Ask an agent to research a company and it will return a fluent profile almost immediately. Whether that profile is right depends on where its facts came from. Company research draws on five kinds of sources: official registries, the company's own website, news, regulatory filings and commercial data vendors. Each trades coverage, freshness and provenance in a different way. This guide explains those trade-offs, shows which sources answer which questions, and ends with a checklist for deciding what your agents should rely on.
Three properties to judge every source by
- Coverage. Which companies does the source include, in which countries, and down to what size? A source that is thorough on listed companies may know nothing about a ten-person supplier in a smaller market.
- Freshness. How quickly does a real-world change, such as a new director, a move or a funding round, show up? Keep apart how often the source updates and how often you pull from it.
- Provenance. Can you trace each value to where it came from? An entry in an official register, a sentence on the company's site and a vendor's modelled estimate carry very different weight, even when they show the same number.
These properties pull against each other. Broad coverage usually comes from aggregation, which dilutes provenance. The most authoritative records are often the slowest. The useful question is not which source is best in general but which property matters most for the question in front of you.
Official registries
Company registries record that a legal entity exists, along with its registered address, its officers and, depending on the country, its shareholders and filed accounts. Examples include Companies House in the United Kingdom and, in Türkiye, the trade registry and its official gazette. Aggregators such as OpenCorporates collect registry data from many jurisdictions, and the Legal Entity Identifier system overseen by GLEIF gives a global identifier to entities that take part in financial transactions.
Registries offer the strongest provenance available for legal facts, plus stable identifiers that make deduplication possible. Their limits are just as clear: they say little about what a company actually does, depth varies widely between countries, access rules and reuse terms differ by registry, and in some places records exist only as scanned documents. Use registries to anchor identity. Before enriching a company, resolve it to a registry identifier, so that records for “Acme Lojistik” and “Acme Logistics Ltd” are neither merged nor split by mistake.
Company websites and news
The company's own site
The website is the best source for what a company says it sells, to whom, how it prices and which clients it names. It is often current, because companies update their sites when they launch something. But it is self-description written to persuade, and its quieter corners go stale; an About page that still lists a former chief executive is common. An agent needs a fetch or scraping step to read sites, so check robots rules and terms, and keep the URL and retrieval date with every value taken from a page. What web scraping is covers the mechanics.
News and press releases
News reports events: acquisitions, leadership changes, contracts, layoffs, disputes. It is the freshest source for change and a weak one for baseline facts, because coverage follows what is newsworthy rather than what exists; many small private companies never appear at all. Press releases are the company speaking again through another channel. For agents, the main traps are duplication, where one announcement syndicated across many outlets gets counted as many signals, and dates, since the day an article appeared is not always the day the event happened. For alerting on events, see company and industry news alerts.
Regulatory filings
Listed companies, and in many countries larger private ones, file periodic reports: annual accounts, financial statements, ownership disclosures and notices of material events. In the United States these appear in the SEC's EDGAR system; in Türkiye, listed companies disclose through the Public Disclosure Platform, known as KAP. Filings combine strong provenance with rich detail and are the right place for financials, ownership and stated risks.
Their limits are scope, since only companies obliged to file appear; lag, since periodic reports describe the past; and format, since the figure you need often sits in a table or footnote of a long document. That makes filings a natural input for extraction that keeps the supporting passage next to each figure, so a reviewer can see exactly which line a number came from.
Commercial data vendors
Data vendors assemble company profiles from many of the sources above, plus their own research and modelling, and deliver them through a database or an API. The category ranges from firmographic and enrichment providers to private market databases such as Crunchbase or PitchBook and business information companies such as Dun & Bradstreet. The appeal is coverage and convenience: one call returns a profile with standard fields.
The questions to ask concern provenance and freshness. For each field, check whether the vendor can tell you where the value came from and when it was last verified, and whether figures such as revenue or headcount are reported or estimated. Estimates are useful for prioritising and risky as facts. Check coverage for the countries and company sizes you actually target rather than relying on general claims. Data enrichment APIs covers how to evaluate and integrate these providers, and web enrichment for sales data covers the sales use case.
Matching sources to questions
| Question | Start with | Confirm with |
|---|---|---|
| Does this legal entity exist, and who runs it? | Registry | Filings, if the company files |
| What does it sell, and to whom? | Company website | News, named client references |
| What has changed recently? | News and press releases | Registry for officer or address changes; the company site |
| How large is it, and how is it doing? | Filings where available | Vendor estimates, labelled as estimates |
| Who owns it? | Registry and ownership disclosures | Vendor data, news of transactions |
| Which companies match my criteria? | Vendor database or a search across sources | Registry and website for each shortlisted company |
The pattern is simple: let broad sources propose and authoritative sources confirm. An agent should never promote a vendor estimate or a passing news mention into a fact without saying where it came from. The last row is the start of a sourcing workflow, described in AI-assisted deal sourcing; for a diligence pipeline built on this split, see automated due diligence research.
Where GoCux fits
GoCux is neither a data vendor nor a crawler, and it does not read the open web. What it adds is control over how facts enter your work. Web material arrives only through a browsing or export step that someone authorized, and it keeps its original page reference and retrieval date, so provenance and freshness travel with each value.
Extraction works on material you supply, such as filings, registry extracts or a vendor export, and returns each value with the passage that supports it; a value it cannot find is left unknown instead of being filled in. Memory is treated as something other than a live feed: a company fact remembered from an earlier project is a reason to refresh the source, not a substitute for it, and the memory documentation explains how proposed and approved memories are kept apart.
Checklist for choosing your sources
- For each question your agents answer, you know which property matters most: coverage, freshness or provenance.
- Every company resolves to a stable identifier, such as a registry number or an LEI where relevant, before enrichment.
- Each stored value keeps its source, its URL or document reference, and its retrieval date.
- Reported and estimated figures are labelled differently.
- Coverage was checked for your real target countries and company sizes.
- Registry and vendor terms allow the storage, display and reuse you plan.
- News signals are deduplicated across syndicated copies, and event dates are kept apart from publication dates.
- Missing values stay empty and visible rather than being guessed.
Bring this workflow into your workspace.
Set the scope. Review the evidence. Approve the steps that need it.