Introducing web search monitors

Every benchmark we're in. Most of them aren't ours.

Five public benchmarks of scraping and search APIs. Four are run by other people, one is ours and open source. Every number links to its source.

  1. 01Ours · open sourcetrace.context.dev · Nov 6, 2026 sample

    87.0%of 2,527 fresh pages returned useful content. First of four.

    Every day about 2,500 never-tested pages go to Context.dev, Keenable, Firecrawl and Exa with caching off, and a judge that never sees the provider name decides whether the page came back. The lead over Keenable is statistically significant, and Context.dev has ranked first in every daily sample so far.

    $0.79
    per 1,000 useful pages
    8.91 s
    p95, the shortest slow tail
    3 of 3
    daily samples ranked first

    Read the write-uptrace.context.devSource on GitHub

    Useful content · 2,527 pages · Nov 6, 2026

    1. Context.dev87.0%
    2. Keenable84.2%
    3. Firecrawl82.3%
    4. Exa79.1%

    95% confidence intervals of ±1.3 to ±1.5 points. New sample every day at 00:15 UTC.

  2. 02Independentscrapingbench.com · Run Oct 19, 2026

    95%of scrape tasks finished. First of ten providers.

    Every API becomes the only tool of the same agent, Claude Haiku 5.5, with the same prompt and limits, and a script checks each answer against a verified key. Context.dev finished 38 of 40 tasks and cost 1.36¢ per finished task, with failures counted, against 1.67¢ for Firecrawl.

    38 of 40
    tasks finished
    1.36¢
    per finished task, failures counted
    99%
    of each page came through

    Read the write-upscrapingbench.comSource on GitHub

    Tasks done · 40 scrape tasks · 10 providers

    1. Context.dev95%
    2. Spider93%
    3. Scrapfly90%
    4. String88%
    5. Linkup85%

    Five more providers from 80% down to 15%. Confident gap over six of the nine.

  3. 03Run by a competitorwebcrawlerapi.com · Published Oct 12, 2026

    92.1%end-to-end success on 806 URLs. First of seven.

    WebCrawlerAPI ran Firecrawl's public dataset through seven APIs on default settings with caching off. We did not run it, and only heard about it when a customer sent us the link. Context.dev returned usable content on the most URLs, and turning on main content mode lifted extraction F1 from 0.176 to 0.567.

    82.7%
    content recall
    2.9 s
    median response
    0.567
    extraction F1 with mainContentOnly

    Read the write-upwebcrawlerapi.comCode on GitHub

    End-to-end success · 806 URLs · 7 tools

    1. Context.dev92.1%
    2. Firecrawl89.7%
    3. WebCrawlerAPI88.0%
    4. Crawl4AI Cloud88.0%
    5. Jina Reader86.8%

    Two more tools: Spider at 84.0% and Crawlbase at 70.0%.

  4. 04Independentopenbenchmarks.com/web-search · Measured Oct 21, 2026

    The same 300 questions, the same extraction model and the same judge, so only the search API changes. Context.dev came second to Exa fast at 99.3%, with answer recall at five of 98.3% and the lowest cost per 1,000 correct answers of the five most accurate endpoints.

    98.3%
    answer recall at 5
    $4.49
    per 1,000 correct answers
    10.9%
    for the model with no search

    Read the write-upopenbenchmarks.comHarness on GitHub

    Accuracy · 300 questions · 13 web search APIs

    1. Exa fast99.3%
    2. Context.dev98.0%
    3. Exa instant97.7%
    4. Perplexity97.3%
    5. Linkup fast96.7%

    Twelve more endpoints down to 87.7%. The site is re-running every vendor in one window.

  5. 05Independentopenbenchmarks.com/company-news · Measured Oct 21, 2026

    98.0%on company-news lookups. The best dedicated news index scored 58.3%.

    Same questions, two kinds of product: general web search APIs and indexes built only for press releases, wires and newsrooms, all under one protocol. Context.dev answered 98.0% from the open web, 39.7 points ahead of the best news index and nearly nine times the model-only baseline.

    39.7 pts
    ahead of the best news index
    9x
    the model-only baseline
    $4.49
    per 1,000 correct answers

    Read the write-upopenbenchmarks.comDataset on Hugging Face

    Accuracy · 300 questions · vs news indexes

    1. Context.dev98.0%
    2. Seltz58.3%
    3. Autobound30.0%
    4. PredictLeads20.7%
    5. Datahyena5.3%

    Model with no search at all: 10.9%. The site ranks the two bands separately.

Reading the numbers.

Every row links to the page it was read from. The request settings are in the docs.

Which of these benchmarks does Context.dev run?

Only TRACE. Scrapingbench and the two Openbenchmarks boards are independent, and the scrape quality benchmark is run by WebCrawlerAPI, a competitor. We did not take part in building or running any of the four.

Why might the numbers here differ from the benchmark sites?

Each row shows the figures as published on the date shown. TRACE draws a new sample every day, and the other sites re-run or add vendors over time, so the live pages are the source of truth. Every row links to its source.

How was Context.dev configured in these runs?

Each site publishes the exact request it sent. Scrapingbench and WebCrawlerAPI ran every provider on its default settings, with caching off where the API allows it. WebCrawlerAPI also ran a second pass with our main content option on. Openbenchmarks called the Search API with ten results and per-result highlights.

Can I reproduce the results?

Yes. TRACE, Scrapingbench, WebCrawlerAPI and Openbenchmarks all publish their harness and scoring code on GitHub, and TRACE and Openbenchmarks publish the data as well. The free tier covers a sample of your own URLs if you would rather test those.

Run your own URLs.

Bring the pages that fail today. The free tier has 1,000 credits a month, no credit card, and the same API every benchmark above called.

1,000free credits

Every month. No credit card required.

Start for freeRead the docs