SOTA2 first-party benchmark

Best Tools for Collecting Data from Websites

We collect product details, pricing, and customer quotes from company websites for the SOTA2 catalog. We compared tools on that task in two ways: reading a specific page and finding relevant pages starting from a company homepage.

Which tools we recommend

  1. AnyAPI

    You already have the page URL

    Read all 192 supplied pages and recovered 340 of 343 expected details.

  2. Context.dev

    You have only the homepage URL and need results fast

    Fastest homepage crawl at 11.75 seconds P95. AnyAPI found the same details for $1.20, compared with $2.874 for Context.dev.

  3. Exa Search + Contents

    You need cited answers to specific questions

    Exa found the original source pages and GPT-5.6 Luna extracted the cited answers in a separate test with three questions on two sites.

How we tested

Both tasks looked for the same 343 details, drawn from 82 of the 100 companies: 82 homepage titles, 75 product page titles, 59 prices with billing units, and 127 customer quotes.

1. Read a specific page
We supplied 192 specific URLs, such as a pricing or customer story page. The tool had to recover the expected details from the supplied page without following links.
2. Follow links from the company homepage
We supplied only the homepage URL for each of 100 companies. Each crawler had to discover the right pages and return the details, using up to 10 pages per site including the homepage.

We checked the text each tool returned against that list. For example, “$500 per month” counts as one pricing detail; “$500” alone doesn’t.

Tools tested AnyAPI · Context.dev · Firecrawl · Exa Contents (API docs) · TinyFish Fetch · Search1API

Scoring rules and timing

We selected the sample from our catalog work. Starting from a homepage, a price could come from another pricing page on the same host, and a quote could appear anywhere on that host. We counted repeated spellings of the same price and unit on a page once.

We ran 2 tasks at a time per tool configuration through AgentRouter for external tools. Results arriving after 90 seconds earned no matches. P95 is the time by which 95% of measured tasks ended, including failed attempts and API overhead.

Exa Contents and TinyFish Fetch took part in the main single-page test. The separate citation test also used Exa Search.

Summary scorecard (JSON). Coverage, speed, and API charges by tool, plus the leaders for each metric.

Individual runs (JSON). The URL each tool received, the pages it returned, the details it found, and the time and cost per run.

AgentRouter charged $13.35695 for page retrieval, including $0.53 for the guided Firecrawl rerun. The results tables show each tool’s charges for the two main tests.

Read a specific page 192 pages

Goal: recover homepage titles, product page titles, prices with billing units, and customer quotes from the pages we supplied. AnyAPI found 340 of 343 expected details and returned text for all 192 pages.

Scroll the table to see all columns.

Known URLs by tool and evidence type
Tool or routePages with textCompany homepage titlesProduct page titlesPrices with billing unitsCustomer quotesTotal expected details foundP95 task timeAPI charge
AnyAPI192/19282/8275/7556/59127/127340/3437.19 s$0.13495
Context.dev191/19279/8272/7554/59122/127327/3433.07 s$0.576
Firecrawl190/19273/8267/7549/59119/127308/3435.00 s$0.96
Exa190/19273/8269/7543/59114/127299/3430.94 s$0.19
TinyFish190/19264/8264/7546/59105/127279/3439.20 s$0.00
Search1API192/19237/8240/7537/5986/127200/3434.01 s$0.192

A tool can return text but miss some details. A single pricing page can contain several prices, each counted separately.

Follow links from the company homepage 100 companies

Goal: start with only the homepage URL, then find homepage and product page titles, prices with billing units, and customer quotes within 10 pages. AnyAPI and Context.dev tied for the most expected details found (234/343).

Expected details found in up to ten pages per company

Product page titles

Context.dev43/75
AnyAPI43/75
Firecrawl26/75
Search1API14/75

Prices with billing units

Context.dev40/59
AnyAPI40/59
Firecrawl3/59
Search1API13/59

Customer quotes

Context.dev72/127
AnyAPI72/127
Firecrawl58/127
Search1API32/127

Scroll the table to see all columns.

Homepage crawls by tool and evidence type
Tool or routeSites with at least one readable pageCompany homepage titlesProduct page titlesPrices with billing unitsCustomer quotesTotal expected details foundP95 task timeAPI charge
Context.dev92/10079/8243/7540/5972/127234/34311.75 s$2.874
AnyAPI92/10079/8243/7540/5972/127234/34327.30 s$1.20
Firecrawl89/10070/8226/753/5958/127157/34331.56 s$4.70
Search1API89/10027/8214/7513/5932/12786/34323.90 s$2.00

AnyAPI and Context.dev matched the same details and source URLs across all 100 sites. The AgentRouter receipts did not identify AnyAPI’s upstream provider. AnyAPI routing documentation.

Returning one readable page counted toward the sites column, even if that page contained none of the expected details.

Why the starting URL matters: Firecrawl on Momentic

We looked for seven prices with their billing units on Momentic’s pricing page, including $500 per month and $0.005 per minute. Given the pricing URL, Firecrawl found all seven in the returned text. Given only the homepage, it returned the homepage and nine documentation pages, missed the pricing page, and found none of the seven prices.

Individual runs (JSON)
Crawl setup and reruns

In a separate 12-company test, we ran the same Firecrawl crawl with instructions to prioritize products, prices, and customer evidence. Pricing hits went from 0 to 15 out of 22. Homepage-title hits went from 12 to 0 out of 12. The instructions changed which evidence surfaced under the same budget.

We resumed 191 Context.dev single-page calls and 93 homepage crawls after the initial provider quota interruption. Inputs and limits stayed the same. The scorecard combines the original and resumed tasks. P95 measures individual task times and excludes the pause between batches.

Can tools return facts with sources?

The main test checks what the returned pages contain. These three questions on two sites check whether a workflow can answer a question and cite the page that supports each fact.

Only a homepage
Search the official domain for the question. Exa Search + Contents found the original source for all three questions and the shared extractor returned 16 of 16 expected facts with supporting citations.
Already have the source page
Start by extracting from the known page and verify the cited URL. The supplied-page control returned 16 of 16 expected facts with supporting citations. In the original pilot, Context.dev passed 15 and Firecrawl Scrape passed 5.

Finding answers and their sources from a homepage

For the Roo question, Exa returned the original customer story first. Firecrawl Search found three correct, citable facts on other Momentic pages. Context.dev’s homepage crawl scored 0/4 on that question.

Scroll the table to see all columns.

Expected facts returned with supporting citations
WorkflowSalient productsMomentic pricingRoo customer storyTotalCost for 3 questions (USD)
Exa search + contents5/57/74/416/16≈ $0.06584AgentRouter receipts $0.051Model est. $0.01484
Firecrawl Search5/57/73/415/16≈ $0.20215AgentRouter receipts $0.18Model est. $0.02215
Context.dev crawl5/57/70/412/16≈ $0.08471AgentRouter receipts $0.06Model est. $0.02471
Supplied page control5/57/74/416/16≈ $0.0127Recorded page fetches $0.009Model est. $0.0037

The follow-up used the same three questions on two sites. The original Firecrawl Agent run used its own extraction and timing, so its result is listed alongside the original pilot.

How we compared search and crawl

We gave each homepage workflow the same three questions. Search was restricted to the company’s domain. Each workflow passed up to 10 pages to gpt-5.6-luna for extraction. The supplied-page control reused recorded page text.

A fact passed when the answer matched the expected detail and the cited official page supported it. A separate official page could support the same claim.

Each total covers all three questions. Retrieval charges come from AgentRouter receipts, including recorded fetches for the supplied-page control. We counted each reused crawl once. Model estimates use recorded token counts and OpenAI list rates, including cache writes. The totals exclude the infrastructure that ran the extractor.

AgentRouter charged $0.291 for the new retrieval calls. See OpenAI’s GPT-5.6 Luna rates and cache pricing.

Answers, citations, and costs (JSON)
Original pilot with Firecrawl Agent, Firecrawl Scrape, and Context.dev

The original pilot used Firecrawl Agent to check whether an agent could find and cite facts starting at a homepage. It returned 16 of 16 expected facts with supporting citations in about 46–96 seconds per task. Firecrawl Scrape and Context.dev received the source URLs.

The citation rules differ: with a supplied page, the citation had to point to that exact page. Starting from a homepage, any official company page that supported the fact could pass.

TaskFirecrawl ScrapeGiven the source pageContext.devGiven the source pageFirecrawl AgentGiven only the homepage
List Salient’s products and their uses5/5 expected facts found0/5 with passing citations5/5 expected facts found5/5 with passing citations5/5 expected facts found5/5 with passing citations
Find Momentic’s prices and billing rules5/7 expected facts found5/7 with passing citations6/7 expected facts found6/7 with passing citations7/7 expected facts found7/7 with passing citations
Find the results Roo reported4/4 expected facts found0/4 with passing citations4/4 expected facts found4/4 with passing citations4/4 expected facts found4/4 with passing citations
Total across all three tasks14/16 expected facts found5/16 with passing citations15/16 expected facts found15/16 with passing citations16/16 expected facts found16/16 with passing citations
How the citation rule affected the Salient result

For Salient, Firecrawl Scrape cited product subpages instead of the supplied product index, so those five citations did not pass. Firecrawl Agent’s product subpage citations passed under the homepage rule.

See the facts we checked in the pilot

List Salient’s products and their uses

  • Taylor handles collections
  • Marshall automates audits
  • Flyn handles insurance claims for totaled vehicles
  • Alex handles chargebacks and disputes
  • Melanie automates chargeoff decisions

Find Momentic’s prices and billing rules

  • Scale costs $500 per month and bills in advance
  • Runs pause when the balance reaches zero
  • Unused included usage does not roll over
  • Scale top-ups range from $50 to $1,000
  • Added funds do not expire
  • Hosted browser has a $0.005 per-minute meter
  • iOS simulator has a $0.1875 per-minute meter

Find the results Roo reported

  • Speaker's full name and role
  • Expected savings are 200+ hours per quarter
  • Test authoring is 4x faster
  • Flaky-test stabilization and maintenance time down about 50%

These pilot tasks cover Salient and Momentic.

Four correct facts, zero passing citations

Firecrawl Scrape returned all four facts about Roo but cited Roo’s homepage for each. The link needed to point to Momentic’s customer story.

  • AgentRouter receipts totaled $0.12 for the 6 single-page calls. The 3 Firecrawl Agent calls showed $0.00 charged in this run.
Original citation test (JSON)

Conclusion

You have the exact page URL

Read the text on a specific page, such as a pricing page or customer story. We tested 192 URLs.

API cost
TinyFish Fetch$0.00 charged in this run, with 279/343 details found. AnyAPI was the cheapest paid option at $0.13495.
Coverage
AnyAPI340/343 details found. It returned readable text for all 192 URLs.
Speed
Exa Contents0.94 seconds at P95.

If you need citations Read the source page first. With the shared model, the supplied-page control returned 16/16 facts with supporting citations. See citation results for known pages.

See single-page results

You only have the company homepage

Find relevant pages across the site. We tested 100 companies with a limit of 10 pages per site.

API cost
AnyAPI$1.20 for all 100 sites. Context.dev cost $2.874 for the same coverage.
Coverage
AnyAPI and Context.dev tiedBoth found 234/343 expected details on the same source pages.
Speed
Context.dev11.75 seconds at P95.

If you need citations Exa Search + Contents with GPT-5.6 Luna. 16/16 facts with supporting citations across three questions on two sites, at an estimated $0.066 total. See citation results from a homepage.

See homepage crawl results
How to cite this benchmark
In SOTA2’s November 15, 2026 company website benchmark, AnyAPI and Context.dev tied for the most expected details (234/343) starting from 100 company homepages with a 10-page limit per site. In a separate 192-URL test, AnyAPI found the most expected details when given the exact pages (340/343 checks).

Best Tools for Collecting Data from Websites · Summary scorecard (JSON)

Back to SOTA2 Benchmarks