Sourceframe

Workflow pattern / Research

Create a company-research workflow from a URL

Start from a single homepage and assemble a structured profile from the company's own public pages.

Intermediate / JSON records / On demand

The problem

Preparing for a call means opening fifteen tabs and copying facts into a document, inconsistently and without sources.

Who it is for

Revenue and strategy teams doing repeatable account preparation.

Teams: Sales / Partnerships / Strategy

Inputs

  • A company homepage URL
  • A profile schema
  • A list of page types to look for

Conceptual process

  1. 01Map

    Crawl a shallow scope to find about, product, pricing, careers, and blog sections.

  2. 02Classify

    Label each discovered page by type so extraction can be page-type specific.

  3. 03Extract

    Pull schema fields from the appropriate page types, with excerpts.

  4. 04Reconcile

    Where pages disagree, keep both values with their sources.

  5. 05Assemble

    Emit one record per company with per-field provenance.

Flow diagram

  1. 01Homepage URL
  2. 02Shallow crawl
  3. 03Classify pages
  4. 04Extract fields
  5. 05Reconcile
  6. 06Company profile

Example output

Illustrative output
{  "company_url": "https://example.com",  "positioning": {    "value": "Workflow automation for logistics operators",    "source_url": "https://example.com/about"  },  "products": ["Route planner", "Dispatch console"],  "pricing_model": { "value": "per seat, published", "source_url": "https://example.com/pricing" },  "hiring_signals": { "open_roles": 12, "source_url": "https://example.com/careers" },  "retrieved_at": "2026-09-14T11:20:00Z",  "not_found": ["headcount", "funding"]}

Data-quality considerations

  • A company's own pages are authoritative about intent and claims, not about outcomes.
  • Never infer headcount or revenue from page content.
  • Keep a not_found list so gaps are visible rather than implied.

Failure modes

  • Marketing language extracted as factual capability.
  • Stale pages presented as current positioning.
  • Subdomain or regional sites missed by a narrow crawl scope.

Suggested architecture

  • Shallow site crawler
  • Page-type classifier
  • Per-type extractors
  • Reconciliation step
  • Profile store

What to test first

  1. 01Run it on a company you know well and count what it got wrong.
  2. 02Check whether not_found is populated honestly.

Related reading