Firecrawl vs ScrapeGraphAI should be evaluated against the output your application needs. Getting page text, discovering pages and extracting named fields are different tasks. Normalize the relevant endpoint output, keep the source and compare missing content explicitly. A valid JSON object is not proof that its values match the original page.
Compare the operation, not the product label
| Task | What to verify |
|---|---|
| Known-page text | Required paragraphs, links, tables and code remain available |
| Structured fields | Field values are supported by visible source content |
| Site discovery | Scope, discovered URLs, duplicates and completion are documented |
| Dynamic page | The chosen endpoint returns required content under a defined ready condition |
Firecrawl’s documentation separates its extraction and crawling operations. ScrapeGraphAI’s documentation describes its own extraction services. Compare the endpoints you actually intend to use. Do not assume an open-source package, a hosted endpoint and a crawl job have interchangeable inputs, pricing or operational responsibilities.
| Current documented interface | Input/output distinction | What to test |
|---|---|---|
| Firecrawl scrape | A URL and selected formats, including Markdown; multiple formats can be requested | Required content survives the chosen formats |
| ScrapeGraphAI SmartScraper | A user_prompt with one of website_url, website_html or website_markdown; optional output_schema | Requested fields and qualifiers match the source |
| SmartScraper response | Its documented result object is not the same envelope as every SDK wrapper | Use the current SDK/endpoint response contract before normalization |
The Firecrawl scrape guide and SmartScraper guide support these interface differences. For existing source text with a targeted extraction instruction, inspect SmartScraper’s text-input route; for page acquisition plus selected output formats, inspect Firecrawl scrape. Test the task before choosing, rather than imply either service is universally more accurate.
Create an output contract before the sample
The executable assessment below accepts normalized Markdown and structured fields. It records missing required text and missing fields separately. The optional SearchCans Reader adapter demonstrates a third known-URL text operation, not a claim that Reader reproduces either vendor’s full product. Install requests for that adapter; its session and key are supplied by the caller.
def assess_output(output, required_text, required_fields):
if not isinstance(output, dict):
raise ValueError("Normalize the endpoint output to an object first")
markdown = output.get("markdown")
structured = output.get("structured")
missing_text = [v for v in required_text
if not isinstance(markdown, str) or v not in markdown]
missing_fields = [k for k in required_fields
if not isinstance(structured, dict)
or k not in structured or structured[k] is None]
return {"missing_text": missing_text, "missing_fields": missing_fields,
"text_acceptable": bool(markdown and not missing_text),
"fields_acceptable": bool(isinstance(structured, dict) and not missing_fields)}
def reader_extract(url, session, api_key):
response = session.post("https://www.searchcans.com/api/v1/url",
headers={"Authorization": "Bearer " + api_key},
json={"t": "url", "s": url, "d": 30000,
"mode": 0, "proxy": 0, "maxCache": 0}, timeout=45)
response.raise_for_status()
payload = response.json()
data = payload.get("data") if isinstance(payload, dict) else None
if not isinstance(payload, dict) or payload.get("code") != 0 or not isinstance(data, dict):
raise ValueError("Reader business error or invalid data object")
markdown = data.get("markdown")
if not isinstance(markdown, str) or not markdown.strip():
raise ValueError("Empty extraction")
return {"markdown": markdown, "structured": None, "source_url": url}
if __name__ == "__main__":
import json
fixture = {"markdown": "# Fixture\nPrice: 12 units",
"structured": {"price": 12, "currency": "fixture_units"}}
print(json.dumps({"fixture_only": True,
"assessment": assess_output(fixture, ["Price:"], ["price", "currency"])}))
Review values, not just field presence
The synthetic price record only tests field presence and a retained label. For a real comparison, check the value against the permitted source snapshot, including units, currency and qualifiers. A present but incorrect value can pass a schema check, so a reviewed reference set must sit beside the automated test.
- Retain the original authorized page or reference snapshot and the extraction timestamp.
- Compare text completeness and structured-field correctness with separate labels.
- Record empty output, blocked pages and business errors as failures rather than usable sources.
- Use the same targets and requirements before reporting any cross-provider result.
Keep the workflow boundaries visible
Use the Firecrawl alternatives map when the question is which layer to replace. the browser-state guide owns actions and session readiness. Reader API is a candidate for reading a permitted known URL. This page owns the Firecrawl and ScrapeGraphAI output decision; it does not publish an unmeasured accuracy or latency winner.
SearchCans publishes this guide. Its product examples are first-party illustrations, not an independent vendor performance study.