How StatsMapped is made
StatsMapped is built and maintained by AI agents, working under one developer's direction. The agents write the code, ingest and check the data, and review each other's work before anything reaches this site — but every number you see here comes from an official publisher (the CSO, the Office for National Statistics, the Central Bank of Ireland, and similar bodies named on each page's own citation), never from the AI itself.
What the AI does, and does not do, here
It writes and maintains this site's own code, and works on what to build next under the developer's direction. On pages that say so explicitly, it also writes short pieces of connecting text. It never invents a statistic, and never overrides a caveat a dataset's own publisher attaches. Every change goes through an independent review step before it ships — a second AI reviewer checks the work, and a further audit pass checks what actually landed.
Where AI-written text appears
StatsMapped's own chat feature answers your questions in real time, using an AI model — it is scoped to answer only from this site's own published figures, never from the model's own general knowledge, and it says so in its own introduction.
How current is each number?
Figures reflect the data as StatsMapped holds it when you ask. When a publisher revises a figure we keep the earlier value, and we have done so since 24 September 2026; the date a figure first appeared on StatsMapped is recorded only from the day that tracking was switched on (shown on each receipt). For anything older, the vintage — what was known, and when — is unknown, and we never present today's figure as an earlier one. You cannot yet ask for the data “as of” a past date: that waits until the record is long enough to be worth trusting.
How a comparison verdict is reached
When you ask whether two statistics can fairly be compared (the comparison check), the answer is one of four verdicts, each with the reasons that decided it. A comparison starts from a pair that a person has looked at: a pair that has not been checked by a person is never better than INSUFFICIENT, however strong the numbers look.
The four verdicts
- SUPPORT
- The pair has been checked by a person, the periods line up, the units and definitions match, there are enough observations, no small-base or population-size problem, and the result survives dropping the outliers.
- QUALIFY
- The pair has been checked by a person and the numbers can be computed, but the evidence carries caveats, which are listed. A comparison across different cadences (for example quarterly against annual) can never be better than QUALIFY.
- REJECT
- A person has ruled this pair out, or the two series share no geography level to join on. The correlation is never consulted, whatever it is.
- INSUFFICIENT
- There is not enough to judge: no geography information, the pair has not been checked by a person, no aligned period, too few paired observations, or no correlation could be computed.
At present SUPPORT cannot be reached by a comparison that rests on a modelled estimate (disposable income per person, a council's income share, purchaser age by area): no sample count is published with those figures, so the small-base check cannot be run, and such a comparison is held at QUALIFY. Reaching SUPPORT also still needs a pair that a person has checked.
Coverage. 17 pairs of statistics have been checked by a person (as of 2026-10-18). The rest are not tested, which is different from being unrelated. A check by a person is a review of whether a comparison is fair to make, not a finding that the two things are related.
A verdict never says that one thing causes another. Every verdict carries the line “causal strength: not established”, and nothing in the check can change it.
The cut-offs, and what they are
The numbers below are assumptions chosen by the people who build this site. They have not been fitted or calibrated against labelled examples, so read them as rules of thumb, not as measured boundaries. They are shown here exactly as the check uses them.
- Minimum paired observations: 8. With fewer than this many paired observations nothing is judged: the answer is INSUFFICIENT.
- Strong correlation: |r| of 0.7 or more. Called "strong" in the verdict's statistical-strength line.
- Moderate correlation: |r| of 0.4 or more. Called "moderate"; below this it is called "weak", and a weak correlation is a reason to QUALIFY.
- Small-base share: more than 25% of observations. If more than this share of observations sit on small bases, data quality is marked low and the verdict is QUALIFY.
- Outlier robustness: a fall of more than 0.2 in |r|. If dropping the outliers makes the correlation fall by more than this, or flips its sign, the result is called fragile and the verdict is QUALIFY.
How a notable change is flagged
A change is flagged when a figure moves by more than a set percentage from the same measure a month, a quarter or a year earlier, depending on how often it is published. Each dataset has its own threshold. We aim for a flag in about 1 in 20 periods. Measured on our stored alerts on 15 October 2026, about 1 in 5 comparable periods flagged, so the thresholds were too sensitive; they are being tuned dataset by dataset, and the first change (house prices, month on month) is already made.
Publishers sometimes revise a figure after the event. An alert whose change a later revision removes is marked as superseded rather than deleted, so the record shows what was flagged and why it no longer stands.
The licence record behind each dataset
Every dataset answer from the API and the MCP server carries a licence record: the publisher of record, the licence, what it requires (attribution, ShareAlike, no advertising), and whether the data may be reused commercially. The record says how far we have checked it: checked means a dated check with an evidence link; registry means we hold the publisher’s licence on our register but have not recorded a dated check; conditional means the series itself carries a note that its licence is unconfirmed, so the answer to “may it be sold?” is open; unchecked means no licence is registered, and we do not assume the data may be reused or sold.
Why this page exists
Some of the text on this site is AI-written, not just AI-assisted — we'd rather say so plainly than have a reader assume otherwise.