How to Find Reliable Data Online: Sources, Methodology and Red Flags

“Data” sounds objective. In practice, most numbers on the internet are either someone’s software’s guess, someone’s marketing slide, or a good source reported badly. The difference between a dashboard that decides well and one that decides badly is rarely the chart library — it is the provenance discipline behind the numbers.

This is a field guide to vetting data online: where to start, what to demand from a source, and the seven red flags that should make you walk away. It is the same checklist we apply to every dataset in the AxioStats matrix.

Prefer primaries over resellers

The internet has a data food chain. At the top sit primaries: the institution legally obliged to produce the statistic — a central bank, a statistical office, a platform’s official API. Below them sit resellers: news sites, aggregators, dashboards and AI assistants that copy numbers downstream, usually with lag and sometimes with errors.

The rule of thumb: every time data passes through another human or another script, it can be corrupted. Ask yourself: did this number reach me directly from the source, or through a chain? For the US inflation rate, the chain is BLS → FRED → (someone’s chart) → (someone’s tweet) → (someone’s dashboard). Prefer links with fewer hops, and always link to the primary.

Demand methodology — in writing

A number without methodology is not data, it is an assertion. A serious source will tell you, in plain language:

1. What is measured (which basket, which universe, which population).
2. When and how often (daily, monthly, quarterly; release calendar).
3. How it is calculated (formulas, base years, seasonal adjustment).
4. Where the raw input comes from (surveys? administrative records? web scraping?).

Two examples from the matrix: our US Inflation Rate page documents that the YoY rate is computed as (CPIₜ − CPIₜ₋₁₂) ⁄ CPIₜ₋₁₂ × 100 from the raw BLS index — not downloaded as a ready-made series. Our LLM benchmarks page states explicitly that scores are vendor-reported and not independently re-measured. Both statements are methodology; one of them changes how you should use the data.

Check the units and the seasonality

The most common data error is not a wrong source — it is a misread unit:

• Index (100 = base year) vs percentage vs percentage point vs rate. 3.3 means three different things in those four systems.
• Level vs change: CPI at 300 points and inflation at 3.3% are the same dataset read two ways.
• Seasonally adjusted vs not: comparing July’s unadjusted number with June’s adjusted one will invent a spike. Adjusted series exist precisely because calendar effects (holidays, weather, school years) are real.
• Nominal vs real: GDP in current dollars vs chained 2017 dollars can differ by more than the growth rate you are trying to measure.

The fix is mechanical: for every series, read the metadata fields before the values. FRED’s units and seasonal_adjustment fields exist because professionals got burned enough times to print the warnings.

Keep provenance with the data

Reproducibility is the cheapest quality control there is. From day one:

• Store source name, source URL, and retrieval date alongside every table you build.
• Version your snapshots (our data layer commits raw JSON snapshots per refresh — a build can never fetch, so it can never silently change).
• When you share a number, include the acquisition chain in a footnote: “BLS CPIAUCSL via FRED, retrieved 2026-09-01, YoY computed by us.”

Six months later, “where did this number come from?” is a question you will be glad you can answer in one sentence.

The seven red flags

When evaluating any dataset, walk away if you see:

1. No source at all — “industry data” with nobody to cite is fiction with a chart.
2. Source is a reseller cited as primary — “according to The Economist chart of FRED data” is not FRED data.
3. Screenshot-only numbers — if the publisher fears you re-verifying, treat the numbers as unverified.
4. Too-precise marketing numbers — “market share 23.47%” from a company that sells market research.
5. Methodological silence — no units, no date, no formula, no definition of what is counted.
6. Cherry-picked windows — a “record” that starts conveniently after a peak, or ends conveniently before a trough.
7. No update cadence — data that never changes while the world changes is data you cannot trust to decide now.

The matrix method

This is the organizing idea behind AxioStats, made explicit: separate what the data is about (domain) from how it was produced (engine). When you know that a number is hard data (official time series from primaries), telemetry (measured by automated tests), sentiment (derived from text) or vox (crowdsourced votes), you immediately know what you can and cannot do with it — and the methodology lives on every page for the cases you need to verify.

Browse the verified sources:

• All datasets with source links and methodology

FAQ

What is the most reliable free data source? For economics, the primaries — BLS, FRED, Eurostat, World Bank, IMF — are the gold standard. For platform data (npm, Wikipedia, GitHub), the platform’s own public APIs beat any mirror.

How do I know if a statistic is fabricated? You cannot always tell from the number alone; you can only judge the chain. If the source, date, methodology and raw inputs are all public and auditable, fabrication is far harder than with a screenshot.

Is AI-generated data reliable? Generative models are a compression of training data, not a measurement instrument — they will confidently reproduce numbers with hallucinated precision. Never cite an assistant’s output as a source; use it to find primaries, then cite the primary.