Vercel's is-agentic.com: agent readability gets a score. Enter a URL, get a grade on how well AI agents can read your site. SmarToolbox ran it on ours, and learned things SmarToolbox didn't want to know. On August 21, the Vercel Developers account posted what looked like a routine product drop. It wasn't one. is-agentic.com is a free tool that measures how well AI agents can read your website: type in a domain, get a score out of 100, see every check that failed, and copy a fix prompt straight into your coding agent. It landed quietly. It shouldn't have. This is the first time agent-readability gets a public scoreboard with serious infrastructure behind it, and the scores people are getting back are humbling. Including mine. What the number actually measures. is-agentic.com is a partnership. Vercel provides the front door; Ora provides the engine. Ora has been running a larger research ranker for months: 127 checks across four layers covering discovery, access, usability, and payments, against more than 16,000 domains. Most websites don't transact with agents and have no API surface, so the two companies carved out a version tuned for normal sites. The scoring model is public on the methodology page. Essential checks share an 80-point pool. Recommended checks share 20. Bonus signals can add up to five points and never subtract. A relevance pass excludes checks that don't apply, so a blog never gets marked down for lacking checkout rails. Scans run live against your server, and completed reports cache for roughly six hours. Look at what sits in the essentials: a real HTTP 404 status for missing pages, server-rendered headings an agent can parse without JavaScript, machine-readable trust pages, structured organization data. None of this is exotic protocol work. It's the plumbing a competent site should already have, finally checked by something that notices. Agents don't bounce off a confusing site. They quietly decide you're broken and go recommend someone else. That asymmetry is what makes this category different from anything before it, and it's worth sitting with for a second. A human who hits a confusing page gets frustrated, maybe closes the tab. An agent does neither. It reclassifies your entire domain as low-confidence and moves on. No bounce event, no exit survey, no angry support ticket. The failure never registers anywhere you can see it. Even Vercel flunks parts of its own test. When I pulled the live report for vercel.com, it scored 85 out of 100. Strong, top-percentile, and still dropping essential points for the least glamorous reason possible: nonexistent paths return HTTP 200 with the app shell. The classic soft-404. An agent probing for resources concludes that every path on the domain exists, which poisons everything downstream. There's a lesson in the fact that the platform company itself hasn't shipped that fix. These failures survive because they're invisible. Humans never hit a soft-404 by accident, and agents don't file bug reports. Nothing in your analytics dashboard flags the moment a crawler misreads your site and gives up. The damage happens entirely off the books. I ran it on smartoolbox.com. Its own site scored 76 out of 100, which the tool labels "Ready with a few material gaps." Material is the right word. The report found that smartoolbox.com doesn't do markdown content negotiation: request the same URL with Accept: text/markdown and you get HTML back, so agents burn tokens parsing navigation and cookie banners instead of reading clean text. There's no when-to-use guidance telling agents which jobs the site is right for. The Organization schema is missing contactPoint and address. There's no Contact page for a machine to verify. Eight findings in total. Each one ships with a fix written as a prompt you can paste into Claude Code or Cursor, plus the curl command to verify it worked. Total research time: about a minute. Doing this manually would have meant weeks of guessing what agents struggle with, based on nothing. That's the actual product here. Not the number, the specificity. "Your site is invisible to machines" is anxiety. "Return a real 404 status code; here's the curl command to prove it" is a Tuesday afternoon task. Why this landed now. The timing isn't accidental. Cloudflare reported earlier this year that fewer than half of all HTML page requests come from human beings. HUMAN Security's 2026 benchmark measured AI agent traffic growing 7,851% year over year. A May 2026 domain audit cited by Similarweb put agent-ready websites under 5% of the web. Meanwhile Chrome shipped an Agentic Browsing category in Lighthouse, and Google and Microsoft pushed WebMCP through a W3C draft toward a Chrome origin trial. Every layer of the stack converged on the same question at the same time: can a machine find, read, understand, and act on what you built? For twenty years the answer didn't matter much. Now there's a number attached to it, and the number is low almost everywhere. My take: this is not SEO reborn. The lazy comparison is "SEO for robots." It's wrong in a useful way. SEO rewarded attention. Rankings, impressions, clicks: all measurable, all exploitable, because a ranking function responds to signals you can manufacture. Agent-readability rewards evaluation clarity: being correctly understood and correctly acted upon. When an agent skips your site, nothing logs it. There is no missed-impressions report. The recommendation simply goes to a competitor whose pricing and specs were easier to extract, and you never find out. You cannot stuff evaluation clarity. Either the machine can pull your prices, your constraints, and your next action, or it can't, and no amount of keyword seasoning changes that. The incentive gradient finally points toward being genuinely legible, which might be the healthiest thing to happen to web marketing in a decade. Run the audit before your competitors do. Most sites sit barely above half, so one deliberate afternoon of essential fixes puts you ahead of nearly everybody. What to watch, and what to do this week. Three things deserve your attention. First, run the audit yourself: npx is-agentic yourdomain.com, or the web form. Free, no signup, about sixty seconds. Fix essentials before touching anything in the recommended tier, and refuse to fake surfaces you don't run; a stubbed MCP server fools the scanner and nothing else. Second, expect these checks to harden into deploy gates. Adjacent tooling already exits nonzero in CI when a readiness score drops below a threshold, and it takes little imagination to see hosting platforms wiring an agent-readiness check into deployment previews. Third, watch whether the score becomes table stakes the way mobile-friendliness did around 2015: ignored, then mocked, then suddenly mandatory. The window matters more than the number. Under 5% of sites have done this work, which means the ones moving now get evaluated, cited, and recommended while everyone else waits for the trend to feel safe. The bar is low right now precisely because almost nobody has tried to jump it.