How does Factually.co select and verify articles and sources?

1. What the study actually says — manual selection for a research dataset

A peer‑reviewed development study of a transformer‑based tool for automated fact‑checking of online health information makes clear that the web pages used in the research were supplied and manually selected by Factually Health Company: “A total of 538 health‑related web pages … were manually selected by Factually Health Company,” and roughly 1,000 pages were initially provided to the researchers on January 31, 2023 ^[1]. The paper explains that the company specializes in identifying reliable health‑content sites and that random sampling within disease categories was applied to reduce overrepresentation of any single disease subject ^[1].

2. What that implies about selection practices in the study

In this project Factually Health acted as an upstream curator: it sourced and filtered candidate pages before handing them to academic researchers, and used manual selection plus random sampling across disease categories to create a balanced dataset ^[1]. That combination — domain expertise guiding selection, then random sampling within categories — is standard for building datasets intended to train or evaluate automated fact‑checking models, because it reduces class imbalance and potential topical skew ^[1].

3. What the paper does not say — limits of available reporting

The published article focuses on the dataset and the model evaluation; it does not map out a comprehensive editorial policy for Factually.co (or Factually Health) beyond the dataset curation described. Available sources do not mention how Factually.co defines “reliable” sites in operational terms, what verification steps or fact‑checking protocols the company applies outside the research dataset, or whether it uses independent experts, primary‑source checks, or ongoing monitoring for accuracy ^[1].

4. How much we can generalize from a research dataset to editorial workflows

Researchers often rely on partner organizations to supply curated datasets, but a dataset‑level manual selection does not automatically reveal an organization’s routine editorial checks, dispute resolution, or source‑ranking criteria. The study documents a defensible method for building a balanced research corpus — manual curation plus stratified sampling — but does not document recurring source verification steps such as cross‑checking claims with primary literature, transparent correction policies, or disclosure of conflicts of interest ^[1].

5. Competing perspectives and implicit agendas to note

The study frames Factually Health as a specialist in identifying reliable health content and treats its supplied dataset as authoritative for model development ^[1]. That positions the company as a gatekeeper for the research; researchers accepted the curated pages rather than independently assembling a comprehensive sample. Readers should consider that partner‑provided datasets can introduce implicit biases reflecting the curator’s priorities or business model — something the paper acknowledges indirectly by describing the sampling method but does not fully audit ^[1].

6. Honest limitations and what to look for next

The sole available source documents only one collaboration and a dataset description; it therefore cannot confirm broader editorial methodologies or verification protocols used day‑to‑day by Factually.co ^[1]. To evaluate Factually.co’s general selection and verification practices, look for primary company documentation — methodology pages, editorial standards, transparency reports, or third‑party audits — none of which are present in the cited study ^[1].

7. Practical takeaway for readers who want to assess Factually.co

Use the study as evidence that Factually Health can curate balanced research datasets and that it participates in academic evaluations of automated fact‑checking ^[1]. For claims about routine article selection, source verification, or editorial transparency, request or seek out the company’s own policies and independent assessments; those elements are not documented in the currently provided reporting ^[1].