# Accuracy or coverage in mapping? What 99.99% really means

> URL: https://hotelmapping.org/en/guide/accuracy-or-coverage-in-hotel-mapping/ · Language: English · Last updated: 2026-11-15

**In hotel mapping, accuracy shows how many of the automatic matches are correct; coverage shows how many of the records that should match actually matched. Both cannot peak at once: raising the threshold reduces wrong matches but lowers coverage and grows the human review queue. A "99.99% accuracy" claim cannot be interpreted without knowing which test set and which denominator it was measured on.**

## Short answer

One of the figures seen most often in hotel mapping providers' marketing copy is "99.99% accuracy." The number looks impressive, but on its own it does not tell you how useful a mapping service will be. That is because mapping quality is described not by a single number but by three interdependent numbers:

- **Accuracy:** What share of the automatic matches is correct? The inverse is the wrong-match rate.
- **Coverage:** What share of the records that truly belong to the same hotel was actually matched?
- **Human review rate:** What share of records could not be decided automatically and fell to an expert for review?

These three numbers move together. A tweak made to improve one often worsens one of the other two. In this article we explain that balance, how threshold choice changes the three numbers, and the questions to ask when evaluating high-accuracy claims. For how hotel mapping works in general, start with the [hotel mapping guide](https://hotelmapping.org/en/guide/what-is-hotel-mapping/).

## Accuracy, coverage and human review: three separate questions

For every supplier record, a mapping engine makes one of three decisions: link the record to an existing hotel, ask an expert, or open a new hotel ID for the record. Each decision can be right or wrong, and the cost of each error is different.

| Outcome | What happened? | Which metric does it affect? | Impact on the agency |
| --- | --- | --- | --- |
| Correct match | Two records of the same hotel were linked | Accuracy and coverage | Prices are compared on the correct hotel |
| Wrong match | Two different hotels were linked to the same ID | Lowers accuracy | A customer is sold one hotel, the booking goes to another |
| Missed match | A record of the same hotel received a new ID | Lowers coverage | The same hotel appears twice in the list (duplicate) |
| Human review | The score fell in the uncertain zone, the record went to an expert | Raises the human review rate | The right decision is made, but time and effort are needed |

A wrong match is usually the most expensive error. The customer pays for one hotel, the booking is forwarded to another property, and the problem often surfaces at the hotel's front desk. A missed match is quieter: the same hotel appears in the list under two different names with two different prices. This lowers conversion and breaks price comparison. We explain how duplicates are cleaned up in the [duplicate cleanup](https://hotelmapping.org/en/glossary/deduplication/) entry and what they cost an agency in the [duplicate hotel cost article](https://hotelmapping.org/en/guide/cost-of-duplicate-hotel-listings/).

Human review, on the other hand, is not an error but a cost. Showing uncertain records to an expert reduces both kinds of error, but each record takes time, and the matching of a new supplier file is delayed.

### Why can't both reach the maximum?

For every candidate pair, the mapping engine produces a [match score](https://hotelmapping.org/en/glossary/match-score/): a value between 0 and 1 indicating the likelihood that the two records are the same hotel. The decision depends on which threshold this score clears.

If you raise the threshold, only records the engine is very sure about are matched automatically. Wrong matches decrease and accuracy rises. But correct matches whose scores are slightly lower also fall outside automatic decisions: they either go to the review queue or receive a new ID and become duplicates. If you lower the threshold, the opposite happens: coverage rises, but wrong matches slip in as well.

That is why the sentence "our accuracy is 99.99%" remains incomplete until the question "what is your coverage?" is answered. A system that matches no records makes no wrong matches; it is technically error-free, but completely useless.

## How does threshold choice change the three numbers?

The decision rule planned at hotelmapping.org consists of two thresholds: a score of 0.97 or above is an automatic match, between 0.80 and 0.97 goes to human review, and below 0.80 gets a new HM ID. The thresholds are tuned against a manually labeled test set. The effect of these two thresholds and of alternative settings can be shown with a worked example.

### Worked example (entirely hypothetical)

**The figures below are examples.** They do not belong to any real measurement, provider or customer; they are constructed only to show the effect of threshold choice.

Suppose we have 10,000 labeled supplier records. 9,000 of them have a true counterpart in the database, and 1,000 are new hotels not yet present in any source. We assume the expert makes the correct decision on records sent to human review. When the same records are processed with four different settings, the table might look like this:

| Scenario | Setting | Automatic accuracy | Automatic coverage | Coverage after review | Human review rate | Wrong matches | Missed matches (duplicates) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| A | Single threshold 0.90, no human review | 99.50% | 97.3% | 97.3% | 0% | 44 | 214 |
| B | 0.97 / 0.80 | 99.90% | 89.9% | 97.6% | 7.5% | 8 | 210 |
| C | 0.99 / 0.80 | 99.99% | 77.8% | 97.7% | 18.5% | 1 | 210 |
| D | Single threshold 0.99, no human review | 99.99% | 77.8% | 77.8% | 0% | 1 | 2,000 |

Four conclusions follow from the table:

1. **Scenarios C and D have the same accuracy: 99.99%.** But in scenario D, 2,000 hotels remain duplicated; roughly one in five hotels that have a counterpart appears twice in the list. The same accuracy figure can describe two very different services.
2. **Scenario C buys 99.99% accuracy with human effort.** 18.5% of records go to an expert. In a file of 10,000 records, that means 1,850 records. When a new supplier is added, it can take days for the file to be fully matched.
3. **Scenario A is fast and broad, but leaves 44 wrong matches.** Each one is a risk of sending a customer to a different hotel.
4. **The lower threshold (0.80) determines the missed matches.** In scenarios B and C, the number of records left duplicated is the same (210) even though the upper threshold changes. To reduce duplicates you need to lower the lower threshold, which grows the review queue even more.

Scenario B in the example corresponds to the thresholds planned by hotelmapping.org. In this example the human review rate comes out at 7.5%; that is, above the planned 5% target. This also shows that threshold tuning alone is not enough. The real way to shrink the review queue is to reduce the number of records the model places in the uncertain zone: better normalization, cleaner coordinates, additional signals such as phone number and web address. The threshold chooses where you stand on the same curve; a better model improves the curve itself.

### What should a measurement report look like?

A quality measurement should show not only the resulting percentage but also the numbers from which that percentage was computed. The JSON below shows how scenario B above might appear in a measurement report. The field names and values are examples; this is not a finalized hotelmapping.org report format.

```json
{
  "example": true,
  "test_set": {
    "name": "ornek-akdeniz-2026",
    "records": 10000,
    "records_with_counterpart": 9000,
    "labeling": "two experts, a third on disagreement",
    "markets": ["antalya", "mugla", "istanbul"]
  },
  "thresholds": { "auto_match": 0.97, "review": 0.80 },
  "decisions": {
    "auto_match": { "count": 8100, "correct": 8092, "wrong": 8 },
    "review": { "count": 750, "matched_after_review": 690, "new_id_after_review": 60 },
    "new_id": { "count": 1150, "missed_counterpart": 210 }
  },
  "metrics": {
    "precision_auto": 0.9990,
    "coverage_auto": 0.8991,
    "coverage_after_review": 0.9758,
    "review_share": 0.075
  },
  "by_market": [
    { "market": "antalya", "auto_match": 3400, "wrong": 6, "precision_auto": 0.9982 },
    { "market": "other", "auto_match": 4700, "wrong": 2, "precision_auto": 0.9996 }
  ]
}
```

Note the `by_market` section. Accuracy that is 99.90% overall drops to 99.82% on Antalya records; 6 of the 8 wrong matches are in this market. Many resorts with the same name, locations close to one another and multilingual names make one region harder than the others. The average figure hides this difference.

## How should you question a "99.99% accuracy" claim?

When you see a high accuracy figure, asking the questions below reveals what the number actually describes. These questions apply to every provider, including hotelmapping.org.

### 1. Which test set was it measured on?

Behind the figure there should be a manually labeled test set. How many records or pairs were labeled? Were the records chosen at random, or are easy-to-match chain hotels over-represented? Was the test set kept separate from the data the model was trained on? How many people did the labeling, and under which rule?

The size of the test set matters too. Even if no errors appear in a test set, that does not mean the rate is conclusively proven. According to the "rule of three" commonly used in statistics, when no errors are seen in n records, the upper bound of the true error rate is taken as roughly 3/n (at a 95% confidence level). Zero errors in 5,000 automatic matches indicates an error rate below about 0.06%; in other words, you can say at most "99.94% and above." To show 99.99% with this method, roughly 30,000 automatic matches would need to come out error-free.

### 2. Which market was it measured in?

A global average may not show the situation in the regions you sell. City hotels and large chains are usually easier to match; Mediterranean resort regions, aparthotels and boutique hotels are harder. Ask for separate figures for regions such as Antalya, Muğla, Dubai or Egypt. Was the figure measured only with names in a single language, or are Russian and Arabic names in the test set too?

### 3. Which denominator was it computed on?

This is the most frequently overlooked question. Accuracy can be computed in three different ways, and all three give very different figures:

- **Over automatic matches:** Correct matches / all automatic matches. This is the definition in this article.
- **Over all records:** Correct decisions / all records. Correct "new hotel" decisions are added to the numerator, so the figure rises.
- **Over all candidate pairs:** Correct decisions / all pairs compared. Nearly all compared pairs are different hotels anyway. If only 9,000 of 1,000,000 candidate pairs are true matches, even a system that matches nothing reaches a 99.1% "correct decision" rate.

If the denominator is not stated, the figure cannot be compared.

### 4. Only automatic matches, or all matches?

Some figures also include matches that have passed through human review. In that case, the high accuracy is the success of the expert team, not of the model. That is not a bad thing, but it is a different thing: human review takes time and determines how long it takes for new supplier files to be matched. Ask separately for the accuracy of automatic matches and the human review rate.

### 5. What are the coverage and review rate at the same threshold?

The accuracy figure should be given together with the coverage and human review rate measured at the same threshold. Scenario D in the worked example shows that 99.99% accuracy can come with 77.8% coverage. If a provider shares only accuracy, ask separately for the other two figures.

### 6. Was it tested on your own data?

The most reliable measurement is a trial using a sample taken from your own supplier data. Label a few hundred records from the regions you sell yourself and compare the provider's result against those labels. This method tells you far more than a published general figure. For other questions to ask when choosing a provider, see the article [12 questions to ask when choosing a hotel mapping API](https://hotelmapping.org/en/guide/12-questions-hotel-mapping-api/).

## Which metric should take priority?

The right balance depends on what you use the mapped data for:

- For **agencies and tour operators that take bookings**, a wrong match is the most expensive error. It makes sense here to prioritize accuracy and send uncertain records to human review.
- For **price comparison sites**, both accuracy and coverage matter. A wrong match leads to a wrong price comparison, and a missed match leads to an incomplete price list.
- For **reporting and analytics**, coverage may matter more; the occasional wrong match does not change the overall statistics much.

Whichever you choose, the decision needs to be visible for every match. In hotelmapping.org's API v1 responses, each match's confidence score is returned in the `confidence` field and the decision taken in the `decision` field. This lets you apply a stricter threshold with your own business rule. For details, see the [matching API](https://hotelmapping.org/en/products/find-match-api-solutions/) page.

## How will hotelmapping.org measure these figures?

The quality of the [hotel mapping](https://hotelmapping.org/en/products/hotel-mapping-solutions/) product will be measured with every release using a test set of at least 5,000 manually labeled hotel pairs. The test set will include Antalya, Muğla, Istanbul, Dubai and Egypt records in particular. The metrics reported will be the three numbers described in this article: accuracy (wrong-match rate), coverage (share of records matched) and the share of records that fall to human review.

No measured results have been published yet. Once the test set is complete, the measured values will be published on the [coverage](https://hotelmapping.org/en/coverage/) page together with the denominator and market information they were measured on. On the site we will include only figures verified by this measurement.

## Summary

Accuracy, coverage and the human review rate are three faces of the same threshold setting. Raising the threshold reduces wrong matches, but it grows either the duplicates or the review queue. A figure like "99.99% accuracy" carries meaning only when the test set, market, denominator and the distinction between automatic and reviewed matches are known, and when it is accompanied by the coverage figure at the same threshold. When evaluating a provider, ask for all three numbers together and, if possible, try it on your own data.

## Frequently asked questions

### Are accuracy and coverage the same thing?

No. Accuracy measures how many of the matches given are correct; coverage measures how many of the records that should match were actually matched. A system that returns no matches makes no wrong matches either, but its coverage is zero. That is why both figures must always be given together and for the same threshold.

### Does a 99.99% accuracy claim mean it is wrong?

It does not have to be wrong, but on its own it is incomplete. Without knowing which test set, which market and which denominator the figure was measured on, and whether it covers only automatic matches, you cannot tell what it means. Ask for the coverage and human review rate at the same threshold as well.

### Can I change the threshold values myself?

In API v1 responses, each match's confidence score is returned in the confidence field. You can use this score to apply a stricter rule on your side, for example accepting only scores above 0.99 automatically. Whether per-customer threshold settings will be offered on the service side has not been finalized yet.

### What are hotelmapping.org's measured accuracy and coverage?

There is no published measurement yet. The plan is to measure with at least 5,000 manually labeled hotel pairs and publish the result on the coverage page.

## Related products

- [Hotel mapping solutions](https://hotelmapping.org/en/products/hotel-mapping-solutions/)
- [Find Match API solutions](https://hotelmapping.org/en/products/find-match-api-solutions/)

## Related articles

- [What is hotel mapping? How it works: a complete guide](https://hotelmapping.org/en/guide/what-is-hotel-mapping/)
- [12 questions to ask when choosing a hotel mapping API](https://hotelmapping.org/en/guide/12-questions-hotel-mapping-api/)

## Related terms

- [Match score](https://hotelmapping.org/en/glossary/match-score/)
- [Deduplication](https://hotelmapping.org/en/glossary/deduplication/)
