Exa
Best live-web search API here for agents that need ranked results, long page text, and cited answers.
Best overall in this benchmark
- you are building an agent, chatbot, research tool, or open-web RAG layer
- you need both search results and a direct answer mode, scored separately
- you care about current facts, freshness, and long extracted page text
- you need a consumer search UI instead of an API
Our take
Exa is the strongest all-around choice in this benchmark. `/search` is the most reproducible of the former top three and best on current-fact retrieval, `/answer` is the best direct-answer mode at 48/52, and both modes return long extracted page text with measured cost. The caveats are real: answer-mode ambiguity is weaker, Q43 shows a correct answer with a false citation, and the billing-dashboard allowance still needs a direct capture.
In-Depth Review
Our detailed analysis of Exa — features, performance, and real-world testing.
Feature-by-Feature Breakdown
Live Web Search with Extracted Page ContentTop-3 tied-best overall, strongest on current-fact queries.27/47▾
Feature tested: Live Web Search with Extracted Page Content
Result: Passed (27/47)
Verdict: Top-3 tied-best overall, strongest on current-fact queries.
Expected behavior: Exa's `/search` mode takes a plain query string and returns ranked live-web results with titles, URLs, extracted page text, published dates, and per-call cost metadata. The returned page text is often long enough to use directly instead of just a snippet.
Test case: Text/code file → Text/code file
Input type: Text/code file
Input used: Input artifact (Text/code file): Same fixed benchmark query set used for the run. — QUERY-SET-ground-truth.csv
Observed output: Output artifact (Text/code file): The run export shows about 38,600 median characters per result and published dates on 94% of results. — EXA-scored-run-export.csv
Input artifact: Input artifact (Text/code file): Same fixed benchmark query set used for the run. — QUERY-SET-ground-truth.csv
Output artifact: Output artifact (Text/code file): The run export shows about 38,600 median characters per result and published dates on 94% of results. — EXA-scored-run-export.csv
What changed: Text/code file transformed into Text/code file
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): This is the main search-mode ambiguity miss in the report: Exa chose SonarSource instead of the intended Perplexity entity. — EXA-per-query-output.md
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): This is the main search-mode ambiguity miss in the report: Exa chose SonarSource instead of the intended Perplexity entity. — EXA-per-query-output.md
What changed: Text prompt transformed into Text/code file
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): The run returned the correct release date and release type for a fresh version-number-collision query. — EXA-per-query-output.md
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): The run returned the correct release date and release type for a fresh version-number-collision query. — EXA-per-query-output.md
What changed: Text prompt transformed into Text/code file
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): Extraction quality averaged 2.00/2 and the needed figure was present every time. — EXA-verification-summary.md
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): Extraction quality averaged 2.00/2 and the needed figure was present every time. — EXA-verification-summary.md
What changed: Text prompt transformed into Text/code file
Why it matters / Conclusion: Use `/search` when you want reproducible retrieval and long extracted text; it is excellent for current facts, but it is not a substitute for opening the citations.
Exa's `/search` mode takes a plain query string and returns ranked live-web results with titles, URLs, extracted page text, published dates, and per-call cost metadata. The returned page text is often long enough to use directly instead of just a snippet.
Cited Answer SynthesisBest direct-answer mode, with citation checks still required.48/52▾
Feature tested: Cited Answer Synthesis
Result: Passed (48/52)
Verdict: Best direct-answer mode, with citation checks still required.
Expected behavior: Exa's `/answer` mode takes a query and synthesizes a direct cited answer rather than just returning ranked results. It was strongest on freshness and multi-source answers, though the citations still need verification.
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): Q43 is graded CORRECT, so the later citation verdict does not come from a wrong answer. — EXA-answer-grades.csv
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): Q43 is graded CORRECT, so the later citation verdict does not come from a wrong answer. — EXA-answer-grades.csv
What changed: Text prompt transformed into Text/code file
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): The citation verdict is FALSE: Exa cited aiohttp v3.11.16 release notes for a Python 3.11.16 question. — EXA-citation-verdicts.csv
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): The citation verdict is FALSE: Exa cited aiohttp v3.11.16 release notes for a Python 3.11.16 question. — EXA-citation-verdicts.csv
What changed: Text prompt transformed into Text/code file
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): Exa cleanly abstained and explained why no answer exists. — EXA-per-query-output.md
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): Exa cleanly abstained and explained why no answer exists. — EXA-per-query-output.md
What changed: Text prompt transformed into Text/code file
Test case: Text prompt → Text/code file
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Text/code file): Exa failed the abstention test by promoting Linkup marketing copy into a measured figure. — EXA-per-query-output.md
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Text/code file): Exa failed the abstention test by promoting Linkup marketing copy into a measured figure. — EXA-per-query-output.md
What changed: Text prompt transformed into Text/code file
Why it matters / Conclusion: Use `/answer` when you want a direct answer rather than a results list, but treat the citation as something to verify rather than trust blindly.
Exa's `/answer` mode takes a query and synthesizes a direct cited answer rather than just returning ranked results. It was strongest on freshness and multi-source answers, though the citations still need verification.
How it scored on the research's own criteria
The 11 evaluation dimensions from our hands-on research on Exa, each judged from recorded runs on 2 test inputs — the same verdicts the ranking page ranks on.
held up partial failed not exercised by this input
| Criterion | Verdict | What the runs showed | Per input | Proof |
|---|---|---|---|---|
| Ambiguity handling | Weak2/5 | Search does okay, but the answer mode is unreliable when names collide and can commit to the wrong entity. That is a clear weakness on ambiguous queries. | — | |
| Answer quality (answer APIs) | Strong5/5 | The answers themselves are excellent across the benchmark, including freshness and multi-source questions, with only a few misses. That is top-tier answer quality. | open proof ↗ | |
| Citation accuracy (answer APIs) | Strong4/5 | Most citations really do back the claims, but the Q43 miss shows it is not perfect. That is strong performance with a small but real grounding gap. | open proof ↗ | |
| Extraction quality | Strong5/5 | It returns rich, directly usable page text instead of thin snippets, so the extracted content is strong enough to feed straight into downstream use. | open proof ↗ | |
| Freshness | Strong5/5 | It consistently handled recent-release questions and did not drift on the freshness probes, so this is top-tier freshness behavior. | open proof ↗ | |
| Long-tail coverage | Mixed3/5 | It can handle niche topics, but the drop on technical and content-depth queries is real enough that long-tail coverage is only middling overall. | open proof ↗ | |
| No-answer behaviour | Strong4/5 | It usually refuses cleanly and even gives good reasons for abstaining, but one hallucinated answer shows the guard is not fully dependable. | open proof ↗ | |
| Relevance @ top-k | Strong4/5 | Results usually show up near the top, and search is especially strong, but the answer mode is noticeably weaker than the search mode. That makes this a strong performer rather than a flawless one. | open proof ↗ | |
| Cost per 1k queries | Strong5/5 | Pricing is explicit, measured, and easy to normalize, with no extra extraction surcharge hidden in the run. That makes cost handling strong and predictable. | open proof ↗ | |
| p50 / p95 latency | Strong4/5 | Latency is consistent enough to be usable, but it is still in the multi-second range. Good tail behavior keeps it from scoring lower. | open proof ↗ | |
| Stability | Mixed3/5 | Search is fairly reproducible, but answer mode moves too much from one run to the next to call the system stable overall. | open proof ↗ |
Verdicts come verbatim from the study's recorded observations, never re-derived at render; a criterion with no recorded run shows Not exercised — this section cannot invent a score.
Measured endpoint rates
Exa bills by credits in its docs; the benchmark measured the public per-call rates directly.
Exa's docs say $20 on signup and $10/month in credits with no payment method required. Third-party trackers reported a free-tier change from 1,000 to 20,000 requests/month on 2026-07-14, but the August billing-dashboard allowance was still not captured here.
Featured in Rankings
Independent rankings where Exa was tested and rated.
Banner Preview
How the embed badge will look on your site
Embed HTML
Copy this code to your website source
Quick Integration Guide
- 1Copy the HTML code block above.
- 2Paste it into your site's HTML or CMS editor.
- 3Banner appears instantly on your page.
- 4Links back to your tool profile here.
Similar Tools
Discover more AI tools like Exa to enhance your workflow.
Comments (0)
Need a custom AI solution for this use case?
If you are looking to build a custom web search, cited answer, or research agent system for your business or internal workflow, email us at contact@futuresmart.ai.
Found something inaccurate or missing? We try to keep our AI research accurate and useful. If you found outdated information, an issue, or have a suggestion, email us at collaborate@aidemos.com.