AI agents are already reading your site. See what they see.
Paste a URL and get a graded report: which AI crawlers you block, what an agent can actually read with JavaScript switched off, and where your structured data breaks, with the exact fixes.
Building agents? Explore the live monitoring demo →
How agents actually fail on real sites
Measured, not imagined: the failures below were detected on real public scans of 278 domains in the last 30 days. Every one looks fine in a browser, and no analytics event fires when an agent leaves.
See the domains behind these numbers →
strongest area: Performance
weakest area: Structured data
weakest area: Discoverability
Every score is a permanent public URL at agentspeed.com/score/<domain>, with a badge you can embed.
AI agents are shopping. Can they buy from you?
Shopping assistants read your store with no JavaScript, no clicks, and no patience. Three gates decide whether they can buy, and a store can pass the first two and still lose the sale at the third.
Your robots.txt lets shopping agents in, and a sitemap tells them where the products are.
Price and availability survive without JavaScript. This is where most stores fail: the price renders for a human and does not exist for an agent.
No login wall or consent banner between an agent and your checkout.
Three ways this pays for itself
A redesign moves the price into a JS-rendered component. Humans see it fine. Every AI assistant now quotes your competitor, and nothing in your analytics says why.
One malformed tool response sends an agent into a retry spiral overnight. Success rate barely moves (it retried until it succeeded) but the bill quadruples.
A prompt tweak ships. Nothing errors. Runs still return 200. But the answers get subtly worse, and you find out from a customer three weeks later.
The 24 AI user agents we track
For each bot: is your robots.txt allowing it, blocking it, or silently challenging it? Most sites disallow at least one by accident.
ChatGPT-UserOAI-SearchBotGPTBotClaude-UserClaude-SearchBotClaudeBotPerplexity-UserPerplexityBotGoogle-ExtendedApplebot-ExtendedYouBotAmazonbotBytespiderCCBotcohere-aiDiffbotMeta-ExternalAgentFacebookBotApplebotBaiduspider· CNSogou· CNPetalBot· CNYandexBot· RUYeti· KR9 drive your score · 10 training crawlers reported · 5 regional engines checked only when your site declares their markets (hreflang or lang).
Coverage tracks each operator’s public documentation. Run the per-bot access matrix →
Four things only machines see
Which AI crawlers your site allows, blocks, redirects, or challenges. Real fetches as GPTBot, ClaudeBot, Perplexity, and friends.
run this check →Is the file present, valid, and actually useful to an agent? The free analyzer grades it on six axes, separate from your scan score, which deliberately does not count llms.txt.
run this check →Most agents never run your JavaScript. See what a bot reading raw HTML actually understands about your page.
run this check →The JSON-LD machines rely on: which types you ship, what fails validation, and what is missing for your page type.
run this check →The full scan grades your site against our public rubric. Powered by Free tools →
A check is a snapshot. A watch is monitoring.
A stray robots rule or a new bot wall can cut agents off silently: nothing errors, nothing shows in analytics. Turn any check into a watch and find out the moment it changes.
Hourly, daily, or weekly. The same check, the same URL, without you remembering to look.
A robots rule edited, a bot newly challenged, a JSON-LD block dropped, a score that slid. Every snapshot is kept.
Slack, webhook, or email the moment a meaningful change lands. No alert storms; a 30-minute cooldown is built in.
Start one from any tool result: look for “Watch this URL”. Watches are included on paid plans (from $19/mo).
Building agents? Monitor those too.
Agents don’t crash. They quietly get slower, pricier, and wrong. Trace every run down to the token and the dollar, and get paged the moment something breaks.
Click a run and read it top to bottom: every LLM call, tool call, and retrieval with its own timing and tokens, plus live p50/p95 per agent.
Tokens and dollars by agent, by model, by run. One stuck retry loop can spend $40 before you finish your coffee. You will see it the same minute.
Send a quality score with each run, a number you already compute. Watch it per agent and per model, and catch the slide where success holds at 99% but answers quietly got worse.
Describe a critical flow in plain steps. On a schedule we probe the live site and verify each step is still supported by what it actually serves, and page you when one breaks. Bot walls get reported, never bypassed.
Threshold on failure rate, latency, cost, quality, or silence. Every failure keeps its error type, so a creeping RateLimitError cannot hide behind a healthy average.
Hand customers a status page with 30-day uptime bars. One link, always current.
| Status | Started | Latency | Model | Tokens | Cost | Error |
|---|---|---|---|---|---|---|
| running | just now | — | claude-opus-4-8 | — | — | |
| succeeded | 2m ago | 3.1s | claude-opus-4-8 | 1,540 | $0.021 | |
| succeeded | 4m ago | 2.7s | claude-sonnet-4-6 | 1,120 | $0.009 | |
| failed | 6m ago | 9.0s | gpt-4o | 2,310 | $0.038 | RateLimitError |
| succeeded | 8m ago | 3.4s | claude-opus-4-8 | 1,705 | $0.024 | |
| timeout | 11m ago | 30.0s | claude-opus-4-8 | 980 | $0.014 | TimeoutError |
| succeeded | 13m ago | 2.9s | claude-sonnet-4-6 | 1,260 | $0.010 |
Free for 10,000 events/month · No credit card · See how it all works →
One HTTP endpoint, from any language or framework that can send JSON. No SDK required.
The agent shipped. The hard part started.
The agent started failing overnight. You found out from a support ticket the next morning, with a screenshot you can’t reproduce.
This month cost more than last, and nobody touched the pricing. Something is retrying in a loop and there’s no quick way to see which agent.
Success rate slid from 98% to 91% after a small wording change. It shipped last Tuesday. You’re finding out now.
A 200 from the endpoint means the server is awake. It says nothing about whether the agent actually did the job.
If two of those landed, you’re who we built AgentSpeed for. Start watching →
One POST and you're live
Create a project and generate an API key. We hash it at rest, and the secret shows once.
Send one request from anything that speaks HTTP. Spans and events are optional. A new agent provisions itself the first time it reports.
Health appears right away. Set your thresholds, publish a status page, and get on with your day. No SDK to install.
curl -X POST https://agentspeed.com/api/v1/runs \\\n -H "Authorization: Bearer $AGENTSPEED_API_KEY" \\\n -d '{ "agent": "checkout-assistant", "status": "succeeded",\n "model": "claude-opus-4-8", "inputTokens": 1200,\n "outputTokens": 340, "costUsd": 0.0185, "score": 0.92 }'Basic, Deep, or Continuous: same rubric
A quick free look, a one-time deep audit, or ongoing monitoring: all three score against the same public rubric, so results are directly comparable.