How current is private company data — and how would you tell
Three dates in every company field. Knowing which is the source date, which is the observation date, and which is the filing period tells you exactly how current the evidence is.
Before using a private company record in a target screen, a valuation exercise or an AI-assisted search, ask what date each field reflects. Most records contain three distinct dates — source date, observation date and filing period — and vendors rarely separate them. Confusing the three is one of the most reliable ways that current-looking evidence turns out to be stale.
Three dates in every field
Source date is when the originating document was published or formally accepted. For a statutory filing it is the date Companies House accepted the document. For a company website it is the last-modified timestamp, which is rarely visible and often unreliable.
Observation date is when a researcher checked the source and recorded the value. A researcher can observe a 2022 filing in September 2026. The source date is 2022; the observation date is 2026. Both matter: the source date tells you how fresh the primary document is, while the observation date tells you how long ago someone actually looked at it.
Filing period applies to financial data specifically. A set of accounts covers a defined accounting year — for example, the year ending 31 March 2024. That 12-month window is the filing period. In the United Kingdom, private companies must file their accounts within nine months of the accounting reference date; a filing period ending March 2024 can legitimately appear in the register in January 2025, by which point the underlying financial data is already over a year old.
All three dates are different, and only the filing period typically appears on a vendor’s methodology page.
How different field types age
Not all fields age at the same rate.
Statutory identity fields — company number, incorporation date, legal form — are permanent or change only on formal events such as a re-registration or name change. Their source date is the date of the original registration or the most recent formal event.
Legal status changes only via a filing, but the practical consequence of a stale read is acting on a company that has since entered dissolution or changed form. An observation date for status matters more than for most other fields.
Trading address and current activity come from the company’s own website. They are only as current as the researcher’s last visit. A website observation from 18 months ago may reflect a company that has since moved premises, changed its trading name or ceased operations — none of which would yet appear in a registry filing.
Filed financials are retrospective by design. The figures cover a past period and arrive in a public register after a statutory delay. The financial data in any record is always at least one full accounting year behind the observation date, and often more.
Modelled fields — revenue estimated from proxy variables, headcount inferred from a LinkedIn scan, ownership inferred from a group chart — have no observation date from a primary source. They were calculated, not observed. There is no filing to check, no website to revisit, and no traceable event that would trigger an update.
Why the filing delay matters for financial screening
UK private companies must file their annual accounts within nine months of the accounting reference date. A company with a year ending 31 March 2024 has until 31 December 2024 to file. A researcher checking in January 2025 would be working with data covering trading activity from over a year earlier, even if the filing itself is recent.
For revenue-based screening, this matters. A target that appears to sit above a threshold in filed accounts may have contracted significantly in the intervening period. The record is not wrong — it accurately reflects what was filed — but treating it as a current figure rather than a trailing one changes what conclusions it supports.
For AI-assisted research, the distinction between “revenue from filed accounts, year ending 31 March 2024, filed January 2025” and “revenue from model estimate, basis and observation date unknown” is precisely the metadata the model needs to qualify its output rather than present a stale or estimated figure as a present fact.
What a well-provenanced record shows
The Basel Committee’s principles for effective risk data aggregation (BCBS 239) distinguish accuracy, completeness, integrity and timeliness as separate properties to design into a data flow, not check at the end. Timeliness in that framework means the data is available when it is needed and reflects when it was captured — not when it was last queried.
Applied to private-company data, each field in a well-provenanced record should show:
- the source type (statutory register, filed accounts, company website, public announcement);
- the source URL or document identifier;
- the source publication or filing date;
- the researcher observation date;
- the filing period, for any financial value; and
- a reason when the field is blank, not just a null.
GLEIF’s data quality framework recognizes currency — data is maintained and up to date — as a distinct quality dimension alongside accuracy and completeness. Without observation dates at the field level, currency cannot be evaluated for individual fields.
Why AI retrieval needs observation dates specifically
NIST’s generative-AI risk framework describes confabulation as producing confidently presented erroneous content. A retrieval-augmented system reduces that risk when the evidence layer carries enough metadata for the model to qualify its claims. Without observation dates, a model cannot distinguish a status field checked yesterday from one checked two years ago.
A well-formed response from an AI assistant working over provenanced company data might read: “Companies House showed this entity as active as of the observation date of January 2025; the most recent filed accounts cover the year ending March 2024.” That is a narrower and more honest claim than “this company is active and reported revenue of £X million” — and it is the difference a properly labelled record makes possible.
For the implementation of provenance metadata in an AI-ready dataset, see how to prepare private-company data for AI research. For the taxonomy of blank field reasons and how each should be encoded, see what blank fields in a private company record actually mean. For the four source types and what each one can establish, see where private company data actually comes from.
What to ask a data provider
A headline “refreshed monthly” claim describes when the database was last processed, not when specific fields were last checked against primary sources.
Three questions cut through that.
What is the observation date for each material field? A provider who can answer this field by field has thought about provenance. A row-level timestamp may reflect a bulk re-import rather than individual field verification.
How are modelled and filed financial values distinguished? If they appear in the same revenue column without labelling, the file cannot support evidence-led research. Ask for a sample record that shows the field-type metadata alongside the values.
What does a blank field mean, and when was it last checked? A blank is informative when it records a reason and an observation date. Without those, it is unknown whether the researcher looked and found nothing, or never looked at all.
SeekCompanies records each field with the source and observation date it was built from. Registered office and trading address remain separate fields; financial values appear only when a filing supports them; blank means not established from public evidence, with the reason recorded. The coverage methodology explains the evidence standard. The current access path explains how to register interest in a market.
Frequently asked questions
What is the difference between a source date and an observation date in a company record?
The source date is when the originating document — a registry filing, a company accounts document, a website page — was published or last updated. The observation date is when a researcher checked and recorded the value from that source. A researcher can observe a 2022 filing in 2026; the source date is when the filing was accepted, and the observation date is when the record was built. Both matter because the source date tells you how fresh the primary document is, while the observation date tells you how long ago someone last checked it.
How quickly can a private company record go out of date?
It depends on the field type. Company number and incorporation date are permanent. Registry status can change at any time — a solvent company can enter voluntary liquidation without immediate public notice. Trading address and activity are only as current as the website visit the researcher last made. Filed financial data covers a past accounting period, which may be 12 to 18 months before the observation date. Modelled financial fields have no update trigger at all — they were estimated once and carry no observation date.
Are modelled financial fields as reliable as filed figures?
No. A filed figure is traceable to a specific document, accounting period, currency and jurisdiction. A modelled figure was estimated without a primary source; it carries no observation date, no filing period and no audit trail. For evidence-led research — screening, valuation or AI-assisted analysis — a modelled revenue figure at a threshold is not the same claim as a filed revenue figure at the same threshold. The two should appear in separate, labelled fields.
How do you check when a private company database was last updated?
Ask for field-level provenance, not a headline refresh date. A headline "updated monthly" claim tells you when the database was last processed, not when any specific field was last checked against a primary source. Request an example record and look for a source URL or document identifier, a source publication date and a researcher observation date for each material field. If a record has only one row-level timestamp, the file does not carry field-level provenance.
Why do AI systems need observation dates in company data?
Without observation dates, a retrieval model cannot distinguish a field checked last month from one checked three years ago — it treats all supplied values as equally current. NIST's generative-AI risk framework identifies confabulation as a core risk; structured provenance metadata reduces that risk by letting the model qualify its claims with source and date rather than presenting values as timeless facts. A well-formed retrieval answer can then report when each field was observed, not just what it says.
Sources
- Annual accounts and filing requirements — Government of the United Kingdom; accessed September 25, 2026
- GLEIF Data Quality Management — Data Quality Checks — Global Legal Entity Identifier Foundation; accessed September 25, 2026
- Artificial Intelligence Risk Management Framework — Generative AI Profile — National Institute of Standards and Technology; accessed September 25, 2026
- Principles for effective risk data aggregation and risk reporting (BCBS 239) — Basel Committee on Banking Supervision; accessed September 25, 2026
Tell us which market you need.
Research is under way. No market file, price or release date is available yet. Review the planned licence, then email the market you need to ask about its research status.