AI Detections

WebDecoy provides comprehensive detection of AI-related activity on your website, including server-side AI crawlers, client-side browser extensions, LLM referral tracking, and suspicious signals that indicate AI-assisted browsing.

For the crawl-to-referral model and reporting capabilities, see the AI Traffic product overview.

Overview

AI activity falls into six main categories:

Category	Detection Method	Needs a server-side sensor?	Examples
AI Scrapers	User-Agent analysis	Yes	GPTBot, ClaudeBot, CCBot
AI Browser Plugins	DOM/JavaScript analysis	No	Claude extension, ChatGPT sidebar
Vision AI Bots	Form behavior analysis	No	Claude Computer Use, Stagehand, Browser Use
LLM Referrals	Referrer header analysis	No	Visitors clicking links shared by ChatGPT, Perplexity, Claude
AI Tools	Behavioral signals	No	Automated form filling, content extraction
Suspicious Signals	API monitoring	No	Wrapped fetch, modified observers

┌─────────────────────────────────────────────────────────────────┐
│                    AI DETECTION LAYERS                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│   SERVER-SIDE                    CLIENT-SIDE                    │
│   ───────────                    ───────────                    │
│                                                                  │
│   ┌─────────────────┐            ┌─────────────────┐           │
│   │  AI Scrapers    │            │  Browser Plugins │           │
│   │  (User-Agent)   │            │  (DOM Detection) │           │
│   └────────┬────────┘            └────────┬────────┘           │
│            │                              │                     │
│            │                     ┌────────┴────────┐           │
│            │                     │  Vision AI Bots │           │
│            │                     │  (Form Behavior)│           │
│            │                     └────────┬────────┘           │
│            ▼                              ▼                     │
│   ┌─────────────────┐            ┌─────────────────┐           │
│   │  AI Scraper     │            │  AI Plugin +    │           │
│   │  Score (0-100)  │            │  Vision Flags   │           │
│   └────────┬────────┘            └────────┬────────┘           │
│            │                              │                     │
│            └──────────────┬───────────────┘                     │
│                           ▼                                     │
│                  ┌─────────────────┐                           │
│                  │  Detection      │                           │
│                  │  Record         │                           │
│                  └─────────────────┘                           │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

AI Scrapers (Server-Side)

AI scrapers are bots that crawl websites to collect content for training AI models. WebDecoy detects these through User-Agent analysis and assigns an AI Scraper Score.

The score in these tables is the agent’s base score in WebDecoy’s agent registry (186 known agents; browse them all in Bot Directory). A matched user agent gets that score as its AI scraper score.

Training Crawlers (High Score)

These crawlers collect content for AI/ML model training:

Crawler	Company	Score	User-Agent Pattern	Purpose
GPTBot	OpenAI	85	`GPTBot`	Training GPT models
ClaudeBot	Anthropic	85	`ClaudeBot`	Training Claude models
Anthropic	Anthropic	85	`anthropic-ai`	Anthropic’s general crawler
CCBot	Common Crawl	80	`CCBot`	Open dataset used for AI training
Google-Extended	Google	80	`Google-Extended`	Gemini AI training
Cohere	Cohere	80	`cohere-ai`	Cohere model training
ByteSpider	ByteDance	75	`Bytespider`	TikTok and AI training
Meta-ExternalAgent	Meta	75	`Meta-ExternalAgent`	Meta AI training
Applebot-Extended	Apple	75	`Applebot-Extended`	Apple Intelligence training
YouBot	You.com	75	`YouBot`	You.com AI search training
Amazonbot	Amazon	70	`Amazonbot`	Alexa and Amazon AI
FacebookBot	Meta	70	`FacebookBot`	Meta AI training

AI Search Crawlers, Assistants and Commercial Scrapers

Crawler	Company	Category	Score	User-Agent Pattern
OAI-SearchBot	OpenAI	AI search crawler	80	`OAI-SearchBot`
PerplexityBot	Perplexity	AI search crawler	80	`PerplexityBot`
ChatGPT-User	OpenAI	AI assistant (user-triggered fetch)	85	`ChatGPT-User`
Diffbot	Diffbot	Commercial scraper	70	`Diffbot`

Search Crawlers (Lower Score)

Traditional search engines that may also feed AI features:

Crawler	Company	Score	User-Agent Pattern	Note
Googlebot	Google	30	`Googlebot`	Primary search indexing
Bingbot	Microsoft	30	`bingbot`	Bing search + Copilot
DuckDuckBot	DuckDuckGo	30	`DuckDuckBot`	Privacy-focused search
Applebot	Apple	30	`Applebot`	Siri and Spotlight
YandexBot	Yandex	35	`YandexBot`	Russian search + AI
Baiduspider	Baidu	40	`Baiduspider`	Chinese search + AI

SEO and Analytics Crawlers

Crawler	Company	Score	User-Agent Pattern
SemrushBot	Semrush	50	`SemrushBot`
AhrefsBot	Ahrefs	50	`AhrefsBot`
MJ12bot	Majestic	50	`MJ12bot`
DotBot	Moz	50	`DotBot`

HTTP Client Libraries and CLI Tools

Matched by User-Agent like any other agent:

Library	Pattern	Score
Python Requests	`python-requests`	55
Go HTTP Client	`Go-http-client`	55
Java HTTP Client	`Apache-HttpClient`	55
Scrapy	`Scrapy`	65
cURL	`curl/`	45
Wget	`Wget/`	45

Where crawler activity is reported

Individual crawler hits land in the Detections table like anything else, but the report built for them is AI Crawlers & Agents (AI Crawlers in the sidebar, at /verified-agents). It answers a question a score alone cannot: whether the crawler was who it said it was.

Requests that claimed a known agent are split three ways, at the request level:

Bucket	Meaning
Proven	a Web Bot Auth signature verified, or the address matched the operator’s published ranges with forward-confirmed reverse DNS
Unproven	the claim stands alone, with nothing corroborating it either way
Forged	the claim was made and the proof failed, which is an impersonation attempt

The report’s Overview carries those totals for the whole property, its Agents tab breaks them down per crawler, and its Handshakes tab reports clients whose TLS handshake contradicts the identity they claimed. See Verified Agents for how each proof is produced.

A GPTBot row with a high AI scraper score therefore means two different things depending on which bucket it fell into, and only the report tells you which.

AI Browser Plugins (Client-Side)

WebDecoy’s Detection Script detects AI browser extensions installed by visitors. These extensions inject DOM elements, global variables, and content scripts that can be identified.

Primary AI Assistants

High-confidence detection of major AI assistant extensions:

Extension	Detection Flags	Score Impact
Claude Extension	`claude_extension`	+15 points
ChatGPT Extension	`chatgpt_extension`	+15 points
GitHub Copilot	`copilot_extension`	+15 points
Gemini/Bard Extension	`gemini_extension`	+15 points
Perplexity Extension	`perplexity_extension`	+15 points

Secondary AI Tools

Other AI-powered browser extensions:

Extension	Detection Flags	Score Impact
Monica AI	`monica_extension`	+10 points
Jasper AI	`jasper_extension`	+10 points
WriteSonic	`writesonic_extension`	+10 points
Merlin AI	`merlin_extension`	+10 points
Sider AI	`sider_extension`	+10 points
MaxAI	`maxai_extension`	+10 points

Detection Methods

WebDecoy uses multiple techniques to detect AI browser plugins:

1. DOM Element Detection

AI extensions inject UI elements with identifiable patterns:

// Claude extension patterns
'[class*="claude"]'
'[data-claude]'
'[data-anthropic]'
'[id*="claude-"]'
'claude-extension'

// ChatGPT extension patterns
'[class*="chatgpt"]'
'[class*="openai"]'
'[data-chatgpt]'
'chatgpt-sidebar'

// Copilot patterns
'[class*="copilot"]'
'[data-copilot]'
'github-copilot'

2. Global Variable Detection

Extensions often expose global objects:

// Claude
window.__claude__
window.__CLAUDE__
window.__anthropic__

// ChatGPT/OpenAI
window.__chatgpt__
window.__openai__
window.chatGPTExtension

// Copilot
window.__copilot__
window.githubCopilot

// Generic AI
window.__aiAssistant__
window.__AI_EXTENSION__

3. Shadow DOM Detection

Modern extensions use Shadow DOM for isolation. WebDecoy inspects:

Custom element tag names (<claude-assistant>, <chatgpt-sidebar>)
Shadow root content for AI-related patterns
Nested shadow DOM structures

4. Custom Element Registry

Extensions register custom HTML elements:

// Checked custom elements
'claude-assistant'
'claude-sidebar'
'claude-popup'
'chatgpt-assistant'
'chatgpt-sidebar'
'gpt-popup'
'copilot-assistant'
'copilot-suggestion'
'perplexity-assistant'
'gemini-assistant'
'monica-assistant'
'merlin-assistant'
'sider-assistant'

Vision AI Detection

Vision AI bots represent a new category of AI-assisted browsing where AI agents directly control browser interactions. Unlike traditional browser extensions that assist users, Vision AI bots operate the browser autonomously by manipulating the DOM directly.

Vision AI detection is part of the standard detection script’s behavioral phase. There is no separate script tier.

What are Vision AI Bots?

Vision AI bots are AI systems that:

See the page visually using screenshots or DOM snapshots
Interact programmatically by setting form values directly via JavaScript
Skip human input patterns by not generating keyboard/mouse events
Operate autonomously to fill forms, click buttons, and navigate sites

┌─────────────────────────────────────────────────────────────────┐
│                 VISION AI BOT INTERACTION                        │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│   Human User                      Vision AI Bot                  │
│   ──────────                      ──────────────                 │
│                                                                  │
│   Keyboard → keydown event        DOM Manipulation               │
│            → keyup event          element.value = "text"         │
│            → input event          ↓                              │
│            → value changes        input event only               │
│                                   (no keyboard events!)          │
│                                                                  │
│   ┌─────────────────────────────────────────────────────────┐   │
│   │  Detection: Field has content but ZERO keyboard events  │   │
│   │  Score Impact: +40 points per field                     │   │
│   └─────────────────────────────────────────────────────────┘   │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Known Vision AI Bots

Bot/Tool	Type	Detection Method
Claude Computer Use	AI Agent	DOM manipulation, no keyboard events
Stagehand	Browser Automation	DOM manipulation, synthetic events
Browserbase Agents	Cloud Browser	DOM manipulation, headless signals
Playwright AI	Testing Framework	DOM manipulation, automation flags
Browser Use	AI Agent Framework	DOM manipulation, programmatic clicks

How Detection Works

WebDecoy uses two complementary techniques to detect Vision AI bots:

Ghost Click Detection - Clicks at positions where the mouse was never near
FormAnalyzer - Form fields with content but no keyboard events

Ghost Click Detection

Real users must move their mouse to a position before clicking there. Vision AI bots using element.click() or browser automation APIs can click anywhere without prior mouse movement.

┌─────────────────────────────────────────────────────────────────┐
│                    GHOST CLICK DETECTION                         │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│   Human User                      Vision AI Bot                  │
│   ──────────                      ──────────────                 │
│                                                                  │
│   1. Move mouse to button         1. Find element via DOM       │
│   2. Hover over button            2. Call element.click()       │
│   3. Click button                 3. No mouse movement at all!  │
│                                                                  │
│   Mouse path recorded:            Mouse path recorded:           │
│   ▸ (100,50) → (150,80) → click   ▸ click at (150,80)           │
│                                     (no prior movements!)        │
│                                                                  │
│   ┌─────────────────────────────────────────────────────────┐   │
│   │  Detection: Click with NO mouse activity within 100px   │   │
│   │  Score Impact: +45 points                                │   │
│   └─────────────────────────────────────────────────────────┘   │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Key features:

Tracks mouse position history with timestamps
Checks if mouse was within 100px of click position in the 2 seconds before click
Immediate report: after 2 clicks with no nearby mouse activity, the script sends a detection right away rather than waiting for the behavioral phase
Works even if page navigates away immediately after click

FormAnalyzer (DOM Manipulation Detection)

WebDecoy’s FormAnalyzer tracks keyboard events per form field and compares them against actual content:

1. Event Tracking

// FormAnalyzer tracks per-field keyboard activity
{
  "email_field": {
    keydownCount: 0,      // No keydown events
    keyupCount: 0,        // No keyup events
    inputCount: 1,        // One input event (from DOM manipulation)
    pasteCount: 0,        // No paste events
    contentLength: 25     // But field has 25 characters!
  }
}

2. Pattern Detection

Each suspicious field is reported in metadata.behavior.formAnalysis.suspiciousFields with a reason, and WebDecoy’s server adds to the score per field:

`reason`	Detection	Score
`content_without_keyboard`	Field has content but no keydown and no paste events	+40
`high_dom_manipulation_ratio`	Value changes without recent keyboard activity outnumber half the keystrokes	+25
`abnormal_keydown_keyup_ratio`	Keydown/keyup ratio outside 0.5 to 2.0	+15

When totalDomManipulations reaches 3 or more, another +20 is added.

3. Typing Rhythm Analysis

Real humans have variable typing rhythms. Bots often have:

Zero variance in keystroke timing
Superhuman speed (< 30ms between keystrokes)
Perfect consistency across all fields

Detection Flags

Click patterns are reported as flags (stored in the detection’s metadata.triggers). Form analysis is not a flag: it travels in metadata.behavior.formAnalysis and is scored as described above.

Flag	Meaning	Score Impact
`ghost_clicks`	Clicks at positions with no prior mouse activity	+45
`instant_clicks`	Clicks with no delay after mouse stops	+30
`clicks_no_pre_movement`	Clicks without preceding mouse movement	+25

Example Detection Record

{
  "id": "det_xyz789",
  "source": "bot_scanner",
  "unified_score": 85,
  "threat_level": "HIGH",

  "metadata": {
    "triggers": ["ghost_clicks"],
    "behavior": {
      "clickAnalysis": {
        "ghostClicks": true,
        "ghostClickRatio": 0.67,
        "instantClicks": false,
        "noPreMovement": true
      },
      "formAnalysis": {
        "suspiciousFields": [
          {
            "id": "email",
            "reason": "content_without_keyboard",
            "contentLength": 24,
            "confidence": 0.9
          },
          {
            "id": "message",
            "reason": "content_without_keyboard",
            "contentLength": 156,
            "confidence": 0.9
          }
        ],
        "totalDomManipulations": 2,
        "fields": {
          "email": {
            "contentLength": 24,
            "keyCount": 0,
            "noKeyboardEvents": true
          },
          "message": {
            "contentLength": 156,
            "keyCount": 0,
            "noKeyboardEvents": true
          }
        }
      }
    }
  }
}

Distinguishing from Legitimate Use Cases

Not all DOM manipulation is malicious. WebDecoy accounts for:

Scenario	How It Is Handled
Copy/paste	A field with a paste event is not reported as content without keyboard
Password managers and browser autofill	Can fill a field without keyboard events and so can look like DOM manipulation. Review the rest of the detection before acting on form analysis alone
Form pre-population	A field your own page pre-fills has content without keyboard events and can be reported the same way

What to do with these signals

Ghost clicks and form DOM manipulation are recorded as flags on the detection and raise its score. There is no separate rule language for flags. To act on them:

Review them: open a detection to see its flags and form analysis.
Protect the paths that matter: add your signup, login or checkout routes as protected paths, so a client that cannot earn clearance is challenged or refused there.
Route them elsewhere: a webhook or SIEM integration receives the detection, flags included, so you can alert or act in your own tooling.

Filtering Vision AI Detections

The Detections table and API filter by source and score, not by individual flag:

Filter	Value	What It Shows
Source	`bot_scanner`	All client-side detections, where Vision AI flags come from
Minimum unified score	`70`	High-confidence activity

Open a detection to see which flags it carries.

API Example (the management API takes the Auth0 session token the dashboard uses, not an sk_live_ API key; see API Reference):

# High-confidence client-side detections
curl -H "Authorization: Bearer $WEBDECOY_SESSION_JWT" \
  "https://api.webdecoy.com/api/organizations/$ORG_ID/detections?source=bot_scanner&min_unified_score=70"

Browser API observations

Generic API wrapping is retained as an observation. It does not identify an AI extension or add risk by itself; security and analytics integrations can wrap the same browser APIs.

API Wrapping

Signal	Flag	Score Impact	Meaning
Wrapped Fetch API	`wrapped_fetch`	0 points	`window.fetch` wrapping observed; no AI plugin attribution
Wrapped XHR	`wrapped_xhr`	0 points	`XMLHttpRequest.open` wrapping observed; no AI plugin attribution
Modified MutationObserver	`modified_mutation_observer`	+5 points	Observer constructor has been patched

The detection script places wrapped_fetch and wrapped_xhr in aiBrowserPlugins.observations, separately from aiBrowserPlugins.plugins. When a report is sent, the flags remain in metadata.triggers for diagnostics. The ingest scorer also assigns them zero risk, including flags from older tags. Independent extension and automation evidence still contributes to the score.

Detection Logic

// Check if fetch has been wrapped
var fetchStr = window.fetch.toString();
if (fetchStr.indexOf('[native code]') === -1) {
  // Observation only; the wrapper's purpose is unknown.
  flags.push('wrapped_fetch');
}

// Check if XHR has been wrapped
var xhrStr = XMLHttpRequest.prototype.open.toString();
if (xhrStr.indexOf('[native code]') === -1) {
  // XHR has been wrapped
  flags.push('wrapped_xhr');
}

Signal	Flag	Meaning
AI Context Menu	`ai_context_menu`	Extension added AI options to right-click menu

Understanding Detection Flags

When AI activity is detected, the script’s flags are stored on the detection in metadata.triggers:

Flag Categories

┌─────────────────────────────────────────────────────────────────┐
│                    DETECTION FLAGS BREAKDOWN                     │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│   Primary AI Extensions (15 pts each)                           │
│   ────────────────────────────────────                          │
│   claude_extension, chatgpt_extension, copilot_extension,       │
│   gemini_extension, perplexity_extension                        │
│                                                                  │
│   Secondary AI Extensions (10 pts each)                         │
│   ─────────────────────────────────────                         │
│   monica_extension, jasper_extension, writesonic_extension,     │
│   merlin_extension, sider_extension, maxai_extension            │
│                                                                  │
│   Suspicious Signals (5 pts each)                               │
│   ───────────────────────────────                               │
│   modified_mutation_observer                                   │
│                                                                  │
│   Diagnostic API observations (0 pts)                         │
│   wrapped_fetch, wrapped_xhr                                    │
│                                                                  │
│   Other Signals (8 pts each)                                    │
│   ─────────────────────────                                     │
│   generic_ai_extension, ai_context_menu                         │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Example Detection Record

{
  "id": "det_abc123",
  "source": "bot_scanner",
  "ip_address": "192.168.1.100",

  "unified_score": 45,
  "threat_level": "MEDIUM",

  "metadata": {
    "triggers": ["claude_extension", "wrapped_fetch"]
  },

  "score_components": {
    "ai_scraper_score": 0
  }
}

Filtering AI Detections

In the Detections table, filter specifically for AI-related activity:

Filter	How to Use	Effect
Source = bot_scanner	Select source filter	Show client-side detections, where AI browser plugin flags come from
Minimum unified score	Set minimum score	Show only high-confidence activity
Impersonation only	`?agent_impersonation=true` on the detections URL	Show only requests that claimed a known agent and failed the proof
One agent	`?ai_scraper_name=GPTBot` on the detections URL	Show every request that claimed that identity, proven or not

The forged counts in the agent trust report are links that set exactly these two parameters, so clicking a number lands you on the requests behind it rather than on a filter you have to rebuild.

Individual flags such as claude_extension are not a filter: open a detection to see its flags.

Score Impact

AI browser plugin detection contributes to the overall bot score:

Scoring Breakdown

Detection	Points Added	Rationale
Single primary AI extension	+15	Strong indicator of AI-assisted browsing
Multiple primary AI extensions	+15 each	Compounding evidence
Secondary AI extension	+10	Common AI productivity tools
Fetch/XHR wrapping	0	Diagnostic observation; does not identify an AI extension
Modified MutationObserver	+5	Separate constructor-patching signal
Generic AI signals	+8	Unspecified AI activity

Score Examples

Scenario	Flags	Added Score
Claude extension only	`claude_extension`	+15
ChatGPT + wrapped fetch	`chatgpt_extension`, `wrapped_fetch`	+15
Multiple AI tools	`claude_extension`, `monica_extension`, `copilot_extension`	+40
Wrapped APIs only	`wrapped_fetch`, `wrapped_xhr`	0

Use Cases

AI browser plugin flags raise a detection’s score; there is no separate rule language for them.

Content and forms: add the routes you care about (premium content, signup, checkout) as protected paths. Enforcement there is per site and starts in Monitor mode.
Alerting: send detections to a webhook or SIEM integration and alert on the flags in your own tooling.

WebDecoy does not report on visitors who are not automated, so it is not a tool for measuring AI adoption among your human audience.

API Access

Access AI detection data via the WebDecoy API:

Detection Response Fields

Client-side flags are in metadata.triggers. The AI scraper fields are inside score_components, and ai_scraper_category and ai_scraper_name appear only when a known agent matched:

{
  "metadata": {
    "triggers": ["claude_extension", "wrapped_fetch"]
  },
  "score_components": {
    "ai_scraper_score": 0
  }
}

Filtering API Requests

# Client-side detections, where AI browser plugin flags appear
curl -H "Authorization: Bearer $WEBDECOY_SESSION_JWT" \
  "https://api.webdecoy.com/api/organizations/$ORG_ID/detections?source=bot_scanner"

# Every request that claimed to be GPTBot
curl -H "Authorization: Bearer $WEBDECOY_SESSION_JWT" \
  "https://api.webdecoy.com/api/organizations/$ORG_ID/detections?ai_scraper_name=GPTBot"

The API does not filter by flag; read metadata.triggers on each detection returned.

Privacy Considerations

AI browser plugin detection is designed to be:

Non-invasive: Only checks for presence, doesn’t read content
Privacy-respecting: Doesn’t capture what users type into AI tools
Transparent: Detection signals are clearly documented

LLM Referral Tracking

When a visitor arrives at your site by clicking a link in an AI product (a ChatGPT conversation, a Perplexity answer, a Claude response), WebDecoy counts the visit as an AI referral. Human visits are counted, never stored as detections.

How It Works

The detection script sends an install heartbeat once per session. On the session’s first page it adds where the visit came from, reduced before it leaves the page: the referrer’s origin only (and only when it is another site), plus the landing path with just the utm_source, ref and utm_medium parameters. WebDecoy classifies that against its AI platform table and adds one to a daily counter for the platform and landing path. Nothing else about the visit is kept.

Visitor clicks link in ChatGPT → Lands on your site
    │
    ├── document.referrer = "https://chatgpt.com/..."
    │
    ├── Heartbeat carries { r: "https://chatgpt.com", u: "/landing-path" }
    │
    └── WebDecoy counts one ChatGPT referral for that day and landing path

When there is no Referer, a campaign tag that names a known platform (for example ?utm_source=chatgpt.com) is used instead. A Referer from any other site always wins over a tag. Server-side sensors (edge, Vercel log drain, WordPress plugin, Node SDK) can report the same counts; when more than one source reports for a site on a day, WebDecoy keeps one so visits are not counted twice.

Likely referrals (Vercel log drain)

Many AI apps strip the referrer, so a person who clicks a link in an answer often arrives looking like a direct visit. The Vercel log drain sees the other half of that click: when someone asks an assistant about a page, its user fetcher (ChatGPT-User, Perplexity-User, Claude-User, Gemini-User, Meta-ExternalFetcher) fetches the page from your site moments before.

A visit is counted as a likely referral from that assistant when:

it arrived with no referrer, no utm_ tag and no ad click id (gclid, fbclid and similar),
exactly one assistant’s user fetcher fetched the same page (query string ignored, trailing slash ignored) in the 15 minutes before, or in the same minute,
and the visit was not itself reported as a bot.

Training and search crawlers never count, because nobody is waiting on them. Claude Code’s fetches never count, because they run for a developer. If two assistants fetched the page in the window, the visit is not counted at all rather than credited to one of them. Each visitor counts once per page per delivery.

Likely referrals are an estimate and are shown as one: on the Referrals tab as “+N likely” beside each platform’s counted visits, with an explanation. They are never added into the referral count, the dashboard total, or the crawl-to-referral ratio. A fetch can be for someone who read the answer and never clicked, and a direct visit a minute later can be someone else. User-fetcher IPs are not verified on this path, so a forged fetcher could raise the estimate. Only counts are stored, never anything about a visitor.

Supported AI Platforms

Platform	Domains Tracked
ChatGPT	`chat.openai.com`, `chatgpt.com`
Perplexity	`perplexity.ai`, `www.perplexity.ai`
Gemini	`gemini.google.com`
Claude	`claude.ai`
DeepSeek	`deepseek.com`, `www.deepseek.com`, `chat.deepseek.com`
Copilot	`copilot.microsoft.com`
You.com	`you.com`, `www.you.com`
Phind	`phind.com`, `www.phind.com`
Kagi	`kagi.com`, `www.kagi.com`
Meta AI	`meta.ai`, `www.meta.ai`
Grok	`grok.com`, `www.grok.com`

Viewing LLM Referral Data

AI referrals appear on the AI Traffic page and in the dashboard’s AI referral summary, which show:

Referrals per AI platform
Referrals per day
Top landing pages receiving AI referral traffic

Referral fields on detections

A referral is not a detection. But when a request that is stored as a detection (a bot that scored high enough, a decoy hit) carries an AI platform referrer or campaign tag, ingest records it in the detection’s metadata:

Metadata key	Value
`is_llm_referral`	`"true"`
`llm_platform`	Platform name, e.g. `ChatGPT`
`llm_category`	`ai_assistant` or `ai_search`
`llm_referral_signal`	`referer`, or the campaign parameter that matched (`utm_source`, `ref`, `utm_medium`)

AI Scraper Scoring - Detailed AI crawler scoring methodology
Verified Agents - Proven, unproven and forged agent claims, and the report that splits them
Detection Script - JavaScript-based bot detection
Threat Scoring - Overall threat assessment
Viewing Detections - Working with detection data

AI Detections

Overview

AI Scrapers (Server-Side)

Training Crawlers (High Score)

AI Search Crawlers, Assistants and Commercial Scrapers

Search Crawlers (Lower Score)

SEO and Analytics Crawlers

HTTP Client Libraries and CLI Tools

Where crawler activity is reported

AI Browser Plugins (Client-Side)

Primary AI Assistants

Secondary AI Tools

Detection Methods

1. DOM Element Detection

2. Global Variable Detection

3. Shadow DOM Detection

4. Custom Element Registry

Vision AI Detection

What are Vision AI Bots?

Known Vision AI Bots

How Detection Works

Ghost Click Detection

FormAnalyzer (DOM Manipulation Detection)

1. Event Tracking

2. Pattern Detection

3. Typing Rhythm Analysis

Detection Flags

Example Detection Record

Distinguishing from Legitimate Use Cases

What to do with these signals

Filtering Vision AI Detections

Browser API observations

API Wrapping

Detection Logic

Context Menu Modifications

Understanding Detection Flags

Flag Categories

Example Detection Record

Filtering AI Detections

Filter Options

Score Impact

Scoring Breakdown

Score Examples

Use Cases

API Access

Detection Response Fields

Filtering API Requests

Privacy Considerations

LLM Referral Tracking

How It Works

Likely referrals (Vercel log drain)

Supported AI Platforms

Viewing LLM Referral Data

Referral fields on detections

Related Documentation