Claude Prompt Actor - GEO – API + Web Search, Sources & Citations

Ask Claude anything, with live web search, and capture the answer and every source it cites.

This Actor sends your prompts to Claude through the official Anthropic API and returns each answer as text and Markdown, together with the cited sources (URL, title, domain) and an aggregated list of cited domains — the core metric for Generative Engine Optimization (GEO).

Claude's login-only chat UI can't be scraped reliably, so this Actor uses the API directly. By default it runs with a compact, Claude.ai / Claude Desktop-style system prompt and Anthropic's server-side web search tool, so the answers and citations resemble what a real Claude Desktop user sees.

Use it to:

• Monitor AI search visibility — is your brand mentioned and cited in Claude's answers?
• Track competitors — which domains does Claude cite for your money keywords?
• GEO & SEO reporting — measure AI answer coverage across hundreds of prompts
• Content research — see exactly what Claude summarizes for a topic, with sources

Bring Your Own Key (BYOK)

This Actor is BYOK: you provide your own Anthropic API key in the encrypted apiKey input.

Where do I get a key?

1. Go to https://platform.claude.com/settings/keys (sign in / create a free Anthropic account).
2. Click Create Key, give it a name, and copy the value — it starts with sk-ant-.
3. Paste it into this Actor's apiKey input field.

See Pricing below for what this costs.

Is it safe to enter my key here?

Yes. The apiKey field is a secret input, which means:

• It's stored encrypted. Apify encrypts secret inputs at rest; the value is not shown in the run's input view, is not written to the run log, and is not saved into the output dataset. This Actor never prints or stores it.
• It's only used to call Anthropic. The key is passed straight to the official Anthropic API (api.anthropic.com) to answer your prompts — nothing else, no other destination.
• You keep full control. Tokens are billed to your Anthropic account at cost (no markup), you can set spend limits in the Anthropic console, and you can revoke or rotate the key any time at https://platform.claude.com/settings/keys — that instantly disables it everywhere.
• No reselling. This Actor does not resell Anthropic access; you run on Anthropic's own terms with your own account.

Tip: create a dedicated key for this Actor (and optionally a monthly spend limit). Then you can revoke just that key later without affecting anything else.

Pricing

This Actor uses pay per event + usage pricing:

• One prompt-answered event per successfully answered prompt. Prompts that fail (refusal, empty response, API error) are not charged.
• Your own Anthropic token cost (BYOK) is billed to you directly by Anthropic — this Actor never marks it up or resells it.
• Apify platform usage (compute) is billed separately to you at cost.

So a run costs: the per-answered-prompt event fee + your Anthropic tokens + a little Apify compute. Lower maxSearchUses/maxTokens or pick Sonnet/Haiku to reduce the token part.

Input

Field	Type	Description
prompts	array (required)	Questions/prompts to send to Claude. Each runs in its own independent conversation.
apiKey	string (required)	Your Anthropic API key (sk-ant-…). Stored encrypted (BYOK).
model	string	claude-opus-4-8 (default, most capable), claude-sonnet-5, claude-haiku-4-5, claude-fable-5.
overrideSystemPrompt	string	Empty = use the built-in Claude.ai-style prompt (applied in code). Set only to replace it. {{CURRENT_DATE}} → today.
webSearch	boolean	Enable Anthropic's server-side web search tool (default true). Off = model-knowledge only, no sources.
maxSearchUses	integer	Max web searches per prompt (default 5).
extendedThinking	boolean	Adaptive extended thinking (default false). Ignored for Haiku 4.5.
maxTokens	integer	Max answer length in tokens (default 8192). Requests are streamed, so large values are safe.
maxConcurrency	integer	Parallel prompts (default 3). Bounded by your Anthropic rate limits; no extra memory needed.
maxRequestRetries	integer	Retries per prompt on transient API errors (default 2). Reported per item as attempts.

Output

One dataset item per prompt. The schema is aligned with the other Actors in this project (ChatGPT, Google AI Overview) so results combine easily.

• prompt: the input prompt
• success: true if an answer was returned, false otherwise
• error: null on success, otherwise an error marker (REFUSAL: …, EMPTY_RESPONSE, an API error message, …)
• attempts: number of tries it took (including the final one)
• model: the model that answered
• responseText: the answer as plain text
• responseMarkdown: the answer with Markdown formatting preserved
• sources: array of { position, text, url, domain, type }, deduplicated by normalized URL
  • type: "citation" — a source Claude cited inline in its answer
  • type: "source" — a page returned by web search (considered, maybe not cited inline)
• sourceDomains: unique host names across sources (aggregated GEO metric; full hostname, no www stripping)
• webSearchCount: how many web searches Claude ran for this prompt
• usage: token usage { input_tokens, output_tokens, … } — handy for tracking your Anthropic cost
• stopReason: why generation stopped (end_turn, max_tokens, refusal, …)
• responseTimeMs, promptIndex, totalPrompts, timestamp

Notes

• Cost is on your Anthropic account. Web search and larger models cost more tokens; lower maxSearchUses, pick Sonnet/Haiku, or reduce maxTokens to save.
• Answers take a while. A single web-searched Opus answer typically runs ~10–40s (Claude searches, reads pages, then writes). This is normal. To finish a batch faster, raise maxConcurrency so prompts run in parallel, or use a faster model (Sonnet/Haiku). The run log and Run-tab status update after every prompt so you can follow progress.
• stopReason: "max_tokens" means the answer was cut off — raise maxTokens.
• Answers are non-deterministic — wording, sources and counts vary between identical runs, and a prompt that needs no current facts may return 0 sources.
• Fable 5 requires 30-day data retention on your Anthropic org (not available under zero-data-retention) and may occasionally refuse a request (reported as success: false, error: "REFUSAL: …").