To ground LLM answers with citations, use a search API that returns text the model can answer from and a URL for each part of that text. For the most correct answers, use a Google SERP API with the AI Overview: on 100 benchmark questions, the AI Overview took grounded answers from 86% to 99% correct, and the API returns it with its own source links. When the model must quote the page text, or the user waits for the answer, use Exa or Tavily. When you want a finished answer with citations, use Exa's /answer endpoint, Perplexity Sonar or Brave's Answers plan.
- What you can cite: every search API below returns a URL for each result. They differ in the text next to the URL: a snippet, an excerpt, the full page text or a written answer.
- Accuracy: on 824 multi-hop questions, the two Google SERP APIs answered 80.2% and 77.4% from one search. Parallel answered 65.3%, Exa 64.0%, Brave 53.9% and Tavily 48.4%.
- Speed: Exa had the fastest median search, 0.74 seconds. The two Google SERP APIs were the slowest, at about 8 seconds.
What does each search API hand back for citation?
A citation is a claim and the page that supports it. So for each passage the model needs the text that holds the fact and the URL of its page.
| API and endpoint | Text the model reads | Full page text in the same call | Written answer with sources |
|---|---|---|---|
| Google SERP API (cloro, SerpApi and others) | The AI Overview and one snippet per result | No. You fetch the pages yourself | The AI Overview, with Google's source links |
Exa /search
|
Highlights or full text | Yes, with text or highlights
|
/answer returns answer and a citations array |
Tavily /search
|
A short content description |
Yes, with include_raw_content
|
include_answer adds an answer string. The API spec does not link it to results |
| Parallel Search API | Excerpts of each page | Excerpts. Full pages come from its Extract API | Not in the Search API |
| Brave Search API | A description and extra snippets | No | The Answers plan returns answers with citations |
| Perplexity Search API | Page text in each result's snippet
|
Yes, in snippet
|
Sonar returns citations and search_results arrays |
Every row also returns the URL of each result. The Google SERP API row has one more source of citations: the AI Overview is a written answer, and Google attached the pages it used. This is a trimmed AI Overview from a live Google results call for "how tall is the eiffel tower" on November 14, 2026:
"aioverview": {
"text": "The Eiffel Tower is 330 meters (1,083 feet) tall from its base to the tip of its antennas. ...",
"sources": [
{"position": 1, "url": "https://www.toureiffel.paris/en/the-monument/key-figures"},
{"position": 2, "url": "https://en.wikipedia.org/wiki/Eiffel_Tower"}
]
}
The same response can return the AI Overview as markdown, with a link after each claim. Then each sentence has its own source.
Which part of a search result makes the grounded answer correct?
A public benchmark tested this. It kept cloro's saved Google responses for 100 SimpleQA Verified questions and changed only what the reader model could see. The search results stayed the same in every row.
| What the model read | Correct answers | Tokens per question |
|---|---|---|
| AI Overview, then 10 results with title, URL and snippet | 99% | 843 |
| AI Overview and snippets only | 99% | 528 |
| 10 results without the AI Overview | 86% | 747 |
| 10 results without the AI Overview, plus the text of the top result pages | 91% | 6,239 |
The AI Overview carries most of the accuracy. Without it, the model lost 13 of 100 questions and gained none. The full page text did not close that gap, and it used about 7 times as many tokens.
URLs did not change the score. Without them it stayed at 99%. Keep them anyway, because they are the citations.
The test scored the answers, and it did not score the links. The reader got the AI Overview text without its source links. So the 99% shows that the answer is in the AI Overview. It does not show that each link points to the best page.
Two limits apply. In a separate test of 500 queries with mixed intent, Google showed an AI Overview on 64% of them. For the other queries, the 86% row is the closer guide. Also, when Google showed an AI Overview, its cited domains matched the organic top 10 only 51% of the time. So put the AI Overview's sources in your citation list, next to the organic links.
When are Exa, Tavily, Parallel or Perplexity the better pick?
- The model must quote the page. A Google SERP API returns snippets, and you fetch the pages yourself. In the benchmark, only 56% of the result pages gave text: robots.txt blocked 22% and 17% returned an error. Exa, Tavily, Parallel and the Perplexity Search API return page text or excerpts in the same call.
- The user waits for the answer. The median search took 0.74 seconds on Exa, 1.21 on Brave, 2.00 on Tavily and 2.99 on Parallel. The two Google SERP APIs took 7.96 and 8.20 seconds.
- The questions are short and factual. On 1,000 SimpleQA Verified questions, Parallel (97.8%), Exa (97.3%) and the top Google SERP API (96.9%) were within the margin of error of each other. Then price and speed decide.
-
You want a finished answer in one call. Exa
/answer, Perplexity Sonar and Brave Answers return an answer and its sources. Their model picks the sources and writes the claims.
The Google SERP API is the better pick for multi-hop questions. On FRAMES, Parallel, Exa, Brave and Tavily replied "not found" on 151 to 295 of the 824 questions, because their 10 results did not hold every fact.
How do you make the model cite its sources?
- Put the AI Overview first, then one snippet per result. With the AI Overview, five results scored the same as 10.
- Give every source a number, and keep the URL list in your code.
- Tell the model to answer from the sources only, to put the source number after each claim, and to reply "NOT FOUND" when the sources do not hold the answer.
- Check that every number in the answer maps to a source. Reject an answer with no citations.
This function does steps 1 and 2. The field names are cloro's, and other Google SERP APIs return the same parts under other names. A result link can be empty, so the function skips it.
def build_context(result, max_results=5):
"""Number every source so the model can cite it as [n]."""
blocks, sources = [], []
aio = result.get("aioverview") or {}
if aio.get("text"):
ids = []
for s in aio.get("sources") or []:
sources.append(s["url"])
ids.append(f"[{len(sources)}]")
blocks.append(f"AI Overview, sources {' '.join(ids)}:\n{aio['text']}")
for r in (result.get("organicResults") or [])[:max_results]:
if r.get("link") and r.get("snippet"):
sources.append(r["link"])
blocks.append(f"[{len(sources)}] {r['snippet']}")
return "\n\n".join(blocks), sources
What does each search API cost per 1,000 searches?
| API | What the price buys | Price per 1,000 |
|---|---|---|
| cloro (Google SERP API) | Organic results and the AI Overview | $2.00, or $2.80 as a sync request |
| SerpApi (Google SERP API) | One search on the Searcher plan, $725 for 100,000 | $7.25 |
| Exa | Up to 10 results / an answer with citations | $7 / $5. Page text is $1 per 1,000 pages for each content type |
| Tavily | A basic search, pay as you go | $8 |
| Parallel Search API | 10 results with excerpts | $5, or $1 in fast mode |
| Brave | Search results / an answer with citations | $5 / $4 plus $5 per million tokens |
| Perplexity Search API | Ranked results with page text | $5, or $1 in fast mode |
The cloro row uses the Hobby plan, $100 a month for 250,000 credits, or $0.40 per 1,000 credits. A Google search is 3 credits, the AI Overview adds 2 and a sync request adds 2 more. In the benchmark, SerpApi needed a second search on 97% of SimpleQA questions to load the AI Overview. Perplexity Sonar is priced per token and is not in this table.
How this was checked
This comparison is published by cloro, which sells one of the Google SERP APIs in it. The accuracy and latency figures come from an open benchmark of six web search APIs on 2,182 questions in November 2026: one search per question, 10 results, and the same reader model and grader for every API. The harness and every graded answer are in the benchmark repo. The 64% and 51% AI Overview figures come from a 500-query test against Google's live results on July 8, 2026. The response fields and prices of the other APIs come from each vendor's own docs and pricing page, read on November 14, 2026.
The full comparison of web search APIs, with how close each one gets to Google's results, is in Best Web Search API for AI Agents.
Top comments (0)