<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by C Logesh on Medium]]></title>
        <description><![CDATA[Stories by C Logesh on Medium]]></description>
        <link>https://medium.com/@logesh_c?source=rss-38a0f20609fc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by C Logesh on Medium</title>
            <link>https://medium.com/@logesh_c?source=rss-38a0f20609fc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 02:32:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@logesh_c/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[We built a website for AI agents and measured it. Two results were good and two were not.]]></title>
            <link>https://medium.com/@logesh_c/we-built-a-website-for-ai-agents-and-measured-it-two-results-were-good-and-two-were-not-7847e815399f?source=rss-38a0f20609fc------2</link>
            <guid isPermaLink="false">https://medium.com/p/7847e815399f</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[seo]]></category>
            <dc:creator><![CDATA[C Logesh]]></dc:creator>
            <pubDate>Sat, 21 Nov 2026 09:32:19 GMT</pubDate>
            <atom:updated>2026-11-21T09:32:19.119Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What 730 test runs showed about agents, MCP, generative UI and getting cited by AI assistants.</em></p><p>Businesses are starting to ask whether their website needs to “work with AI”. Most of the answers we found were opinions or vendor claims, and we did not have one we could defend.</p><p>So in October 2026 we did two things. First we wrote a research study from public sources. Then we built prototypes to test the claims in it, with pass or fail criteria written down before each run.</p><p>This article is about the second part, because that is where our own numbers are. Some of them contradict what we expected.</p><h3>The starting point</h3><p>The study is called <em>Websites for the AI Era</em>. It draws on 139 sources, and it argues that a business website now serves several kinds of visitor at once: people, search engines, AI assistants that summarise the site, and software agents that act for a person.</p><p>Two figures from it explain why we cared. By Similarweb’s US measure, 68% of Google searches now end without a click to a site outside Google. By Cloudflare’s and Imperva’s counts, more than half of web traffic is automated.</p><p>An outside reviewer told us the first version leaned too much on secondary sources. We went back and re-checked every source. 104 held up, 24 needed a caveat, 4 did not support what we had written, and 11 could not be read any more. We published the corrections and the full list of graded claims next to the study.</p><p>That review ended with one piece of advice: stop reading and start measuring. So we did.</p><h3>The setup</h3><p>We built a reference site for a made-up business, Northwind Looms, which sells industrial sewing machines. It has three products, one service and a few FAQ answers. All content sits in one store, and every page is served as HTML for people, as structured data for search engines and as Markdown for assistants.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DSH-PLNBLP7jMPgEfAry3A.png" /><figcaption>The architecture our study proposes: one content store behind a human site, an agent surface and a generative layer. The prototypes test parts of it.</figcaption></figure><p>Then we wrote a test harness. It gives an AI agent a task, such as “find the delivery time for this machine” or “request a quote for two units”, and records whether the agent finished, how many steps it took, how many tokens it used and whether it did anything wrong.</p><p>Every run used one model, Claude Sonnet 5.5. Keep that in mind for everything below.</p><h3>Result 1: a well-built ordinary website already works for agents</h3><p>Our first test had no agent features at all. It was a clean, accessible site, and an agent driving a browser.</p><p>The agent completed 98% of 150 runs. One of the failures came from a form field whose label was not correctly linked to its input. A person would not have noticed. The agent could not tell what the field was for.</p><p>Browser agents read a page largely through the same structure that screen readers use. Accessibility work that many teams treat as a compliance task is also what lets an agent use the site. If you do one thing after reading this, audit your forms.</p><h3>Result 2: an agent API did not save tokens. It saved actions.</h3><p>Next we built what we call an agent gateway. The site’s operations (search products, check availability, look up business details, request a quote, book a service visit) are defined once, then exposed both as a REST API and as an MCP server.</p><p>We ran 300 more executions: the same five tasks through the browser, through REST and through MCP.</p><p>All three paths completed 100% of tasks. That surprised us less than what came next.</p><p>The API was more expensive on simple tasks. The browser agent answered read questions using about 5,500 tokens. The API agents used about 8,100 for the same questions, because the definitions of all five tools are sent with every turn. On a three-product site the page already contains the answer, so reading it is cheaper than calling a tool.</p><p>The API won on the task that changes something. Requesting a quote took the browser agent nine actions through a form. Through the gateway it took two calls. There is no form to break, and a wrong call comes back as a clear error and not as a wrong click.</p><p>REST and MCP were the same to the agent. Completion, steps and tokens matched within 1%. The model sees the same tool descriptions either way. Which one you offer decides which clients can connect. It does not change how the agent behaves.</p><p>Speed was not the argument either. A scripted client finished a task through the API in 6 milliseconds. A real agent took about four seconds on every path, and almost all of that was the model thinking.</p><p>We also wanted to know whether a person stays in control. When an agent files a quote request, the gateway does not submit it. It returns a link. The person opens it, sees exactly what the agent asked for, and approves or declines. In 20 of 20 runs that took one screen. We also tried 20 ways to get around the permission rules, including replayed signatures and requests for another business’s data. None got through.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9Fs1crqy2sKCMKZKqjfsTA.png" /><figcaption>The operator dashboard after the test: every call by operation, path and outcome.</figcaption></figure><p>We expect the token result to reverse on a site with hundreds of products, where the answer no longer fits on one page. We have not measured that yet, so we are not claiming it.</p><h3>Result 3: generative UI did not help anyone finish</h3><p>The third prototype was a chat assistant that answers by assembling real interface components (a product card, a comparison, a quote form) from a fixed set of nine. The model cannot invent a layout, and every number it shows must trace back to a record in the content store.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3YYATTGdqM6OcNNVHMsXDw.png" /><figcaption>The assistant answering a comparison question with a component from the fixed set. Each figure comes from a product record.</figcaption></figure><p>Our pass criterion was that people or agents complete more tasks with the assistant than with normal navigation.</p><p>They did not. Across 280 runs, completion was 100% with the assistant and 100% without it. On a site this small, navigation is already easy, so there was nothing to improve. By the rule we wrote down in advance, that criterion was not met.</p><p>What we did learn is the cost of running it:</p><ul><li>A typical response took 2.9 seconds. One in ten took more than 5.</li><li>A conversation cost under one US cent at the token price we assumed.</li><li>8 of 140 answers were rejected by our own checks before reaching the visitor, 7 of them because the model stated a number it could not trace to a record. All 8 were corrected on the second attempt.</li></ul><p>That last point is the one we would keep. The checks caught an invented number about once in every twenty answers. Without them, those numbers would have been shown to a customer.</p><h3>Result 4: formatting content “for AI” got us zero citations</h3><p>There is a lot of advice about writing content so that AI assistants cite it: put the answer first, show a date, quote statistics with sources.</p><p>We tested it. We wrote ten pairs of pages. Each pair holds exactly the same facts, checked by automated tests. One version follows the advice. The other is ordinary prose. We published all twenty on a public site and asked an assistant the ten matching buyer questions on a weekly schedule.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r4ALdLcPSL37wVj8RtcZrA.png" /><figcaption>One of the ten pairs: the shaped version (left) and the plain version (right), with the same facts.</figcaption></figure><p>After three runs and 30 answers, the shaped pages had been cited 0 times. So had the plain pages.</p><p>The assistant cited at least four sources in every answer. They were marketplaces, directories and established dealers. Our pages were never among them.</p><p>This does not show that the advice is wrong. It shows that the advice made no difference for a new site that nobody links to. Our reading, which is an interpretation and not a finding, is that an assistant decides whether to trust a source before formatting has any chance to matter. The experiment has also only run against one assistant so far, and it is still running.</p><h3>What we cannot claim</h3><ul><li>Every run used one model from one vendor.</li><li>The business is fictional and has three products.</li><li>All agents were our own. No outside assistant has discovered and used the gateway.</li><li>Costs are calculated from measured tokens at a price we assumed.</li><li>None of this has been tried on a live business site, or with real customers.</li></ul><p>These are engineering results from a test bench. They are not proof that any of it earns a business more money.</p><h3>What we would do today</h3><ol><li>Fix accessibility and structure first. It is the only work here that paid off in every test, and it is required by law in a growing number of places anyway.</li><li>Add an agent API for actions, not for reading. If an agent can already read your page, an API for the same content adds cost. An API for booking, quoting and ordering removes fragile forms.</li><li>Do not pick between REST and MCP on performance. Define the operations once and offer both.</li><li>Keep a person in the approval step. One confirmation screen was enough, and the agent passed the link on every time.</li><li>Treat generative UI as unproven on small sites. If you build it, validate every number against your own data before showing it.</li><li>Do not pay for “AI formatting” on a site with no reputation. Being mentioned on the platforms assistants already trust looks like the larger factor.</li></ol><h3>What comes next</h3><p>The next measurements are the ones that could change these conclusions: a large catalogue, more than one model, and a live business site with real visitors. We will publish those results in the same way, including the ones that go against us.</p><p>If you have run similar tests and got different numbers, we would like to compare.</p><p>This research was done at Sysbot Technologies. The full study, with every source and its grade, is here: <a href="https://sysbottech.com/research/websites-for-the-ai-era/">https://sysbottech.com/research/websites-for-the-ai-era/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7847e815399f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>