publishers when their content shows up in an answer. So if you're a publisher behind MCP, content but at a price. Services organizations, things like universities, government agencies, health providers, banks, airlines, the agent is the go -between. So it sits between them and the decision that happens somewhere else like an enrollment or a benefit claim.

5:22The closest thing that's been built for them on Cloudflare is the visibility tracking and that AEO dashboard we saw this morning as well which can tell you whether an assistant mentioned you but not whether what it said was right. So for services organizations the website is shaping the decisions that bring the money in even though the payment might not always be happening on that actual website. So choosing a degree takes weeks or months rather than the minutes it takes to buy shoes and the website is where people check those facts along the way. Think of government websites if you've ever gone there you're going there for information. And if that information isn't accurate and useful and easy to find, the university drops out of that decision and it loses that student enrollment.

6:11Same goes for banks if someone's trying to choose a credit card or a super fund if someone's deciding whether they want to switch. So the money is still changing hands somewhere else, but the website needs to help decide where it's going. Services organizations are better placed than the panic is kind of suggesting. many of their most valuable visits are actually those logged in moments that no answer engines can touch right now. So when is my library book due? Hey can you extend it for me please? ChatGPT can't see your loans. WebMCP tools running in the visitors own logged in session so everything Cloudflare shipped for agents that preview actually fits services really really well. But to make the most of it they need a couple of things.

7:01They need to be correct, which almost matters more than being mentioned. They really need to be able to keep their own voice, their own brand as well, because the agent is rewriting everything. And they need tools for those logged in moments with a clear line between answering and acting. So the difference between checking the specials and actually booking a table. So telling a student the due date for one of their assignments is kind of low risk as long as it's right. But extending the loan. an application or accepting terms and conditions those tools need to sit behind human confirmation

7:37so an agent sees three things from you they see the tool names the tool descriptions and the tool responses and it decides what tool to call by reading the descriptions that you give it so a description is the first piece of writing that an agent ever reads so it's kind of like your agent dating profile So on the left, this is actually the default from Cloudflare AI Search. The name is Search, and its whole bio is find exactly what you're looking for, which is kind of the tool version of ask me anything, lol, on a dating profile. There's nothing about what it brings, what it's looking for, or why anyone would want to swipe right on it. So the agent either swipes left or it matches and then asks it questions it was never really meant to answer. So on the other side, that is the MCP server we run.

8:25for an our own tool which serves answers that squiz content intelligence generates from our site and it validates it says what it does where the answers came from and that every answer comes with its source page the owner and the last review date it also says what it's looking for and when to swipe left which is more than most dating profiles manage honestly so this is the first place your brand voice is meeting an agent so we tried it ourselves We built a small test site. We imported some content over from one of our own university customers and indexed it with Cloudflare AI Search, which gives you that MCP endpoint over the end of it. So without tools, an agent loads pages and reads them and kind of guesses its way through. With tools, it asks one question and gets the answer back.

9:12The retrieval worked really, really well, but it did return the mistakes on the site along with everything else. So these are three. real responses from that test. These are real questions that we asked the Cloudflare MCP server and one of them should worry you or worry the university we got all this content from. But it's really hard to tell them apart because they're all formatted the same way. So at C, we asked what are the fees for a nursing course. The top result was the fees section from the physiotherapy page. The second result was from the occupational therapy page and then the nursing page's own figure. of 36 ,500 came third. So AI search did what it was meant to do. It found the fee section that matched the question and our test rate test site did contain all those figures.

9:57But finding the relevant content doesn't tell the agent what source is authoritative and nothing in the response did either. So the response is that third thing the agent sees and it's where most of your signals really are disappearing. So have a look. here you've got the actual Bachelor of Nursing page no one would mistake this for anyone else if humans looking at it great big Bachelor of Nursing along the top but in the response the page is cut into these chunks and ranked against everything else on the site and the nursing page came third so your page carries a lot of cues a person uses to judge it the layout the design the date the tone your brand the tool response keeps the text and drops the rest And then the agent rewrites that text in their own voice.

10:49So some of this can come down to configuration. Our tests use the defaults, vector search only. We didn't do any re-ranking or query rewriting. How the site is crawled, the chunk size, the overlap, hybrid search. There's a lot of things that you can change what comes back. But it is still really important. You need to fix the content first and then you can tune the retrieval and test both with real questions.

11:15the reason this is so important is your organization is going to wear the mistakes so when the european broadcasting union and the bbc they tested ai assistance on news sites 45 percent of the answers had a significant issue it was wrong and that had the effect of 42 percent of people not trusting the original source ai made the error but the organization carried the risk you So there is a good model for this in Cloud Play Zone Preview. They've got the Content Credentials Pack, which is reading on image provenance and marks every result as a signature not verified, which is handy if you want to sell feet pics online. It can tell your agent who claims to have taken the photo, but it won't tell anyone whose feet it is, so it's fine.

12:01So the pack tells the agents how much to trust what it's getting. Your content tools can do the same. It can send back the source URL, when the content was last reviewed, who owns it. who it applies to and what and what effects it should have. These are the cues the page used to carry. AI search does send one date through but it was the system timestamp so in our test it was the day we set up that example site not the time anyone checked the content. So Cloudflare does have this provenance pack for images it doesn't have it for any other type of content so Cloudflare people that would be a great feature request if it did structured content. Thank you, thank you. Okay, there are a couple of other ways that you can also keep your own voice. First, if you give the agent an approved answer to quote, if it's written and checked by the people who own the content, so it's got less reason to actually paraphrase that answer.

12:52The second thing you can do is use MCP apps. So this is an MCP extension that lets a tool return an interactive interface inside the chat, so where your customers are. It already ships in Claude, ChatGPT, VS Code. so Cursor and GitHub Copilot, and millions of Shopify stores were sending those interactive components into chats before it even became a standard. So for organizations, it means that you can put things like your official fee tables or your course card can show up inside ChatGVT looking like you in your own words. One thing that we did is on our real site, we started with an audit. So we have a product called Squiz Content Intelligent. It asks thousands thousands of questions a real visitor would ask and then it flags the answers that were wrong or incomplete or contradicted on another page. This is one of them.

13:38I did admittedly pick a fairly low stakes one but one page said disputes with Squiz in New Zealand go to the Arbitrators and Mediators Institute and one page said they go to the New Zealand courts. Kind of low stakes but for every problem you end up finding you really need to make sure you either fix it or exclude it from the tools. We ended up finding on our site which has about 40 pages, 147 high priority conflicting or incorrect information. We fix them all, but we regularly re-order to make sure all new content is really kept up to date.

14:17So there are four things that you can do before you turn on that MCP toggle. The first, you have to decide what you want the agents to do for you, for your organization. Some questions they can answer on their own. own and some tasks like submitting an application or changing an account, you want a person to confirm. So think about how you want agents to serve your organization. Make sure you get your tool descriptions and responses accurate. They need to be in the format the way an agent would and make sure every response is carrying the date, the owner and the source.

14:54This might be a little bit specific to us, but check what the switch is adding. pages. So features like WebMCP preview, it injects the script at the edge. But if you're using a lot of workers like we do at Squiz, sometimes the two can fight, which is exactly what happened when we turned it on. So if you're in the same spot, feel free to ask PlaidFlare for the ability to inject these scripts into your own workers instead of having to use theirs so you can control and run in what order that they will run in. If you don't use a lot of workers, you probably won't have this problem.

15:28and audit your content check it ask the questions look through your search logs ask the real questions that people would be asking your content if they're asking an agent over here and make sure they're getting the answer you want them to get

15:43so cloudflare is making all of this stuff easier every month that roadmap has been coming in the matter of sort of months not years and it's built a lot of things into this space you The problem is stakes. If the Blender agent gets it wrong, you get a really, really ugly dragon. If an agent gets your organization wrong, people are going to hold you responsible. So agents are going to read your site, whether you flip this switch or not. So you may as well do the audit and then turn it on and make sure you've got all the descriptions in there. And it'll go a lot easier for every one of your customers who is trying to use your site. Thank you.