Last updated: August 23, 2026
Websites detect bots in 2026 by combining four layers of signals rather than any single check: the network handshake (TLS/JA4 fingerprint and IP reputation), HTTP behavior (header order and HTTP/2 frames), browser fingerprints (canvas, WebGL, fonts, navigator.webdriver), and behavioral biometrics (mouse and keystroke patterns). The single strongest signal is inconsistency between those layers, such as a client that claims to be Chrome over a handshake that is not. Proxies like Thunderproxy operate only on the network and IP layer; they do not change your browser fingerprint or behavior.
The short answer: how websites detect bots in 2026
Modern bot detection is a scoring system, not a single test. A detection platform gathers dozens of signals across four layers, weighs them, and assigns a confidence score to each visitor. No single signal decides the outcome, which is why patching one giveaway rarely changes the result.
The layer that matters most is consistency. A real Chrome user produces a matching story at every layer: the TLS handshake, the HTTP headers, the browser APIs, and the way the mouse moves all agree that this is Chrome on a real device driven by a person. When those layers disagree, the mismatch itself is the signal. Detection engines are built to notice that a visitor claiming one thing at the HTTP layer looks like something else at the network layer.
The bot-detection stack: four layers
Detection fires in sequence, starting before your first byte of HTTP and continuing through every interaction. The table below shows each layer, example signals, when it runs, and whether a proxy affects it. Read the last column carefully: most of the stack is untouched by proxies.
| Layer | Example signals | When it fires | Can a proxy change it? |
|---|---|---|---|
| Network / TLS | JA4 fingerprint, IP reputation, HTTP/2 SETTINGS frame | At connection | IP only, not the TLS fingerprint |
| HTTP | Header order, Client Hints, User-Agent consistency | On each request | Indirectly, via IP and location |
| Browser | Canvas, WebGL, fonts, navigator.webdriver | On page load | No |
| Behavioral | Mouse acceleration, keystroke cadence, scroll physics | During interaction | No |
The practical takeaway: a clean IP addresses one layer of four. It is necessary for many legitimate jobs, but it is not what makes automated traffic look human.
The network layer: IP reputation and TLS/JA4 fingerprinting
The first thing a website judges is the connection itself, before any page loads. Two signals dominate here.
IP reputation asks whether the address has a history. Datacenter ranges, addresses tied to prior abuse, and IPs shared by many unrelated users all carry weaker reputations than a residential or ISP address that looks like a normal home connection. This is the one layer a proxy genuinely changes.
TLS fingerprinting reads the handshake. JA3 and JA4 are fingerprints derived from the TLS ClientHello message that a client sends at the start of every HTTPS connection, before HTTP headers and before encryption is fully established. JA3 hashed the TLS version, cipher suites, extensions, and elliptic curves into an MD5 value, but Chrome 110 introduced extension-order randomization in January 2023, which weakened JA3. JA4 fixes this by sorting ciphers before hashing and by including the ALPN value, so it can tell HTTP/1.1 and HTTP/2 apart. By 2026 the move from JA3 to JA4 is essentially complete across major anti-bot platforms.
The important point for anyone running legitimate automation: the TLS fingerprint is produced by your client, not your proxy. A proxy forwards the connection; it does not rewrite the handshake. So if an HTTP library presents a default OpenSSL handshake while its headers claim to be Chrome, the mismatch stands regardless of which IP it comes from.
The browser layer: fingerprinting and HTTP/2
Once a page loads, JavaScript collects a device fingerprint. A modern fingerprint combines 15 or more signals, including a canvas hash, the WebGL renderer string, screen resolution, timezone, installed fonts, and navigator properties, into an identifier that tends to persist across sessions. Each signal is weak alone, but together they describe a device with enough specificity to recognize it again.
Alongside the browser APIs, the HTTP layer carries its own fingerprint. The order of HTTP headers, the set of Client Hints, and the HTTP/2 SETTINGS frame all vary by client, and detection engines check that they match the browser the visitor claims to be. A request whose header order does not match any real browser build is a straightforward signal.
None of this layer is affected by the network path. Changing your IP does not change your canvas hash, your font list, or your header order.
Headless and automation signals
Automated browsers leave traces that fall into roughly four tiers, from easy to hard to address.
The easiest is navigator.webdriver, which headless Chrome sets to true by default. This one is trivial to patch, so on its own it now catches only unconfigured tools. Harder tiers include GPU and rendering fingerprints, transport-level TLS and HTTP/2 signatures, and the deepest tier, behavioral motion.
A durable modern signal is the Chrome DevTools Protocol (CDP) artifact. When an automation framework drives a browser, it leaves side effects that are inherent to how CDP works: modified error stack traces, non-native property descriptors, subtle timing differences, and changes to console API behavior. These are difficult to remove because they are a byproduct of the control mechanism itself, not a setting.
There is a second-order effect worth knowing. Tooling that patches many signals tends to create its own recognizable fingerprint, because the specific combination of patches is itself distinctive. Adding more countermeasures can make an environment easier to classify, not harder.
The behavioral layer: mouse and keystroke dynamics
The final layer watches how a visitor acts. Human cursor movement has momentum: you accelerate toward a target and decelerate as you arrive, tracing a curved, slightly noisy path. Synthetic input tends to jump instantly or move at uniform speed. Keystroke dynamics measure the dwell time on each key and the flight time between keys, a rhythm that varies from person to person.
Behavioral analysis is powerful because it is expensive to imitate. Peer-reviewed research on behavioral biometrics reports classification accuracy above 99% when distinguishing human from automated input using mouse trajectories or keystroke timing alone (Behavioral Biometrics: Keystroke Dynamics vs Mouse Trajectories for Bot Detection). This is the layer that no proxy, and no simple browser patch, addresses.
Where proxies fit, and where they do not
A proxy changes one thing: the network path and the address a request comes from. Used for legitimate work, that is valuable. A clean residential or static ISP IP with a stable location helps genuine traffic present a consistent, well-reputed network footprint, which is exactly what you want when the account or session is entitled to exist and you simply do not want a shared or low-trust address dragging it down. This is the network layer of the stack above, and it is the layer proxies own.
What a proxy does not do is equally important to state plainly. It does not alter your TLS/JA4 fingerprint, your browser fingerprint, your header order, or your behavior. Those are produced by your client and your interaction, above the network path. Anyone who tells you an IP alone makes automated traffic indistinguishable from a human is describing one layer of four.
For legitimate use cases like QA, geo-testing, ad verification, and price monitoring, the correct approach is to match the tool to the layer: a clean IP for the network footprint, a real browser environment for the fingerprint layer, and realistic pacing for the behavioral layer. For the network part, see the static ISP plans for persistent identities and rotating residential for geo coverage, and the rotating residential docs for how to hold or rotate an address.
FAQ
How do websites detect bots in 2026?
Websites score visitors across four layers: the TLS/JA4 handshake and IP reputation, HTTP header and HTTP/2 patterns, browser fingerprints like canvas and WebGL, and behavioral biometrics such as mouse and keystroke timing. No single check decides it. The strongest signal is inconsistency between layers, like HTTP claiming Chrome over a non-Chrome handshake.
What is TLS fingerprinting with JA3 and JA4?
JA3 and JA4 are fingerprints built from the TLS ClientHello a client sends at the start of an HTTPS connection. JA3 used an MD5 hash weakened by Chrome’s 2023 extension-order randomization. JA4 sorts ciphers and adds the ALPN value, and by 2026 it is the standard across major anti-bot platforms.
Can a proxy hide that I am using automation?
No. A proxy changes the IP address and network path, which is one layer of detection. It does not change your TLS fingerprint, browser fingerprint, HTTP header order, or behavior, and it cannot make automated traffic indistinguishable from a person. Treat a proxy as network hygiene, not a detection solution.
Does navigator.webdriver still detect automation?
Only partly. Headless Chrome sets navigator.webdriver to true by default, so on its own it now catches mainly unconfigured tools. Detection in 2026 leans on harder signals: Chrome DevTools Protocol artifacts, GPU and rendering fingerprints, transport-level TLS patterns, and behavioral motion, which are far more difficult to alter.
What is the most reliable bot-detection signal?
Cross-layer inconsistency is the most reliable signal. A visitor whose HTTP layer claims one browser while the TLS handshake, header order, or behavior says otherwise gets a high automation score. Behavioral biometrics and CDP artifacts are also strong because they are expensive to imitate and hard to remove.
Do proxies change my browser fingerprint?
No. Your browser fingerprint, including canvas hash, WebGL renderer, fonts, and navigator properties, is generated by the browser on the page, above the network layer. A proxy only changes the address the request travels through. Changing IP does not change your canvas hash, your font list, or your header order.
Start with the 256MB free trial, no card required, and use a clean, stable IP as the network layer of a legitimate setup, not as a substitute for the other three.