Valyu icon
developer-tools

Valyu

Returns usable web text for agents, but the citation layer is too error-prone to trust.

Visit Valyu
Search APIAnswer API4 false citations25k-char cap
TL;DR — our verdictUpdated August 2026

Strong retrieval, unsafe citations

Where it wins
  • You need an API that returns large extracted web content your agent can read without a second crawl.
  • You are willing to independently verify citations and source domains before trusting an answer.
  • You care more about usable content payload than low-latency answer-mode convenience.
Main limitation
  • Your product depends on citation honesty without manual checks.

Our take

Valyu's /search mode returns long, usable page text and can support answer-making from the retrieved content, but /answer repeatedly cites the wrong entity and posts the benchmark's worst citation record. In this run it also measured far above the quoted web-source rate and streamed slowly enough that it is hard to justify as a default grounding layer for an agent.

In-Depth Review

Our detailed analysis of Valyu — features, performance, and real-world testing.

AD
AI Demos Team
Expert Reviewer
Verified Review

Feature-by-Feature Breakdown

Ranked Web Retrieval with Extracted Page Text
▾
Test Summary
Feature tested: Ranked Web Retrieval with Extracted Page Text
Result: Partial

Feature tested: Ranked Web Retrieval with Extracted Page Text

Result: Partial

Expected behavior: Valyu’s /search mode returns ranked web results plus large extracted page text, rather than short snippets. In the benchmark it was exercised on current-fact, niche-technical, multi-source, content-depth, and freshness probes, with a median payload of about 24,912 characters.

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Why it matters / Conclusion: Useful when you need a search API that hands back a lot of readable web text; the retrieval is only middling, but the extraction quality is good enough that the model can work from the payload.

Valyu’s /search mode returns ranked web results plus large extracted page text, rather than short snippets. In the benchmark it was exercised on current-fact, niche-technical, multi-source, content-depth, and freshness probes, with a median payload of about 24,912 characters.

INPUT
INPUT — current-fact query from the benchmark: Apollo GraphOS pricing (Q23).
↓
OUTPUT
Search mode returned relevant Apollo-related material with enough extracted text to support an answer, showing that the content layer was not the limiting factor in this run.
INPUT
INPUT — content-depth query from the benchmark, where the answer sits deep in a long page.
↓
OUTPUT
Search mode returned full extracted page text rather than a short snippet, which is consistent with the reported strong extraction score and the ~24,912-character median payload.
INPUT
INPUT — freshness probe from the benchmark.
↓
OUTPUT
Search mode surfaced recent material on some freshness probes, but the run still showed only 33% top-3 freshness performance for /search, so freshness was mixed rather than dependable.
Bottom Line
Useful when you need a search API that hands back a lot of readable web text; the retrieval is only middling, but the extraction quality is good enough that the model can work from the payload.
Cited Answer Synthesis
▾
Test Summary
Feature tested: Cited Answer Synthesis
Result: Failed

Feature tested: Cited Answer Synthesis

Result: Failed

Expected behavior: Valyu’s /answer mode streams a synthesized answer with citations. In the benchmark it answered question sets directly and was evaluated on correctness and citation behavior, including wrong-entity citations and an abstention that still pointed to sources.

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Test case: Text prompt → Text prompt

Input type: Text prompt

Input used: Input artifact (Text prompt): Input

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Text prompt): Input

Output artifact: Output artifact (Text prompt): Output

What changed: Text prompt transformed into Text prompt

Why it matters / Conclusion: Do not trust the citations blindly; if you use Valyu, verify the domains and the claim-to-source match yourself.

Valyu’s /answer mode streams a synthesized answer with citations. In the benchmark it answered question sets directly and was evaluated on correctness and citation behavior, including wrong-entity citations and an abstention that still pointed to sources.

INPUT
INPUT — current-fact query: Apollo GraphOS pricing (Q23).
↓
OUTPUT
The answer cited Apollo.io pricing instead of Apollo GraphOS pricing, so it looked sourced while pointing at the wrong company.
INPUT
INPUT — citation-honesty check: Perplexity Sonar / third-party SonarCloud docs (Q26).
↓
OUTPUT
The answer cited third-party SonarCloud docs for Perplexity Sonar, another wrong-entity attribution rather than a supporting source.
INPUT
INPUT — question about the size of Linkup's index (Q52).
↓
OUTPUT
The answer was graded as a correct abstention, but the citation still pointed to worldwidewebsize.com and did not support the claim; that is a citation failure even though answer-level scoring passed.
INPUT
INPUT — citation-honesty check: unrelated dental-company press release (Q44).
↓
OUTPUT
The answer pointed to a press release for an unrelated dental company, reinforcing the benchmark's finding that Valyu often retrieves the right material but attributes it to the wrong entity.
Bottom Line
Do not trust the citations blindly; if you use Valyu, verify the domains and the claim-to-source match yourself.

How it scored on the research's own criteria

The 11 evaluation dimensions from our hands-on research on Valyu, each judged from recorded runs on 4 test inputs — the same verdicts the ranking page ranks on.

held up  partial  failed  not exercised by this input

CriterionVerdictWhat the runs showedPer inputProof
Ambiguity handlingWeak1/5When a query can point to multiple entities, it can confidently lock onto the wrong one instead of pausing or disambiguating, which is a core failure.—
Answer quality (answer APIs)Weak1/5Even when it can do the arithmetic, it may combine figures that are not comparable, so the final answer can be decisively wrong despite looking numeric and careful.open proof ↗
Citation accuracy (answer APIs)Weak1/5It does not just make citation mistakes; it can attach a real, plausible page to the wrong claim, which makes the answer look sourced when it is not.open proof ↗
Extraction qualityMixed3/5The returned text is usually usable and substantial, but the hard 25k ceiling and missing publication dates keep it from being a clean, fully faithful page feed.open proof ↗
FreshnessWeak1/5Exact version lookups are where freshness matters most, and here it missed the specific release even when the broader topic was easy to name, so the recent-content performance is poor.open proof ↗
Long-tail coverageWeak2/5It can reach niche material sometimes, but a one-third top-3 hit rate is still thin for technical or long-tail queries, so this lands in the weak range.open proof ↗
No-answer behaviourMixed3/5It usually knows to step back on impossible questions, but the partial-answer case shows it can still drift into unsupported detail after a correct abstention.open proof ↗
Relevance @ top-kWeak2/5It usually finds a relevant source somewhere in the top results, but top-1 is too weak and top-3 is only around 30%, so this is below average rather than strong.open proof ↗
Cost per 1k queriesWeak1/5Measured spend is an order of magnitude above the published headline rate, so this is not just a little expensive; it is materially off-budget.open proof ↗
p50 / p95 latencyWeak1/5Both modes are slow, and the answer mode is especially heavy at the tail, so agents would feel the latency even more than the median suggests.open proof ↗
StabilityStrong5/5Ranking held up well across reruns: the overlap stayed very high and the changes stayed inside the reported noise band, so this is genuinely reproducible.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.

✓ Use This If
●You need an API that returns large extracted web content your agent can read without a second crawl.
●You are willing to independently verify citations and source domains before trusting an answer.
●You care more about usable content payload than low-latency answer-mode convenience.
✕ Skip This If
●Your product depends on citation honesty without manual checks.
●You need low-latency or predictable cost at scale.
●You need strong freshness or abstention performance in answer mode.
developer-toolssearch-enginetextOther
It returns extracted content, not just snippets. In this run the median payload was about 24,912 characters per result, and output was hard-capped at 25,000 characters.
Poorly. Valyu produced 4 of the 5 false citations in the benchmark, including wrong-entity citations such as Apollo GraphOS being cited with Apollo.io pricing and Perplexity Sonar being cited with third-party SonarCloud docs.
Not automatically. Q52 was graded as a correct abstention at the answer level, but the citation still pointed to worldwidewebsize.com and did not support the claim, so citation verification is still necessary.
valyu/search measured p50 at 6,311 ms and p95 at 12,098 ms. valyu/answer measured p50 at 14,040 ms and p95 at 24,015 ms, making it the slowest completing mode in the benchmark.
The docs quoted roughly $1.50 per 1,000 web-source calls, but the measured all-in costs from total deduction were $14.54 per 1,000 for /search and $31.81 per 1,000 for /answer.
It did well on current-fact answers, where it was 10/10 correct, and its extraction score was top-group. It was much weaker on freshness and abstention, with 2/6 fresh answers correct and 3/5 unanswerable queries handled cleanly.

Banner Preview

How the embed badge will look on your site

Valyu featured on AI Demos

Embed HTML

Copy this code to your website source

<a target="_blank" href="https://aidemos.com/tools/valyu?utm_source=valyu_embed" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> <img src="https://aidemos-website-images.s3.amazonaws.com/featured.png" alt="Valyu | Featured on AI Demos" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> </a>

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

Similar Tools

Discover more AI tools like Valyu to enhance your workflow.

🤖
Exa
Best live-web search API here for agents that need ranked results, long page text, and cited answers.
AI Tool
🤖
Jina
Rich web search results for agents, but this free-tier run was incomplete and not comparable.
AI Tool
🤖
Linkup
Standard mode is the usable tier: strong extraction, weak ranking; deep is slower, pricier, and weaker.
AI Tool
🤖
Olostep
A search-only baseline with mid-pack retrieval, weak extraction, and no usable answer mode.
AI Tool
🤖
OpenAI web_search
A citation-honest control arm for live-web answering, but not a reliable retrieval layer for fresh content.
AI Tool
🤖
SerpAPI
Fast, zero-error live web search for agents, with stable mid-pack retrieval and snippet outputs.
AI Tool
🤖
Serper
Fast, cheap raw web search for agents — strong on ambiguity, but too snippet-thin to replace a scrape.
AI Tool
🤖
Tavily
A web search API that returns full-page text well, but usually buries the best result.
AI Tool
🤖
You.com
Best-in-benchmark top-1 web retrieval for agent queries, with full-page text and published dates — but at a measured high all-in cost.
AI Tool

Comments (0)

Please Log in to join the discussion.

Built by FutureSmart AI — the team behind AI Demos

Need a custom AI solution for this use case?

If you are looking to build a custom web text extraction, citation verification, or retrieval pipeline for your business or internal workflow, email us at contact@futuresmart.ai.

Get a custom build

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.

Back to Top