Surface fidelity: what each engine result actually measures

May 26, 2026 · View on GitHub

Every tool response includes an interpretation_note field that summarizes this table in one sentence. Read it before drawing conclusions.

Surface types

Surface	What it means
`consumer_scrape`	Result proxied from a real consumer-facing AI search product. Closest to what your users see.
`api_proxy`	API call to a search-enabled LLM. May differ from consumer product — different model tier, no UI-level ranking, no personalization. Use as a directional signal, not ground truth.
`web_rank`	Traditional SERP rank. Measures whether a URL appears in search results, NOT whether an LLM cites it.
`static_signal`	Offline signal computed from public data. No live query.

SERPAPI_KEY — SerpAPI scrapes the real Google AI Overview (the AI-generated summary at the top of google.com search results). This is pixel-accurate to what google.com desktop users see. Requires a SerpAPI key with AI Overview access enabled.

When to use: Best proxy for Google AI Mode citations. Use this as your benchmark if you have a SERPAPI_KEY.

Caveat: SerpAPI is a scrape of a live page, so results vary with locale, personalization, and test timing. Geographic variation can be significant.

`perplexity` — `consumer_scrape` ★★★★☆

PERPLEXITY_API_KEY — sonar-pro API with a consumer-equivalent system prompt. Perplexity's API is the closest public API to the Perplexity.ai product.

interpretation_note: "sonar-pro API with a consumer-equivalent system prompt. Real perplexity.ai users may see sonar-reasoning-pro with multi-turn follow-ups."

When to use: Default auto engine when no SERPAPI_KEY is configured. Very good fidelity for Perplexity.ai citations.

Caveat: Perplexity.ai's consumer product uses sonar-reasoning-pro for complex queries and does multi-turn refinement. The API does one-shot queries. Citation differences are real but usually minor.

`openai` — `api_proxy` ★★★☆☆

OPENAI_API_KEY — gpt-4o via the Responses API with the web_search_preview tool. OpenAI retired the gpt-4o-search-preview alias; base gpt-4o + the tool is the supported path and produces equivalent grounding.

interpretation_note: "gpt-4o + web_search_preview tool via the Responses API. Real chatgpt.com users get gpt-4o + SearchGPT with different ranking and UI-level re-scoring."

When to use: Reasonable proxy for ChatGPT web search citations. Better than gpt-4o-mini (the old default), worse than a direct ChatGPT product scrape.

Caveat: ChatGPT's consumer product applies UI-level ranking and re-scoring that the API doesn't expose. Citation behavior can differ materially for ambiguous queries. Treat as directional.

`claude` — `api_proxy` ★★★☆☆

ANTHROPIC_API_KEY — claude-sonnet-4-7 via the Anthropic Messages API with web_search tool enabled (max 5 uses).

interpretation_note: "claude-sonnet-4-7 API with web_search tool (max 5 uses). Real claude.ai users get a different model tier and citation UI."

When to use: Directional proxy for Claude.ai web search citations.

Caveat: Claude.ai's consumer product uses different model routing, does multi-turn refinement, and has a distinct citation UI. API results are directional — useful for relative comparisons but not for absolute "Claude cites me" claims.

`gemini` — `api_proxy` ★★★☆☆

GEMINI_API_KEY — gemini-2.5-pro via the Generative Language API with google_search grounding.

interpretation_note: "gemini-2.5-pro API with google_search grounding. Real gemini.google.com uses the same grounding index but different re-ranking."

When to use: Directional proxy for Gemini.google.com citations.

Caveat: Gemini's consumer app uses the same Google Search grounding index but different re-ranking and result selection. Citation patterns are similar but not identical.

`bing_serp` — `web_rank` ★★☆☆☆

BING_API_KEY — Bing Web Search API v7.

interpretation_note: "Bing Web Search API — traditional SERP rank, NOT LLM citation behavior."

When to use: Comparing web SERP rank against LLM citation rank for the same query. Not for "does Bing AI cite me?" questions.

Caveat: am_i_cited refuses this engine — it only measures LLM behavior. Use check_citations directly with engine=bing_serp for SERP rank data.

`brave_serp` — `web_rank` ★★☆☆☆

BRAVE_API_KEY — Brave Search API.

interpretation_note: "Brave Search API — traditional SERP rank, NOT LLM citation behavior."

When to use: Same as bing_serp — SERP rank comparison only.

Caveat: Same as bing_serp. Free tier: 2000 queries/month; built-in 1.1 s delay between calls to avoid 429.

Honest use of this data

Never report API-proxy results as "ChatGPT/Claude/Gemini cites you" without the interpretation_note caveat. The API and the consumer product behave differently.
Prefer google_ai_mode + perplexity when you want results closest to real consumer behavior.
Use citation_provenance for multi-engine consensus. A URL with engine_count=4 (cited by 4 of 5 engines) is a strong signal. engine_count=1 is engine-specific and may not generalize.
Use am_i_cited with engine=auto to get the full picture for your domain in one call.
Watch for recency bias. LLM search engines favor recently-indexed content. Compare results across different days before concluding something is or isn't cited.