Skip to content
GeoReady Docs

geo citations: LLM Providers & API Setup

geo citations asks a real AI answer engine the questions your customers ask, then checks whether your brand is mentioned and your domain is cited as a source. It needs an API key for whichever provider you want to query. This page covers provider-specific setup that doesn’t fit in the README Quick Start: wire formats, regional endpoints, and model capabilities.

Perplexity is the recommended default (PERPLEXITY_API_KEY) because Perplexity Sonar returns real web citation URLs, not just parametric knowledge. OpenAI, Anthropic, and Groq reveal what the model already knows about your brand from training data, which is a different (and weaker) signal — see the README’s Additional tools section for the distinction.

MiniMax

Section titled “MiniMax”

MiniMax supports either wire format (OpenAI-compatible or Anthropic-compatible) and both a global and a China API root:

Terminal window
export MINIMAX_API_KEY="your-api-key"
export MINIMAX_API_FORMAT="openai"  # or "anthropic"
export MINIMAX_API_BASE_URL="https://api.minimax.io/v1"
geo citations --provider minimax --brand "YourBrand" --domain yoursite.com

For the China endpoint, use https://api.minimaxi.com/v1 with the openai format, or https://api.minimaxi.com/anthropic with the anthropic format. The global messages-format root is https://api.minimax.io/anthropic. The default model is MiniMax-M3.1-Flash. MINIMAX_THINKING accepts adaptive or disabled; Flash normalizes either explicit setting to adaptive, while M3 preserves the selected mode. Leaving it unset preserves the API defaults: Flash adaptive; M3 adaptive on OpenAI and disabled on Anthropic. Effort is not configured by this client, so Flash uses the API default (max).

Set GEO_LLM_MODEL to select M3 or an explicit custom/legacy model ID:

Model Context window API input modalities Thinking
MiniMax-M3.1-Flash (default) 1,000,000 tokens Text, image, video adaptive only
MiniMax-M3 1,000,000 tokens Text, image, video adaptive or disabled

geo citations currently sends text prompts only. The modalities column lists what each model’s API accepts; query_llm currently types text and image content parts, so video input needs a schema addition before it can be passed. See the official MiniMax pricing page for current rates.

Gemini

Section titled “Gemini”

Gemini is checked directly against the Gemini API, not just simulated via crawler user-agents:

Terminal window
export GEMINI_API_KEY="your-api-key"
geo citations --provider gemini --brand "YourBrand" --domain yoursite.com

Set GEO_LLM_MODEL to pick a specific model (default: gemini-3.7-flash). This calls generativelanguage.googleapis.com directly, not Vertex AI, so no extra dependency is needed.

DeepSeek

Section titled “DeepSeek”

DeepSeek covers citation checks against the Chinese AI answer-engine ecosystem, which Western-only providers miss entirely:

Terminal window
export DEEPSEEK_API_KEY="your-api-key"
geo citations --provider deepseek --brand "YourBrand" --domain yoursite.com

Fully OpenAI-compatible wire format against api.deepseek.com. Default model deepseek-v4-flash; set GEO_LLM_MODEL=deepseek-v4-pro for the higher-capability tier.

SerpBase (Google SERP + AI Overview)

Section titled “SerpBase (Google SERP + AI Overview)”

Google AI Overviews isn’t an LLM you can query — it’s a SERP feature — so --provider serpbase takes a different path from every other provider on this page: instead of asking a model a question, it runs a real Google search via serpbase.dev and checks whether your brand appears in the organic results or the AI Overview block (when Google renders one for the query), and whether your domain is among the cited sources.

Terminal window
export SERPBASE_API_KEY="your-api-key"
geo citations --provider serpbase --brand "YourBrand" --domain yoursite.com

Bring-your-own-key, pay-as-you-go: 100 free searches, then $0.30/1k (pricing). Never auto-detected — SERPBASE_API_KEY alone does nothing unless --provider serpbase is passed explicitly, since it’s a different kind of check (SERP observation, not an LLM answer) and a paid one past the free tier.

Two things this provider does differently from the rest:

  • --runs is ignored. A Google SERP isn’t resampled the way a non-deterministic LLM answer is — asking again immediately would mostly return the same page, just at the cost of another API call. Every check is a single live snapshot per query.
  • The AI Overview block isn’t always present. Google only renders one for some queries, and serpbase’s docs don’t publish its internal field schema, so ai_overview is parsed defensively — a query without one still runs the citation check against the organic results, it just won’t have an AI Overview signal to add. How often it actually surfaces is logged (serpbase: AI Overview present in N/M queries) rather than promised.