Search the guides

Traffic fundamentals

Is this request a person or a bot?

No single browser or network signal answers that question reliably. Learn to combine evidence and choose a proportionate response.

Start with what you can actually observe

Many similar requests arrive just after your ad starts. They might be interested visitors, automated checks, or unwanted activity. An address or browser name alone cannot tell you which. Look first at what the requests do and whether that behavior causes a problem.

A request asks the server for something: a page, an image, or a form submission. Headers are information sent with that request. One header, User-Agent, describes the client. It is a claim from the sender, not proof that a human is using the page.

Decide what you need to protect. Reading a public guide is different from submitting the same form repeatedly. The evidence needed for a complete rejection should reflect the harm you are trying to prevent and the cost of rejecting a real person.

Use layers of evidence

Network information helps locate a request, but does not explain its purpose. An office can share one outward-facing address. A hosting network can run an approved monitor. Automation can also use a residential connection.

Next look at behavior: which pages or actions repeat, in what order, and over what period? A refresh may be a customer trying to recover from an error. A burst may follow a successful advertisement. The same pattern can have more than one explanation.

Check the effect on the service. Repeated rejected forms, attempts to reach administrative pages, and an overloaded resource give you more specific problems to solve than a low-quality label. Keep enough evidence to reproduce the issue without retaining unrelated visitor details.

A decision should become stronger as evidence accumulates

Combine observations rather than turning one unusual feature into a verdict. A network label becomes more useful when you can connect it to repeated behavior and a harmful result.

The layers below support an assessment, not certainty about a person's intentions. When evidence is incomplete, keep that uncertainty visible and choose a response suited to the affected action.

Signal layersCombine clues before restricting access
  • Network and client

    Origin and the client's own description.

  • Behavior

    Which actions repeat.

  • Impact

    Overload or unwanted records.

An unusual network plus repeated harmful actions is more informative than the network label alone.

Recognize authorized automation separately

Some automation is useful: search crawlers, advertising checks, and uptime monitors need to request pages. List the services your site depends on and use each service's documented identification method.

For Google, official verification includes published address ranges and a DNS procedure. DNS connects hostnames with network addresses. The manual procedure finds the name for an address and then checks that the name resolves back to that address. Use the guidance for the relevant crawler category.

Verified identity can justify a narrow access exception. It does not justify showing that service a different offer. Retest exceptions when the service's published infrastructure or your own access rules change.

Match the response to the action

A public page should not become inaccessible merely because a weak signal looks unfamiliar. If the observed problem is repeated costly submissions, a limit on that action may address it more directly than a site-wide block. An extra challenge also adds friction, so test it with legitimate users.

In a hypothetical example, several people read a product page through an office network, while one source repeatedly submits the form. Protecting submission targets the problem without treating every reader on the network as hostile.

Make recovery understandable. An error explaining what happened, a safe retry, and a request reference can help support find the decision. A blank screen gives neither the visitor nor the operator a useful next step.

Decision pathChoose a proportionate response
What do you observe?

Connect the signal to an action.

  • Ordinary reading

    Preserve page access.

  • Insufficient evidence

    Observe and investigate.

  • Repeated harmful action

    Limit the affected endpoint.

Target the observed risk rather than guessing every reader's identity.

Separate classification from enforcement

Separate the assessment from its consequence. You can flag an unusual request for review while allowing the public page. That lets you study a new signal before it rejects paid visitors.

Keep the reason and action separately visible: “unusual network, page allowed, form rate limited” is more informative than “bot blocked.” It explains why someone could read the offer but could not submit.

Use clear states, including insufficient evidence. Reading an article, recovering an account, and paying for an order have different consequences if access fails. One threshold for every action may be convenient to configure but poorly matched to those consequences.

Work through an ambiguous burst

Imagine a sudden burst after a campaign launch. A rate rule flags it. Before calling it an attack, compare the timing with the advertisement and inspect the resources being requested.

Many independent visits requesting the page and its normal assets could reflect genuine interest. Repeated requests to one expensive action, accompanied by service disruption, identify a more specific concern. You can protect that action without claiming to know who is behind every request.

After the restriction, check whether the service recovered, ordinary customers could finish, and the unwanted behavior moved elsewhere. Those questions measure protection. A suspected-bot total only measures how often the system applied its label.

A visit without a sale is not evidence of a bot. A person may simply decide against the offer. Likewise, a purchase does not replace separate operational or payment checks.

Build a small review practice

Regularly review a manageable sample from important decision reasons. Include allowed and rejected requests. Look for reproducible patterns and mark cases where the available record cannot establish intent.

Use your own tests as reference points, not as a model of the whole audience. Your account or office network may have exceptions. Include a fresh visitor and a legitimate shared-network or restricted-browser scenario.

Revisit rules after audience and infrastructure changes. A new region or app may introduce ordinary behavior that used to be rare. Record what evidence would make you change the rule, who owns it, and which diagnostic records are necessary.

A short review is useful when it produces a clearer rule, confirms a benefit, or identifies missing evidence. A large report that never changes a decision does little for the next blocked customer.

Review decisions against real outcomes

Measure decisions against the problem you meant to solve. Did harmful submissions fall? Can a genuine customer still use the route? Are support complaints repeating the same rejection reason?

No header proves a person's intentions. A workable process identifies known services, responds to observed harm, and leaves a way to correct mistakes. Keep that process more visible than the confidence implied by a bot score.

Questions answered

Can JavaScript execution prove that a visitor is human?

No. Some automated browsers run page scripts, while legitimate users may restrict them. Treat script execution as one observation among others.

ADSTRAFFIC / ENAll guides →