What 'Accurate' Really Means for an Impressions Number: MAPE and Confidence Intervals

When a competitor stamps "99.5% accurate" on a measurement product, what does that actually mean — and should you believe it? Understanding MAPE and confidence intervals is how a small OOH operator tells a real accuracy claim from a marketing badge, and how you talk about the reliability of your own numbers without over-promising.

This is shared engine-trust material. The deep methodology behind how the counting engine is validated — the benchmark accuracy, the test fixtures, the error math — lives once on the StreetProof methodology hub so it isn't duplicated across brands. This page frames it for impressions and links you there; the canonical version is on StreetProof. It supports the AdWitness OOH audience measurement guide and the practical impressions audit.

Key takeaways

• "Accurate" is not one number. MAPE describes typical error; a confidence interval describes the range around a specific estimate.
• A single "98% accurate" badge hides how accuracy changes with light, density, angle, and sample length.
• A confidence interval reflects count noise and sample size — a wider band means less certainty, not a worse product.
• Even a tight interval needs a representativeness label, because a short sample can look precise and still misjudge a normal week.

Why a single accuracy badge misleads

The instinct to want one number — "how accurate is it?" — is understandable, but accuracy in counting is not a fixed property of a product. It changes with conditions:

• Lighting. A well-lit daytime scene counts more cleanly than a dim, rainy dusk.
• Crowd density. Sparse pedestrians are easy; a dense, overlapping crowd is genuinely harder, and honest tools flag it rather than hide it.
• Camera angle and distance. A clear side-on view of the visibility zone beats a head-on or partly blocked one.
• Sample length. A longer, better-spread sample projects a week more reliably than a short one.

A "99.5% accurate" badge collapses all of that into one figure that can't possibly hold across every condition. It is a claim you cannot check — the exact thing the audit questions are designed to expose. The honest alternative is to report the uncertainty of each estimate, which is where MAPE and confidence intervals come in.

MAPE and confidence intervals explained

Two measures do the honest work that a single accuracy badge only pretends to. One describes typical error across many estimates; the other describes the uncertainty around a specific number. Together they let you talk about reliability without over-promising.

MAPE in plain language

MAPE stands for Mean Absolute Percentage Error. It answers: on average, how far off are the estimates from the truth, in percentage terms?

If you compared a set of counts against a known ground truth and the gaps averaged 8%, the MAPE is 8% — estimates typically land within about 8% of the real number. Two things to hold onto:

1. It's an average, not a promise. A MAPE of 8% doesn't mean every single count is within 8%; some are closer, some further. It describes the typical case.
2. It needs a ground truth to compute. You can only state a MAPE by comparing against a hand-verified count. That is validation work done against test footage — and it belongs on the StreetProof methodology hub, where the engine's benchmarks are documented once for every brand that shares it.

This is why AdWitness does not print a headline accuracy percentage on your certificate. Instead of a badge that pretends to apply everywhere, it shows the uncertainty that actually applies to your specific estimate — a confidence interval.

Confidence intervals in plain language

A confidence interval is the range the true value most likely sits within, at a stated confidence level. On a certificate it reads like:

Weekly impressions: ~42,000 (most likely 34,000–50,000).

The band comes from the natural variation in the counts and how much you sampled. Sparse data widens it; more data narrows it. A wider band is not a worse product — it is an honest signal that you should film more before quoting a hard number. The mechanics of building that band from your footage are in estimating impressions from a week of footage.

The trap: precision is not representativeness

Here is the subtlety that catches people out. A confidence interval captures count noise — the wobble in how many people pass. It does not, on its own, tell you whether the hours you sampled looked like a normal week. Film ten minutes at a freak-busy moment and you can get a mathematically tight interval around a number that badly overstates the week.

That is why an honest estimate needs two things, not one:

• a confidence interval for the count noise, and
• a plain-language confidence label — low, medium, or high — for how representative the sample was.

AdWitness marks very short samples as indicative and withholds longer projections (like a monthly figure) from them, precisely so a tight-looking band never gets mistaken for a reliable forecast. That guardrail is the difference between a number that survives a buyer's scrutiny and one that quietly misleads them.

What to say about your own numbers

You don't need to be a statistician to talk about this credibly. Three habits are enough:

1. Quote the range, not just the headline. "About 42,000 a week, most likely 34,000–50,000."
2. Say the confidence level. "High confidence — sampled across five days and every daypart."
3. Point to verification. The QR code and method note let a buyer check the rest.

That is a more persuasive pitch than any accuracy badge, because it demonstrates the honesty a badge only claims.

Go deeper

For the full validation methodology — how the engine is benchmarked and where its error comes from — read the canonical write-up on the StreetProof methodology hub. To see confidence intervals on a real face, certify one face for $99 or view pricing, and return to the OOH audience measurement guide for the wider picture.