April 15, 2026 · Report

We measured 145 bots. 27% were hostile. 1 site is already enforcing.

Over two weeks we observed every bot that touched botconduct.org and scored each one with our proprietary behavioral engine. The data is unflattering for most — and it's why ImportSignals just became the first production site integrating the Bot Conduct Standard API.

145Distinct bots
40Hostile
83Acceptable
33Exemplary

What "hostile" actually means

We don't judge intent. We score observed behavior against a proprietary rubric that spans multiple behavioral dimensions and is refined continuously as new bot behaviors emerge.

Bots with low scores land in the hostile category: the site would be well within its rights to block them. Bots at the top earn acceptable or exemplary ratings.

The rubric and weights are not public. What is public is the verdict for any given bot, which is what our API returns to sites that integrate BCS.

The exemplary list: bots that do it right

Twelve named operators hit a perfect or near-perfect score by behaving the way a standard expects:

BotOperatorScore
GPTBotOpenAI100
ClaudeBotAnthropic100
BingbotMicrosoft100
BytespiderByteDance100
BaiduspiderBaidu100
YandexBotYandex100
FacebookMeta100
redditbotReddit100
AwarioSmartBotAwario100
toolhub-bottoolhub24.ru100
GooglebotGoogle92
AhrefsBotAhrefs90

These are the reference implementations. If your bot can't match what Googlebot does, that's a product problem.

The hostile list: the bottom of the distribution

On the other extreme, we found 40 bots landing in the hostile category. Most fall into three archetypes:

1. Security scanners not declaring themselves

L9Explore (LeakIX) scored 0 with 118 requests, hammering dozens of security-relevant paths aggressively in a short window. LeakIX is a legitimate security research company, but this specific scanner makes no attempt to identify itself as research rather than attack.

2. Stealth bots at scale

We logged 60+ distributed crawlers exhibiting hostile behavioral patterns consistent with scanning and credential probing. One bot from a single cloud IP sent 2,562 requests in a single day.

3. Abusive probing

A set of operators targeted endpoints that legitimate crawlers have no reason to visit — the kind of paths attackers hit when looking for unsecured infrastructure. Any bot probing those is, by definition, not a legitimate crawler. 12 distinct operators triggered this category over two weeks.

"The hostile bots aren't 'AI agents gone wrong.' They're scanners. And the line between 'legitimate scraper' and 'scanner' is exactly what this standard measures."

Why ImportSignals flipped the switch today

ImportSignals — an e-commerce intelligence platform for importers — gets hit by hundreds of automated requests a day. Most are harmless; some are not. As of today, every request to api.importsignals.com/api/* is classified in real time against the BCS registry:

Today the middleware runs in observe mode: we're measuring, not blocking. In one week we move to warn mode. In two weeks, enforce mode. Operators have plenty of time to certify their bots before anything breaks — if they even use them.

Running a bot, scraper, or AI agent?

Level 1 takes 30 seconds. Or run the full behavioral test and get a real score.

What comes next

We're looking for three things in the next 30 days:

The web needs a shared vocabulary for "this bot is polite" vs "this bot is trying to break in." BCS is one attempt. We'd rather be proven wrong than be ignored.

archive internal archive internal