Both kinds exist. Brave, Exa, Parallel and Perplexity say in their own docs that they crawl and run their own index, while SerpApi, Serper and the other Google SERP APIs return Google's live results page. Almost nobody resells Bing any more, because Microsoft retired the Bing Search API in August 2025. Tavily does not say where its results come from. The search tools built into OpenAI, Claude and Gemini sit in a fourth group: the model vendor picks the index, and only Gemini names it.
- Own index: Brave says "there is no fallback to another search engine for any query". Exa, Parallel and Perplexity also describe their own crawlers and indexes.
- Google results: SERP APIs get them by loading Google's public results page. Google's official Custom Search JSON API is closed to new customers and ends on January 1, 2027.
- Measured overlap: in our runs of 500 queries, the share of top-10 URLs that matched Google's top 10 was 55.5% for Tavily, 46.6% for Brave and 32.3% for Exa.
Where does each search API get its results?
The second column states only what the vendor's own pages say, read on November 16, 2026.
| API | Source | What the vendor says |
|---|---|---|
| Brave Search API | Own index | "Every result comes from Brave's own independent index, crawled and ranked by Brave." (brave.com) |
| Exa | Own index | Search "the largest index of public web and private information", and "crawl billions of documents per day". (exa.ai) |
| Parallel | Own index | "Powered by our own proprietary web scale index." "Parallel operates a web crawler." (parallel.ai) |
| Perplexity Search API | Own index | "An index covering hundreds of billions of webpages", the same one behind Perplexity's answer engine. (blog, Sep 25, 2025) |
| Tavily | Not stated | "Tavily Search API reviews multiple sources". No index or engine is named. (help) |
| SerpApi | Live Google, plus Bing, DuckDuckGo, Yandex and others | "Scrape Google and other search engines from our fast, easy, and complete API." (serpapi.com) |
| Serper | Live Google | "Google search results in 1-2 seconds." (serper.dev) |
| cloro | Live Google, plus AI answer engines | Google's results page as JSON, loaded in logged-out sessions |
| Grounding with Bing Search | Bing, through Azure agents only | "It uses Bing to search over public web data and return relevant chunks." (Microsoft Learn) |
OpenAI web_search
|
Not stated | The docs do not name an index. They list "real-time third-party feeds" for sports, weather and finance. (docs) |
| Claude web search | Not stated by Anthropic | TechCrunch reported in March 2025 that Anthropic lists Brave Search as a subprocessor |
| Gemini Grounding with Google Search | Google Search | Grounded results use "results from Google's search engine". (terms) |
Many agents also scrape DuckDuckGo, which sells no web results API. Its help page says it has its own crawler, DuckDuckBot, and that its traditional links come "largely" from Bing (DuckDuckGo).
"Own index" is a vendor claim. None of these vendors publishes an index size that a third party has audited. The overlap test below is a way to check the claim against your own queries.
Why does almost nobody resell Bing or Google now?
The two official routes to a large index have closed.
- Bing. Microsoft's notice says "Bing Search APIs will be retired on August 11, 2025" and sends customers to Grounding with Bing Search in Azure AI Agents (Microsoft). That tool is not a resale channel: "Developers and end users don't have access to raw content returned from Grounding with Bing Search." The model gets the results and you get the model's answer with citations.
- Google. "The Custom Search JSON API is closed to new customers", and existing customers "have until January 1, 2027" to move (Google).
What is left for Google results is scraping the public page. That is what SerpApi, Serper, cloro and others do, and it has legal risk. Google sued SerpApi in December 2025 for taking "content that Google licenses from others" (Google). The court has dismissed Google's complaint twice. Google can file again by November 30, 2026 (Search Engine Roundtable).
Small wrappers sit on top of these. When a Show HN search API was asked "Which search engine(s) do you source results from?", its developer answered "Currently google". The commenter then lost interest. If the source matters to you, ask before you integrate.
How do you test what an API returns?
Run the same queries through the API and through Google's live results, and compare the top 10. We did this for three APIs on 500 queries across informational, commercial, transactional, navigational and local intent, on the same day as the Google run.
| API | Top-10 URLs also in Google's top 10 | Google's #1 domain in the API's top 10 |
|---|---|---|
| Tavily | 55.5% (URLs normalized) | 92% |
| Brave | 46.6% | 92% |
| Exa | 32.3% | Not reported |
A SERP API that loads Google's page should come close to Google's own top 10 for the same query, country and time, because it is the same page. An independent index lands lower, and that is the point of it. Brave and Tavily return mostly the same domains as Google in a different order. Exa retrieves by meaning and returns a different slice of the web: 2,396 distinct domains on the 500 queries against Google's 1,691. The Tavily study, with the method, is on cloro.dev.
The test is small enough to run yourself:
from urllib.parse import urlsplit
def norm(url: str) -> str:
u = urlsplit(url.lower())
return u.netloc.removeprefix("www.") + u.path.rstrip("/")
def top10_overlap(api_urls: list[str], google_urls: list[str]) -> float:
a = {norm(u) for u in api_urls[:10]}
g = {norm(u) for u in google_urls[:10]}
return len(a & g) / max(len(g), 1)
Normalize before you compare. Google returns some URLs with a trailing slash and other APIs without, and without normalization identical pages count as misses.
Which source should an AI agent use?
| Your agent needs | Source that fits | Trade-off |
|---|---|---|
| What Google shows, with rankings and the AI Overview | A Google SERP API | About 8 s median per search in an open benchmark, against 0.74 s for Exa |
| Speed and results that do not depend on Google | Brave, Exa or Parallel | A different ranking from what users see on Google |
| Pages a keyword search misses | Exa | The lowest overlap with Google in our test |
| One vendor for model and search | OpenAI, Claude or Gemini built-in search | You cannot choose the index, and only Gemini names it |
| No dependence on any one engine | Two sources in a fallback chain | Two sets of terms and keys |
No source is best for every agent. A rank-tracking or SEO agent needs Google's page, because that is what it measures. A research agent that runs many searches in a loop gains more from a fast independent index. The open benchmark of six APIs gives the accuracy, cost and latency of both kinds on the same 2,182 questions.
How this was checked
This article is published by cloro, which is one of the Google SERP APIs in the first table. Every other row quotes the vendor's own page, opened on November 16, 2026, except the Claude row, which cites TechCrunch because Anthropic has not published its search provider. The overlap figures are from cloro's own runs of 500 queries per API against its live Google results, published in June 2026. The Tavily figure normalizes URLs; the Brave and Exa figures come from the same harness. The latency figures come from the benchmark linked above. Questions on Hacker News about index sources and European indexes prompted this comparison.
Top comments (0)