We Scored Over 100 March Madness Sites
for AI Agent Readability.
The Results Will Surprise You.

Every March, millions of people build brackets. Increasingly, they are using AI to help, feeding game data into ChatGPT, building prediction agents that scrape stat sites, asking Claude to analyze matchups, and deploying automated systems to trade on Polymarket and Kalshi. The tournament has become a proving ground for agentic sports intelligence.

There is just one problem. Most of the websites these agents rely on are structured for human eyes, not machine readers. And some of the most important data sources in college basketball are, from an AI agent's perspective, almost completely opaque.

We scored 103 college basketball and sports betting websites on the ARS, Agent Readability Score, measuring how well an AI agent can find, parse, understand, and extract information from each site without human assistance. Scores were collected March 17–18, 2026, using Agentiview's proprietary scoring methodology across ten dimensions covering structural, semantic, technical, and permissions layers of agent readability.

Here is what we found. Very little of it reflects well on the sports industry.

The Winner Nobody Expected

The highest-scoring site in our entire survey is not ESPN. It is not DraftKings, FanDuel, or the NCAA. It is ShotQuality.com, an advanced analytics platform that measures expected shot value in college basketball, scoring 73 out of 100.

ShotQuality is well-regarded among serious analysts. It is rarely mentioned alongside KenPom or Barttorvik in mainstream coverage. Yet it tops 103 sites, including every major media brand, every major sportsbook, and the official governing body of the sport itself.

The reason is straightforward. ShotQuality was built by developers who care about data structure. Its content is clean, its entities are clearly defined, and its architecture makes it easy for agents to navigate and extract information. Nobody optimized it for AI agents deliberately. This is simply what good content structure looks like, and it happens to be exactly what AI agents reward.

The sites built by developers who care about data structure tend to be more agent-readable than sites built by marketing teams optimizing for human engagement. The former accidentally creates machine-readable content. The latter deliberately creates the opposite.

Agentiview Research, March 2026

The Analytics Paradox: Best Data, Worst Readability

The college basketball analytics community has produced some of the most sophisticated publicly available sports data anywhere. KenPom's adjusted efficiency margins, Barttorvik's T-Rank ratings, EvanMiya's Bayesian Performance Ratings, these are genuinely valuable, cited constantly during tournament season by coaches, analysts, and serious bettors.

They are also, from an AI agent's perspective, among the least readable sites in our entire survey.

KenPom scores 44/100. When we crawled kenpom.com for an AI agent's perspective, 7 of 8 crawled pages resolved to the same registration screen, register-kenpom.php?frompage=1. Every interior navigation attempt hit a subscription wall. On a second scan run, KenPom blocked our scanner entirely, producing no score at all. Intermittent blocking may be worse than consistent blocking: an agent that succeeds today may fail tomorrow. If you are building a tournament agent that needs efficiency metrics, you need the API directly, the site itself is unreachable.

EvanMiya scores 38/100. Warren Nolan scores 39/100. The Power Rank scores 42/100. Haslametrics scores 45/100. The pattern is consistent across all four of the sport's most sophisticated analytics properties: the people who build the best basketball data tools build them for human readers, not machine readers. The data is rich. The infrastructure for serving it to agents is not.

The one exception in the analytics category is Barttorvik.com at 59/100, significantly above the survey average and the most agent-accessible option in this tier. If you are building a tournament agent that needs efficiency ratings, Barttorvik's cleaner site architecture makes it the practical choice over KenPom regardless of which methodology you prefer.

The Gambling Industry Accidentally Won

The highest-scoring category in our survey is not analytics. It is not official conference or university sources. It is the sports betting industry, and not because sportsbooks deliberately optimized for AI agents.

Regulatory compliance requirements force betting platforms to build structured, clearly attributed, machine-readable content. Explicit information hierarchies, clear entity definitions, structured odds data, these are what their regulators and users require. The result, accidentally, is precisely what AI agents reward. Covers.com, Dimers.com, and OddsShark all score in the mid-to-high 60s. MyBookie.ag scores 66. The betting analysis category consistently outperforms analytics, media, and official sources.

For developers building tournament prediction agents: use the odds aggregators for your structured data baseline. They have the most consistent, agent-readable content in the entire survey.

The Pac-12 Should Print This on Its Tombstone

The Pac-12 conference dissolved in 2024. Most of its member schools departed for the Big Ten, Big 12, and ACC. The pac-12.com website still exists, still has clean structured content, and scores 62/100, higher than DraftKings (40), tied with ESPN (62), and above the official NCAA website (57).

A dissolved conference's website is more readable to AI agents than the world's largest online sportsbook. The absurdity of this finding is its own commentary on where the industry currently stands.

DraftKings: $6 Billion. ARS: 40/100.

DraftKings is worth approximately $6 billion. It processes billions of dollars in bets annually. Its entire business model depends on data, aggregating it, displaying it, and making it actionable for users making real-time decisions. It scores 40/100 on agent readability, ranking 80th out of 103 sites.

For context: DraftKings scores lower than the dissolved Pac-12, below barttorvik.com (a personal project), below every major media brand in our survey, and below pickle.ai, an AI analytics startup that itself scores only 44/100. The company increasingly deploys AI systems for odds-setting and risk management. It has not made its public-facing content readable to the AI agents it is competing against.

This gap is not currently damaging DraftKings in any measurable way. In three years, as agent-mediated sports research becomes the default, it may be very expensive to close.

The Blue Bloods Are Invisible

Duke. Kansas. UConn. Kentucky. The most storied programs in college basketball all play on university athletics websites built on legacy CMS platforms, typically Sidearm Sports, a white-label athletics site provider used by hundreds of universities and started at Syracuse University. The platform prioritizes human experience, video content, and ticket sales. It does not prioritize machine readability.

GoDuke.com scores 55/100. Kansas's kuathletics.com scores 56/100. Kentucky's ukathletics.com scores 49/100. Michigan's mgoblue.com scores 51/100. The official websites of the programs that will define the next three weeks of sports conversation are structurally invisible to AI agents trying to research them.

An agent asked to compare tournament teams will find richer, more structured information about those teams from a gambling platform than from the programs' own websites. This is a systemic consequence of legacy CMS infrastructure and fragmented digital ownership within large institutions, not a failure of individual schools.

The Official Sources Underperform

The governing bodies and conferences that define the sport have less agent-readable infrastructure than the companies that profit from betting on it.

NCAA.com scores 57/100. The Big Ten's bigten.org scores 52/100. The SEC's secsports.com scores 52/100. The ACC's acc.com scores 54/100. Every one of these falls below the major odds aggregators in our survey. The official scorekeeping infrastructure of college basketball was not built for agent consumption, and it shows.

AP News (59) and ESPN (62) both outperform the official NCAA site. The wire service and the worldwide leader in sports built better agent infrastructure than the sport's governing body. If you are building a tournament agent that needs official data, you will get better structured information from a commercial media property than from the NCAA itself.

Polymarket vs. Kalshi: Which Prediction Market Wins for Agent Builders?

March Madness has become a significant market for prediction platforms. Both Polymarket and Kalshi are seeing record contract volume on tournament outcomes, and both are primary data sources for developers building automated trading agents.

Polymarket scores 63/100. Kalshi scores 57/100. For developers building prediction market agents, this is a directly applicable finding. Polymarket's more agent-readable infrastructure means lower crawl overhead, more reliable data extraction, and fewer parsing failures. If you are building a tournament prediction agent that monitors market prices, Polymarket's cleaner content structure is the practical advantage.

The Irony Award: pickle.ai

pickle.ai, a company with "AI" in its name that builds AI-powered basketball analytics, scores 44/100 for AI agent readability. The irony could not be more complete. A site built on AI, for AI-assisted sports analysis, is nearly invisible to the AI agents that would naturally want to consume it.

This is not an isolated case. It reflects a broader pattern: companies building AI products have not necessarily built AI-readable websites. The product and the marketing infrastructure exist in separate stacks, and agent readability optimization has not yet reached the marketing stack of even the most AI-native sports companies.

The Full Rankings: Top 20 and Bottom 15

What This Means If You're Building a Tournament Agent

The practical guidance from this data is direct. If you are building a bracket prediction agent, a Polymarket trading bot, or any automated system that needs to consume college basketball data during the tournament:

Use the odds aggregators for structured baseline data. Covers.com, Dimers.com, OddsShark, and SportsBookReview all score in the mid-60s. They have the most consistent, agent-readable content in the entire survey, better than analytics sites, better than major media, better than official sources.

Use Barttorvik over KenPom for agent-accessible analytics. Barttorvik scores 59 vs. KenPom's 44. Most of Barttorvik's useful data is accessible without a subscription wall. For efficiency metrics, T-Rank is the agent-accessible choice.

Use Polymarket over Kalshi for prediction market data. 63 vs. 57, Polymarket's cleaner content structure means fewer parsing failures for automated agents.

Do not rely on official sources for structured data. The NCAA, conferences, and university athletics sites consistently underperform commercial sources. The official scorekeeping infrastructure of college basketball was not built for agent consumption.

Why This Matters Beyond March Madness

AI agents are already consuming sports data at scale. Prediction markets are already seeing automated agent activity. Sports analytics is increasingly driven by programmatic data consumption rather than human analysis. March Madness is simply the moment when all of this becomes visible, because the search volume and agent activity surrounding the tournament compress months of ordinary agent behavior into three weeks.

The sites that invest in agent-readable infrastructure, clean HTML structure, proper schema markup, explicit content organization, and permission clarity for AI crawlers, will be discovered, cited, and consumed more frequently as AI agents become the default interface for sports research. The sites that do not will fade from the agent economy's awareness even as human traffic remains stable.

DraftKings has 12 million customers and an ARS of 40, ranking 80th out of 103 sites. Those two facts are not in conflict today. In three years, as agent-mediated sports research becomes the norm rather than the exception, the gap between agent readability and business scale may become very difficult to explain.