How SerpApi, Tavily, Exa, and Firecrawl differ when they return Markdown for LLMs

# Quick summary

# What each provider returns

• SerpApi: converts the search itself and returns the full search engine results page (SERP) as clean Markdown or as structured JSON. That means the Markdown represents the results page a real search engine produced.
• Tavily, Exa, Firecrawl: convert the pages that the search found into Markdown. The Markdown therefore reflects the content on result pages rather than the SERP structure.

This distinction matters. If your agent needs the results page context (rank, SERP features, snippets), SerpApi's Markdown preserves that. If the agent needs the original pages' content (full articles, product pages), the other providers' Markdown gives the source material directly.

# How the comparison was run All providers used their playground defaults with generated answers disabled. Location was set as close to Austin, Texas, as the provider allowed—United States for Tavily and Exa, Austin for Firecrawl. SerpApi used Google Search with location=Austin, Texas, United States plus gl=us and hl=en, and searches were live (no_cache=true).

Three queries tested distinct behaviors:

• coffee — a one-word query used to surface many SERP features.
• how to make delicious coffee — a semantic question representing typical user-forwarded questions.

Token counts were measured using tiktoken with the o200k_base encoding on the full saved response (JSON or Markdown) without stripping fields.

# Representative token numbers and formats

• SerpApi Google Search JSON: 66,237 tokens (full SERP as structured objects)
• SerpApi Google Search Markdown: 13,025 tokens (full SERP as a document)

• SERP-as-Markdown (SerpApi): models see ranking, snippets, structured features (knowledge panels, shopping boxes, etc.) and can reason about the search results themselves—useful for summarization, ranking-aware agents, and workflows that need SERP context.
• Page-as-Markdown (Tavily, Exa, Firecrawl): models read the pages the results link to and can extract and synthesize page-level facts—useful for retrieval-augmented generation (RAG), citation extraction, and tasks that require source text.

# Practical recommendations

• If your agent needs the search context (which results appeared, SERP features, relative ranking), choose the provider that returns the SERP as Markdown.

• Consider token costs: Markdown representations of the SERP can cut token usage compared to full JSON while still giving structured, readable context for LLM consumption.

# Bottom line