Results at a glance
| Starting point | Details found | Runs with readable content | Median time | Time (P95) | API charge | API charge per 100 details found |
|---|---|---|---|---|---|---|
| Exact page URL · Firecrawl Scrape | 308/343 (89.8%) | 190/192 | 2.3 s | 5.0 s | $0.96 | $0.31 |
| Company homepage · Firecrawl Crawl | 157/343 (45.8%) | 89/100 | 16.6 s | 31.6 s | $4.70 | $2.99 |
The homepages and exact URLs were two ways to find the same 343 details. The homepage runs were allowed up to 10 pages per site. The cost per 100 details divides the entire group's API charge, including charges for runs that came back empty, by the number of details found. Published scorecard · Individual runs
Use Scrape when you know the relevant URL. If you only have a homepage, check that the crawl reaches the product and pricing pages before relying on its output for a catalog.
| What we looked for | Exact page URL | Company homepage |
|---|---|---|
| Homepage titles | 73/82 | 70/82 |
| Product-page titles | 67/75 | 26/75 |
| Prices with billing units | 49/59 | 3/59 |
| Customer quotes | 119/127 | 58/127 |
Test 1: Reading a page you already have
Result. Customer quotes had the highest coverage at 119 of 127. Firecrawl also found all seven expected prices on Momentic's pricing page, including per-minute rates and monthly prices. Of the 35 details it missed, 33 came from pages that still returned readable content. Individual runs
| Pricing page | Prices found | Time | API charge | What Firecrawl returned |
|---|---|---|---|---|
| Momentic | 7/7 | 1.9 s | $0.005 | All seven expected amounts with their billing units. |
| Tiptap | 0/6 | 3.2 s | $0.005 | Add-on rates for Track Changes ($249/month) and Compare (from $2,969/month). None of the six expected plan rates matched. |
| Warmly | 0/3 | 1.7 s | $0.005 | An acquisition announcement and an invitation to book a conversation. The three expected annual rates were missing. |
| Vector | 0/1 | 1.8 s | $0.005 | Ad-spend thresholds appeared, but the expected $36,000 annual plan price did not. |
Those three pages accounted for all 10 price misses in the exact-page test. The Tiptap result shows a mismatch with the benchmark reference. The run does not establish why Firecrawl and the reference contained different prices. Individual runs
Setup. We sent 192 URLs to Firecrawl Scrape through AgentRouter and requested Markdown. Each answer was scored against the details expected on that page. A price needed both the amount and its billing unit. We checked whether each expected detail appeared in the returned text. Comparison methodology
Test 2: Starting from the company homepage
Result. Firecrawl found prices at two of the 15 companies with pricing checks. It matched all three prices on the reference pricing pages it returned. Across all categories, 169 of the 186 misses were checks for which Firecrawl did not return readable content from the reference page. It found all expected details for 35 of the 82 companies with checks, even though 89 of 100 homepage runs returned some readable content. Individual runs
Where the page budget went on Momentic
| Firecrawl started with | Prices found | Pages returned | Time | API charge |
|---|---|---|---|---|
| Exact pricing URL | 7/7 | Pricing page | 1.9 s | $0.005 |
| Company homepage | 0/7 | Homepage and nine documentation pages | 20.8 s | $0.05 |
On Momentic, the 10-page crawl did not return the pricing page or the Roo customer story. On InferEdge, Firecrawl also returned 10 pages, including the pricing page. It found both expected prices in 29.8 seconds for $0.05. Individual runs
Setup. We sent 100 homepages to Firecrawl Crawl through AgentRouter. It could choose up to 10 pages per site, with a 90-second observation cutoff and two tasks running at a time. The 343 checks came from 82 of those companies and were the same checks used in Test 1. We scored the text across all readable pages it returned for each company. Comparison methodology
See the 12-site guided rerun
We reran 12 of the original homepages with instructions to prioritize products, pricing, and customer stories. The guided crawl found 15 of 22 prices, up from zero, and 12 of 17 product titles, up from six. It found none of the 12 homepage titles, down from all 12, and one of 19 quotes, down from four. Momentic still came back with zero of seven prices, this time after 62.9 seconds. The guided run cost $0.53 across 12 sites versus $0.60 for the default crawl. Its P95 rose from 28.9 to 62.9 seconds. Paired rerun data
See the three-question citation pilot
On three questions across two companies, we asked Firecrawl to return structured answers with sources. For Agent, any official page that supported the fact counted. For Scrape, the citation had to point to the supplied page.
Agent started at the company homepages and returned 16 of 16 expected facts with acceptable citations in 46–96 seconds per question. Scrape returned 14 of 16 expected facts, and five met its exact-page citation rule. On Salient, Scrape returned all five product facts and cited their individual product pages instead of the supplied index. On Momentic's pricing question it missed advance billing and the rule that runs pause when the balance reaches zero. The pilot receipts total $0.075 for the three Scrape calls and show $0 charged for those three Agent calls. Structured pilot data
When to use Firecrawl for this task
If you have the product, pricing, or customer page URL, start with Scrape. In this run it was quick and recovered most of the expected text. Validate the price and billing unit before saving a record, especially when a page yields few or no rates.
If you only have the domain, give page discovery its own step. For a high-volume catalog, finding the product and pricing URLs first gives you a way to check which pages were reached before you pay to extract their contents. In the same homepage test, both Context.dev and AnyAPI found 234 of 343 details, compared with Firecrawl Crawl's 157. For a few questions that need citations, Firecrawl Agent is worth trying on the strength of its small pilot. Company data tools compared
Data and scoring
The two main tests used AgentRouter. The logs marked 291 of 292 requests as completed. One homepage request was recorded with HTTP 409 and a cutoff error. This study did not run the same cohort directly against Firecrawl, so the results describe the tested route.
- Scope and scorecard has the sample sizes, check categories, charges, and guided rerun.
- Individual runs has inputs, returned URLs, expected details, hits, timing, and charges. The public export omits the full customer quotes. The miss breakdown above groups missed checks by whether a readable version of the reference page was returned.
- Structured pilot has the three questions, returned claims, and citation checks.
Related review
Firecrawl Whole Product ReviewWe tested how Firecrawl reads websites, enriches company and professional records, and retrieves structured data through Alexandria.