Bot or Human? Wrong Question. Detection Starts With Intent

Image Source: unsplash.com

It's time to stop asking bot or human. The question should be what the intent is, as much as who's doing what and where it's coming from. Any type of traffic can technically be a risk. And as we're seeing in 2026 and likely well beyond this year, bot attacks are becoming more sophisticated, and bot detection is becoming more difficult. Read on to find out more.

Why Bot Detection Should Start With Intent

The old binary model just isn't working anymore. Until recently, the assumption was that humans are legitimate and bots are malicious. In some cases, when it's true, obviously that works. But it's all too common that a human can initiate harmful activity through an AI agent. An automated agent can also perform a completely legitimate task for a genuine customer.

Now you can see the predicament we're in. Well, the predicament is that the old binary model isn't as effective as the new way of thinking, which is the question: "What is this request actually trying to do?"

DataDome champions this, believing that bot traffic detection means understanding that traffic type alone doesn't determine risk.

Identity still matters, but it is only the first layer:

• Is this actually the AI agent it claims to be?
• Is it coming from expected infrastructure?
• Has its identity been spoofed?

Even a strongly authenticated, legitimate AI agent can be used maliciously, so confirming identity does not confirm intent. DataDome offers a strong example. Its researchers observed attackers causing legitimate AI infrastructure to send malicious requests. Cases included:

• Perplexity infrastructure being used to probe an e-commerce site with an XSS payload.
• Meta's genuine crawler being triggered to probe a tourism website.
• A request from documented OpenAI infrastructure containing an SQL injection attempt.

Intent detection analyzes everything from behavioral sequencing to network and device signals. Malicious bots, for example, run request rates and volumes no human could match.

And one request often tells you very little. A sequence of requests can reveal a purpose. For example, visiting /login once is completely normal. Repeatedly testing credentials across accounts is not. The endpoint is the same; the behavior and intent are different.

With that in mind, identity still matters. But now, identity + intent matters more.

How AI Agents Act on Behalf of Real Users

AI traffic varies and can serve very different purposes. Two of the most common now are user-prompted agents that operate live on behalf of a person, such as ChatGPT-User and Perplexity-User AI agents. Other AI traffic includes:

• Search and answer bots fetching information to answer queries.
• Crawlers and indexers gathering content for indexing.
• Crawlers for search or training.

So a live AI assistant can increasingly behave like an actual website visitor rather than a traditional crawler; the best example is ChatGPT-User. You ask ChatGPT a question, and ChatGPT-User sends crawlers out to go through websites and give you the answer.

Between July 2025 and June 2026, DataDome recorded 52.7 billion AI agent and LLM crawler requests, with AI traffic growing 82.3%. And the latest Q3 2026 search shows the trend accelerating. DataDome research found almost 25 billion AI agent requests in Q3 alone, and agentic-browser traffic jumped 4.7x in September.

Agentic browsers are more important now because, unlike basic crawlers fetching a page, agentic browsers can render JavaScript and interact with websites through full browser sessions, making them harder to distinguish from ordinary users. That's why it's important to think back to how to detect bots, such as behavior sequencing and network signals.

And interestingly, AI agents are now moving deeper into websites. During H1 2026, DataDome recorded 605.6 million AI requests to high-risk endpoints, which included login pages, forms, payment flows, carts, and account-creation pages.

But to add to the complexity, a legitimate agent can access an account to check an order or loyalty balance for an authorized user. So you see, the back-and-forth we're doing between actions that can be both legitimate and malicious is exactly why intent is so important for detection.

How Bad Bots Mimic Human Traffic

The opposite problem we're seeing is that automation can deliberately try to look human. DataDome found that 29.5% of sampled bad bot traffic was average-complexity automation, often using frameworks including Playwright and Chrome DevTools Protocol, which can simulate legitimate browser environments more convincingly than basic automation.

Some traffic is also advanced enough to interact with active challenges and attempt to evade anti-detection measures. That creates the issue that browser appearance alone isn't proof that there is a human behind the session.

Going back to intent analysis methods, request speed alone might suggest human behavior, whereas the full sequence can reveal automated account creation.

Bad actors can also impersonate trusted AI identities. DataDome found that AI-agent spoofing increased 45% between February and July 2026, and that 80% of AI agents do not properly identify themselves. That creates additional difficulty in separating legitimate crawlers, data collectors, and attackers pretending to be trusted agents.

The biggest issue is that good automation increasingly looks like human activity. Malicious automation deliberately tries to do the same. Appearance is therefore becoming a weaker indicator of intent.

Using Intent to Determine Good and Bad Bot Traffic

Bringing the previous sections together, you can see why we're beyond asking, "bot or human."

Detection is a holistic approach that looks at the resource accessed, whether the behavior matches the visitor's claimed purpose, what happened before and after the request, and what the visitor is actually doing. Did it interact with sensitive endpoints? Or did the speed or volume seem unreasonable for the task?

A great intent-based approach combines server-side and client-side signals. Server-side signals include request headers and traffic patterns, while client-side signals include browser and device telemetry.

The key is then to go beyond individual requests and use session-level and user-level behavioral analysis. This can expose patterns that appear only over time, including coordinated probing and gradual data exfiltration.

With that in mind, it's easy to see how the web is moving toward a point where neither detecting a human nor a bot tells you what's truly happening. You can't tell whether to trust a visitor. Intent is discovering the meaning behind the actions, whether it's a bot or not.