Most AI bot detection still works like a guest list, where you find the AI crawlers by name and decide whether to let them in or keep them out. That made sense when the only AI traffic on your site was a crawler collecting training data, but that's no longer the case.
Some of the AI traffic reaching you now is an agent shopping, comparing or logging in for one of your customers. If you block it, you turn that customer away, and if you allow it by name, you let in anyone who can pretend to be it.
AI bot detection now has to tell you what each request really is, so you can decide what it gets to read and do.
AI traffic isn't one thing

As we wrote in AI, agents, and the trust gap, one AI provider can show up on your site in three ways.
Crawlers collect content, prices and inventory without a person involved. In Q3 2026, KasadaIQ saw AI crawlers send more than 33.7 million requests across monitored sites in a single month, 86% of them from Meta and ByteSpider (Q3 2026 threat report).
Assistants browse for a person who is researching or comparing products.
Agents go a step further and act for someone, signing up, checking out or booking on their behalf.
Each one carries a different risk, and treating them all as one "AI bot" means you either block real customers or let abuse through.
How to detect AI bots
Start with what the request can prove, because some signals are much harder to fake than others.
In order of strength:
- Cryptographic signatures like Web Bot Auth, which let an agent sign what it sends so you know who sent it
- IP ranges published by the vendor, like the ones Google publishes for its crawlers
- A fingerprint built from signals collected from the client, which catches agents that don't identify themselves but has to be done carefully, because a fingerprint botters can reverse engineer becomes an entry point
User agent strings and other ways a request names itself aren't used, because they're easy to spoof.
Meta's Muse shows why that order matters. As our CMO wrote in Muse isn't the user. So who is?, most of its traffic arrives unsigned, and it also comes from several providers and IP ranges, so neither a signature nor an IP list can pin it down. What's left is building a signature distinct enough to tell it apart without false positives or false negatives, which is the same work as regular bot detection. And if the signals behind that signature can be collected and copied, there's no reliable way to tell Muse from a bot pretending to be it.
It's also worth knowing that your marketing analytics may undercount AI traffic, so check your bot management or edge data before you decide anything.
Detection is only half the job
A signed request tells you who sent it, but not what that agent should be allowed to do. As we wrote earlier this year, identity without authorization is just better labeling, and closing that gap is the job of AI agent security.
Two other shortcuts don't hold up either. An agent's own instructions aren't a security boundary, because prompts can be overridden. You also can't count on a challenge, since an agent might solve your CAPTCHA, fail it, or hand it back to the real user halfway through checkout.
How to decide what each AI agent can do
This is where AI bot detection turns into AI bot mitigation, and it's easiest to work through one customer journey at a time. Our guide to AI bot traffic walks through the full policy check.
Start by finding out which agents reach it, how often, and on which pages.
Then weigh the proof, because a signed request is the strongest evidence and a published IP range comes next, while a user agent proves nothing, so give agents with weaker proof less room.
Decide by action, not just by agent. Browsing, search and product pages are one conversation, and login, signup and checkout are a different one. Read-only access is often the right middle ground.
Assume anything you allow will be impersonated, and make sure you can tell the real agent from a copy.
And ask Marketing before you block, because Forrester's recent research found that blanket-blocking buyer-assist agents can kill deals before sales ever sees them. Once you've decided, name an owner and a review date, since agent behavior keeps changing.
How Kasada handles AI agents
Kasada's AI agent trust management keeps a directory of verified bots and AI agents and classifies each one by how strongly it can prove who it is, from Web Bot Auth signatures down to published IP ranges. You set permissions for each agent by HTTP method, from read-only access to specific actions or a full block. Rules run at the edge, before traffic reaches your application.
Kasada has never used CAPTCHAs, and for agents that don't sign their requests or come from a published IP range, it builds a signature from signals collected from the client, so detection doesn't depend on the name a request claims or on agents telling the truth.
When the person behind an agent logs in or signs up, Account Intelligence scores that account on the same request. The agent tells you how someone arrived, and the account tells you who they are.
Dashboards show crawling bots and agentic bots separately, and you can export the logs to your data warehouse. Forrester named Kasada a Leader in The Forrester Wave™ for Bot And Agent Trust Management, Q2 2026.
See which AI agents reach your site. Book a demo.
Rules for agents, not a gate
Think of an AI agent as someone running errands for a friend. You'd happily let them pick up a coffee order, but you'd want more than their word before handing over your friend's bank card.
AI bot detection and mitigation work the same way. Let agents read where it helps your customers find you, and ask for more where money, accounts or data change hands.
More agents are coming, and fewer of them will introduce themselves.

