False confidence - the costliest error in address data. What our report reveals about the reliability of geocoding providers’ quality labels?

When you deploy standardization and geocoding services, the quality label assigned by the provider is often the only thing that decides whether an address moves on through the process or gets pulled aside for extra checks. If that label is wrong, a record that looks perfectly fine slips through automation unchecked and only surfaces later, at a far more expensive, downstream point (for example, the moment a parcel is handed to a courier). We call this phenomenon “false confidence.” In our report “The Address as the Main Key for Linking Data,” we tested 2,500 real addresses to see how far you can actually trust providers’ quality labels - and we show that for most of them, you can only trust them up to a point.

What is “false confidence,” and why does it cost the most?

False confidence is a record that the service has stamped with its highest match status, but which in reality has the wrong standardization, the wrong location, or both. From a process standpoint it’s the worst kind of error there is, precisely because it sets off no alarm at all.

It’s worth contrasting it with an “honest” error. When a service returns low confidence or no match, it is signaling doubt - the record goes to exception handling, and you can estimate the cost of that handling in advance. A falsely confident record sends no such signal: it’s treated as correct, feeds into routing, matching against customer history, or risk scoring, and its consequences only come to light at a later, costlier stage - a failed delivery, a mismatched record, or a distorted risk assessment.

That’s why a service’s real value comes down not just to how often it returns correct versus incorrect results, but also to whether those incorrect results were flagged early enough.

How did we test the reliability of the labels?

We ran the tests on 2,500 real addresses pulled from the system of one of the largest courier operators (December 2025) - addresses carrying the usual quality problems: typos, abbreviations, and inconsistent formatting. To judge whether a given service had corrected an address properly, we needed something to measure against, so for all 2,500 records we built a golden set: a single model, correct version of each address and one reference location, reconciled with the Polish registers (TERYT, EMUiA, and the postal code index). That way, for every record we knew what the correct result should look like, and we could check whether the provider’s label had been honestly earned.

To capture label reliability in a single number, the report calculates a trust index (trust_index). It brings together two measures:

• how often results labeled “highest quality” are in fact wrong;
• and what share of the entire data stream is made up of records that meet all three “end-to-end” conditions at once: highest quality according to the provider, a location error below 50 meters, and fully correct component-level standardization.

We calculated every metric separately for urban and rural areas, because an averaged figure can hide important differences.

A “highest quality” label doesn’t always mean “correct”

Let’s start with the simplest test. The first chart shows what share of the records a service flagged with its top geocoding and standardization quality also turn out to have fully correct standardization. For Data Quality that figure is 98.8%, and for Locit 95.5% - a “highest quality” status almost always means a correct record. For the other providers, that link weakens: Google 62.7%, Mapbox 59.5%, while Emapa, Azure, TomTom, Precisely, Esri, and HERE land somewhere around 38–47%. In other words, for most providers fewer than half of the records carrying the top label have fully correct standardization.

Share of records flagged by the provider with the highest geocoding and standardization quality that also have fully correct standardization (the higher, the more trustworthy the label).

The same pattern shows up in addresses for which no correct answer even exists. From the 2,500 records we singled out 50 unresolvable ones - addresses that couldn’t be matched unambiguously to any point in the registers. We then checked what share of these each service stamped with its highest quality status anyway.

Share of unresolvable addresses to which the service assigned its highest quality status (the lower, the better).

As it turns out, even when there is no correct answer, most services flag a sizable share of these records as highest quality - from 26% (Google) to 54% (Precisely). This is false confidence in its purest form: a high-quality label turning up where there’s nothing to support it. Here too, Data Quality (6%) and Locit (12%) keep their restraint.

What goes into the trust index?

The trust index (trust_index) isn’t a matter of opinion - we calculated it from two specific, measurable components.

Component 1. False positive risk (weight 0.7)

The first component is the weighted risk of false-positive results - a measure of how often a service flags records with a location error of more than 250 m as a “high quality” match.

Weighted risk of false-positive results (fp_weighted) - the basis for the fp_score component (the lower, the better).

Two providers stand out here: Data Quality (0.004) and Locit (0.005). Then comes a clear jump - Esri 0.015, Google 0.019 - with the middle and tail of the pack made up of Azure, TomTom, and Emapa (about 0.037–0.040) along with HERE, Mapbox, and Precisely (0.045–0.073). The gap between the best and the weakest service is almost 18-fold. Rural addresses consistently come off worse: in the countryside, Precisely’s false-positive risk reaches 0.094 and HERE’s 0.066.

Component 2. End-to-end effectiveness (weight 0.3)

The second component is the share of records that satisfy all three conditions at once: highest match quality according to the provider, a location error below 50 m, and fully correct standardization of the address components. This is the toughest test - a record stamped “high quality” has to be right both in how it’s written and in where it’s placed.

Share of records with the highest match quality, a location error below 50 m, and fully correct standardization - the end-to-end component (the higher, the better).

Data Quality holds at 94.8% and Locit at 87.3%, while Google drops to 55.9% and the rest sit in the region of 32–39%. Once you ask a “highest quality” label to deliver both the right point on the map and the correct address, only 32–56% of the records flagged that way are fully correct for eight of the ten services - too few to lean on the quality label alone.

The trust index

We combine the two components into a single figure, giving more weight (0.7) to false-positive risk and less (0.3) to end-to-end effectiveness.

trust_index = 0.7 × false positive risk + 0.3 × end-to-end effectiveness

Trust index for standardization and geocoding results (the higher, the more reliable the quality label).

The picture is the same as in the earlier charts. Data Quality reaches 0.85 and Locit 0.79, and then there’s a cliff - Esri 0.30, Google 0.19, with every other service falling into the 0.10–0.12 band. Only two services reach a level where automation can rely on the quality label on its own. Below roughly 0.3, the label needs its own validation layer - otherwise falsely confident errors go straight into the process.

What does it cost?

Every choice of service carries a real, hidden cost that never shows up on the price list. In a simulation built around a scenario of 100,000 shipments a year, the extra costs range from about PLN 101,000 for the best service to about PLN 897,000 for the weakest - nearly a ninefold difference. They come from a mix of problems: the manual handling of unresolvable and uncertain records, incorrect postal codes, and false-positive results.

Simulation of additional logistics process costs assuming 100,000 shipments / orders annually.

The cost of FP (false positive) errors alone - the serious and critical ones - comes to about PLN 18,000 a year for Data Quality and about PLN 327,000 for Precisely, the service with the highest false-positive risk. A service that’s cheap to buy can prove the most expensive to run, in part because it declares high quality far too readily.

Of all the error types in the simulation, the most expensive is exactly this critical false positive (a location error of more than 1 km), which we price at PLN 60 per record - three times more than an incorrect postal code (PLN 20) and 7.5 times more than a record without top match status that has to be checked by hand.

The reason is that a record stamped “high quality” usually isn’t verified any further and goes straight in, deep into the process, which means it’s caught much later than other kinds of error. So false positives don’t have to be the most numerous to dominate the bill - at Precisely there are almost four times fewer critically wrong results than uncertain records, yet they cost nearly twice as much (about PLN 286,000 versus about PLN 150,000), because each one of these errors is many times more expensive. The practical takeaway: it pays to choose a service that minimizes the number of wrong, falsely confident labels - or to run your own validation mechanisms, which first have to be built and then maintained, adding costs of their own.

How can you cut the cost of false confidence?

• Treat “highest quality” status as a filter only when the data backs it up. In our report, it proved reliable for Data Quality and Locit. For the other services, you need extra validation - a second pass, a cross-check against the registers, or acceptance thresholds set more cautiously.
• Look at the results separately for cities and the countryside. Labels are less reliable for rural addresses, and in nationwide processes it’s precisely those that can generate a disproportionate number of errors.
• Run the numbers on your own data. The report comes with a cost-simulation spreadsheet; just plug in your own volume and rates to see what a given service’s false confidence would really cost your organization.

Accessing the report

We’re making the full report, along with the simulation spreadsheet, available free of charge at this link. If you have questions or would like to talk through the findings, we’d be glad to hear from you - just get in touch.