Chatgpt Prompt Actor - GEO Analyse with Citation Data

ChatGPT Prompt Actor

This Actor automates interactions with ChatGPT by sending prompts and extracting responses. It uses PlaywrightCrawler with Camoufox - a stealthy fork of Firefox - to navigate ChatGPT's web interface and perform automated conversations.

What this Actor does:

1. Navigates to ChatGPT: Opens the ChatGPT web interface at https://chatgpt.com/
2. Handles UI Elements: Automatically closes any modal dialogs that might appear
3. Processes Multiple Prompts: Sequentially sends all prompts from your input array
4. Sends Prompts: Enters each prompt into ChatGPT's text input field
5. Waits for Response: Monitors the page until ChatGPT completes its response (up to 2 minutes)
6. Extracts Results: Captures the AI-generated response text and also citations, and news items
7. Saves Data: Stores all prompts, responses, and metadata in a structured dataset

Input Parameters:

• prompts (array): Messages/questions to send to ChatGPT — each runs in its own fresh conversation
• maxConcurrency (integer, default 1): How many browser sessions to run in parallel (see "Parallel processing" below)
• maxRequestRetries (integer, default 3): Retries per session on failure — each retry uses a fresh browser and IP; already-answered prompts in the session are skipped
• promptsPerSession (integer, default 10): How many prompts to answer in one browser before starting a fresh browser/IP. Higher = faster (a running ChatGPT answers follow-ups in ~5s vs ~25s to start a new browser); those prompts share one IP
• blockResources (boolean, default true): Block images/fonts/media to cut residential proxy traffic (the biggest cost driver). JS/CSS/XHR are kept, so response extraction is unaffected
• proxyCountry (string): Select which country the proxy should be located in (optional, auto-selects if empty)

Parallel processing (chunking)

Prompts are grouped into sessions (chunks) of promptsPerSession. One session = one browser on one IP that answers its prompts one after another (reusing the browser via "new chat" — much faster than starting a new browser per prompt). maxConcurrency controls how many sessions run in parallel.

• Number of sessions = ⌈ number of prompts ÷ promptsPerSession ⌉
• Run in parallel = up to maxConcurrency sessions at a time

Each parallel session needs its own browser (~1 GB RAM), so raise the Actor's memory when you increase maxConcurrency. Parallelism makes the result arrive faster (less wall-clock time) for only a slight increase in usage cost — the total compute is roughly the same.

Examples (25 prompts):

Config	Sessions	Behaviour	Memory
promptsPerSession: 10, maxConcurrency: 1	3 sessions of 10 / 10 / 5	run one after another (1 browser at a time)	2 GB (default)
promptsPerSession: 10, maxConcurrency: 3	3 sessions of 10 / 10 / 5	run all 3 at once — fastest	~4 GB (≈ 1 GB per parallel session)

Note: the chunk size is set purely by promptsPerSession. With promptsPerSession: 10 and 10 prompts you always get one session of 10 — maxConcurrency only matters once there is more than one session. For two parallel sessions of 5, set promptsPerSession: 5.

Output:

The Actor saves one dataset item per prompt:

• prompt: Original input prompt
• success: true if a response was extracted, false otherwise
• error: null on success, otherwise an error message (e.g. EMPTY_RESPONSE, REQUEST_FAILED_AFTER_RETRIES)
• attempts: Number of tries it took (including the final one)
• responseText: ChatGPT's response as plain text
• responseMarkdown: ChatGPT's response with formatting preserved (headings, lists, tables, code blocks) — ideal for further processing
• sources: Web sources as { position, text, url, domain, type }, deduplicated by URL. The actor opens ChatGPT's full "Sources" sidebar, so bundled/hidden citations are included — not just the 2–3 inline pills. The type field tells you where each link came from:
  • citation — cited inline in the answer text
  • source — listed in the sources sidebar
  • more — additionally searched by ChatGPT but not directly cited
• sourceDomains: Unique host names of cited sources (citation/source only, e.g. ["www.nike.com", "bbc.com"]) — the aggregated GEO metric. Host names are kept as-is (no www/subdomain stripping), so www.nike.com and nike.com count separately
• newsListing: Array of news items if present in responses
• responseTimeMs: How long the response generation took
• promptIndex: Index of current prompt (0-based)
• totalPrompts: Total number of prompts in batch
• timestamp: ISO timestamp of when the interaction occurred

Breaking change (v0.1): response was replaced by responseText/responseMarkdown, citations was renamed to sources (now with domain and position, and now includes bundled/hidden citations — not just inline pills), sourceDomains was added, the url field was removed, and success/error are now present on every item.

Features:

• Stealth Browsing: Uses Camoufox to avoid detection and blocking
• Proxy Support: Automatically uses Apify proxy configuration when running on the platform
• Country Selection: Choose proxy location from 35+ countries for geo-specific responses
• Error Handling: Gracefully handles timeouts and extraction failures
• Modal Handling: Automatically closes ChatGPT's welcome dialogs and popups