A panel measures a recruited sample of buyers and weights it up to represent a population. A census measures everything within a defined boundary and claims nothing outside it.
Almost every argument about whose market data is right traces back to this distinction, and it is usually left unstated by both sides.
They fail in opposite directions#
| Panel | Census | |
|---|---|---|
| Measures | A sample, weighted to a population | Everything inside a boundary |
| Claims | The whole market | Only what it observed |
| Error type | Sampling error | Coverage gap |
| Error quantifiable | Yes, with the sample design | Rarely, from inside the data |
| Error grows when | Drilling into small brands or narrow segments | The market moves outside the boundary |
| Can link purchases to a buyer | Yes | Usually no |
| Small-brand reliability | Poor | Full |
The crucial asymmetry: a panel's error is visible and calculable; a census's error is invisible from inside. A panel will tell you its confidence interval. A census cannot tell you what it did not see — that requires an external source.
This makes census data feel more precise than it is. Every figure is exact, and none of them are qualified.
The resolution floor#
A panel represents a brand through however many panellists happened to buy it. For a category leader that is a large number and the estimate is dependable. For a challenger it may be a handful, and an estimate built on a handful carries an error band wide enough to swallow the whole figure.
Census-style observed data has no sample to fall below, so the same challenger is fully visible.
The practical consequence is significant and frequently missed: a category can look stable in panel data while restructuring underneath it, because the brands doing the restructuring sit below the panel's resolution. Anyone whose actual question is about emerging competition should know which kind of data they are reading before drawing a conclusion. See E-Commerce Transaction Data.
Where each is genuinely better#
Panel is better for:
- Total market size including channels no observer covers, especially offline
- Anything about people — penetration, repeat rate, basket composition, switching. See Category Penetration and Repurchase
- Demographics of the buyer, which observed data almost never carries
- Established categories with large, stable brands
Census is better for:
- Product-level and variant-level detail
- Actual selling prices, which is what Price Band Analysis requires
- Emerging brands and long-tail assortment
- Rapid movement, since it is observed continuously rather than fielded
- Anything where the question is confined to the observed channels anyway
"Which is more accurate" is not a question#
It has no answer until the decision is named, and vendors on both sides are content to leave it that way.
For a national category total including offline trade, a panel is closer to the truth. For a specific product's online price last week, an observed dataset counted it while the panel would be estimating from a handful of reported purchases. Both statements are true simultaneously.
The question that does have an answer: which source is more accurate for the specific figure I am about to rely on?
Combining them#
Usually correct, but it requires a documented rule rather than an average, because the two are measured in incompatible units — a weighted population estimate against a direct count within a boundary.
Defensible patterns:
- Split by question. Census for granularity, price and emerging brands; panel for market totals and buyer behaviour. Each figure attributed to its source.
- Calibrate one against the other. Use the panel's total as a denominator for census-observed activity, stating the assumption.
- Publish both where they disagree, with provenance, rather than reconciling silently.
The failure mode is blending them into a single number that carries neither method's error characteristics and cannot be defended when challenged. Data Provenance and Due Diligence covers why the reconciliation rule has to be written down.
Questions to ask a vendor#
- Is this figure sampled or observed?
- If sampled: what is the sample size behind this specific cut, and what is the error band?
- If observed: what is the collection boundary, and what proportion of the category do you believe sits outside it?
- Where you combine the two, what is the reconciliation rule?
- At what level of granularity does this figure stop being reliable?
That last question is the one that most reliably separates vendors who understand their own data from vendors who do not.
Where to look next#
For observed data in depth, see E-Commerce Transaction Data. For the full source taxonomy, see Market Intelligence Data Sources. For testing either kind against reality, see Data Accuracy Benchmarking. For boundary and gap analysis, see Data Coverage and Completeness.
Common questions#
What is the difference between panel and census data?#
A panel measures a recruited sample of buyers and weights that sample up to represent a population, so it makes statements about a whole market on the basis of a part. A census measures everything within a defined boundary and makes no claim beyond it. The difference is where the error lives. A panel's error is sampling error: it grows as you drill into smaller brands and narrower segments, and it is quantifiable. A census's error is a coverage gap: it is precise about what it sees and completely silent about what falls outside, and the gap is usually not quantifiable from inside the data.
Which one is more accurate?#
Neither, in the abstract — they are accurate about different things, and the question is only answerable once you name the decision. For a category total including offline sales, a panel is generally closer to the truth because it is designed to represent the whole market. For a specific product's online price and position, a census-style observed dataset is far more accurate, because it counted rather than inferred. Asking which is more accurate without naming the question is the fastest way to buy the wrong dataset, and vendors on both sides encourage it.
Why do small brands appear differently in each?#
Because a panel has a resolution floor and a census does not. A brand with a small share of a category may be represented by a handful of panellists, and an estimate built on a handful carries an error band wide enough to swallow the figure entirely — so small brands are unreliable in panel data, not absent but not trustworthy. In observed census-style data there is no sample to fall below, so the same brand is fully visible. This is why a category can look stable in panel data while restructuring underneath it, and why anyone researching emerging competition should be careful which source they are reading.
Can the two be combined?#
Yes, and it is usually the right answer, but only with an explicit rule for how. They are measured in incompatible units — a panel produces a weighted population estimate, a census produces a direct count within a boundary — so combining them requires a documented reconciliation, not an average. The common and defensible pattern is to use the census source for granularity and price, the panel for market totals and buyer behaviour, and to publish both figures with their provenance where they disagree. Quietly blending them into one number that means neither is the failure mode.