Which Statistical Measure Should I Use to Compute a Safe Geo-Block Limit from Few Records?

If you have only a few records, do not block a geographic area based on its raw invalid rate or cost per lead. Use a confidence interval—specifically, the upper bound of a Wilson score interval for proportions or a Poisson confidence interval for click counts—to set your geo-block limit. This way, a region is only blocked when the evidence is strong enough that even the most optimistic interpretation of its true performance is still below your threshold.

The problem is sample size. With 10 leads, one bad batch of 4 leads looks like a 40% invalid rate. But the true rate could be anywhere from 16% to 62%. A geo-block limit built on that point estimate will regularly block good regions and miss bad ones. Confidence intervals solve this by giving you a range of plausible values.

Why the Right Statistical Measure Matters

A safe geo-block limit is a threshold. When a region crosses it, you stop spending there. If the threshold is too loose, you waste budget on fraudulent or invalid clicks. If it is too tight, you block regions that contain real buyers. Statistical measures are the only way to balance those risks.

Ignoring them leads to two expensive mistakes. First, you block a city or country based on a handful of bad leads. Second, you let a truly fraudulent region keep running because the noise hides the signal. The right measure turns a guess into a defensible rule. It also helps you explain the decision to a client, a manager, or an ad platform when you request a refund.

The Core Problem: Few Records Mean High Uncertainty

With few records, the variance is huge. A point estimate like "30% invalid" is just the average of what you saw. It does not tell you how confident you should be. If you observed 3 invalid out of 10, the true invalid rate might be 7% or 65%.

This is why statistical significance matters. You need a measure that accounts for the number of records. A confidence interval does exactly that. It widens as the sample size shrinks. A wide interval means "keep collecting data." A narrow interval means "you can make a decision."

How the Math Works: Intervals, Bounds, and Sample Size

A confidence interval gives you a lower bound and an upper bound. The lower bound is the most optimistic true value. The upper bound is the most pessimistic true value. For a negative metric like invalid rate, you care about the upper bound.

For example, a 95% Wilson interval for 4 invalid leads out of 10 runs from about 16% to 62%. The raw percentage is 40%, but the interval tells you the truth could be much better or much worse. If your business threshold is 30%, you cannot safely block the region because the lower bound is below 30%.

The Poisson interval works the same way but for counts. If you observe 5 invalid clicks in a day, the 95% Poisson interval for the true rate is roughly 1.6 to 11.8. You need to compare that upper bound against your daily threshold.

Candidate Measures and Their Trade-offs

Raw percentages are tempting because they are simple. But they are unstable. A single bad lead can move the percentage from 0% to 50%. With a sample size below 30, raw percentages should not be used for blocking decisions.

Standard deviation rules are designed for symmetric, continuous data. Click and lead data are counts, often skewed. Applying a "two standard deviations" rule to a small count distribution will produce misleading limits. It is better to use intervals built for counts and proportions.

Bayesian methods are powerful but require you to choose a prior. The prior introduces subjectivity. Unless you have strong historical data or expert opinion, the added complexity is rarely worth it for a simple geo-block rule.

The Wilson score interval is the best default for most geo-blocking decisions. It is designed for proportions and behaves well even when the sample size is small. It also works when the proportion is 0% or 100%, which breaks simpler formulas.

Decision Rule: How to Choose the Right Measure

Use the Wilson score interval when your geo-block metric is a proportion. Examples include invalid leads divided by total leads, or bounces divided by clicks.

Use the Poisson confidence interval when your metric is a count over a fixed window. Examples include invalid clicks per day or blocked events per week.

Do not use raw percentages when your sample size is below 30. Do not use standard deviation rules when your data is a count or a proportion. If you need a simple business rule, set a minimum sample size. For example: "We only block a geo after 30 or more clicks, and only if the upper bound of the 95% confidence interval is above our quality threshold."

Step-by-Step Framework to Set a Defensible Geo-Block Limit

1. Define the metric. Choose what you are measuring. Common choices are invalid lead rate, cost per valid lead, or click-to-session gap.
2. Set a business threshold. Decide what number means "bad." For example, an invalid lead rate above 30%.
3. Collect records for the geo. Gather clicks, leads, or sessions for the specific region you are evaluating.
4. Calculate the confidence interval. Use a Wilson score calculator for proportions. Use a Poisson calculator for counts. A 95% confidence level is a reasonable standard.
5. Apply the decision rule. Block the geo only if the lower bound of a positive metric (like valid rate) is below your threshold, or the upper bound of a negative metric (like invalid rate) is above your threshold.
6. Require a minimum sample size. Do not make blocking decisions on fewer than 10 records. Ideally, wait for 30 or more.
7. Document and review. Write down the threshold, the interval, and the decision. Review the geo again after a few weeks to confirm the block was justified.

Practical Scenarios: When a Geo-Block Limit Makes Sense

Scenario A: Small City, 10 Leads

A city generates 10 leads. 4 are invalid. The raw invalid rate is 40%. The Wilson upper bound is 62%. Your business threshold is 30%. Do you block the city? No. The interval shows you cannot be confident the true rate is above 30%. The lower bound is 16%. The city might be fine. Collect more data.

Scenario B: Large Country, 500 Leads

A country generates 500 leads. 150 are invalid. The raw invalid rate is 30%. The Wilson upper bound is 34%. The lower bound is 26%. You can be confident the true invalid rate is between 26% and 34%. If your threshold is 25%, you block it. The evidence is strong.

Scenario C: New Campaign, 5 Clicks

A new campaign gets 5 clicks. 2 are suspicious. Do not compute a geo-block limit. With 5 records, any interval will be too wide to be useful. Wait until you have at least 20-30 records before making a decision.

Limitations and When This Advice Does Not Apply

Statistical measures cannot tell you if a click came from a bot or a disinterested human. They only tell you if the number is unusual. A high invalid rate might mean fraud, but it might also mean a poorly targeted ad or a broken landing page. Always pair the math with session-level evidence.

The advice also assumes your tracking data is accurate. If your click IDs, conversion pixels, or CRM data are misconfigured, the calculations are meaningless. Finally, geo-blocking is a blunt tool. Blocking an entire country may cut off valuable traffic. A better approach is to block the specific source of invalid traffic, such as a data center IP range or a suspicious placement.

Key Facts

The following facts come from the BotRefund source pack and provide context for why geo-blocking decisions matter.

Note: The Imperva statistic is industry context, not a claim about any specific advertiser's account. Your own account must be measured on its own evidence.

Frequently Asked Questions

What is a geo-block limit?

A geo-block limit is a threshold. When a specific region's traffic quality crosses that threshold, you block that region from your ad campaigns. It is a statistical safeguard against wasting budget on invalid traffic.

Why can't I just use the average cost per lead?

Averages hide uncertainty. With few records, one bad lead can make the average look terrible. A confidence interval shows the range where the true average is likely to fall. That range helps you avoid false positives.

What is the Wilson score interval?

The Wilson score interval is a formula for calculating a confidence interval around a proportion. It works well with small samples and does not break when the proportion is 0% or 100%. Use it for rates like invalid leads per total leads.

How many records do I need before I can trust a geo-block decision?

Aim for at least 30 records. With fewer than 10, the confidence interval will be too wide to support a safe block. The more records you have, the narrower the interval and the more confident you can be.

What is the difference between the lower bound and the upper bound?

The lower bound is the most optimistic true value. The upper bound is the most pessimistic true value. For a negative metric like invalid rate, use the upper bound to decide whether to block. For a positive metric like valid rate, use the lower bound.

Should I block a geo if the upper bound is above my threshold?

No. If the upper bound is above your threshold but the lower bound is below it, the evidence is mixed. The region could be good or bad. Wait for more data. Only block when the entire interval is on the bad side of your threshold.

What if I don't have enough records but need to act now?

If you suspect fraud but lack data, do not block the entire geo. Instead, pause the specific placement, ad set, or audience that looks suspicious. Or use a session-level tool that can identify bots immediately, rather than waiting for aggregate statistics.