For decades, map labels have commonly been prioritized from the top-right. But when users compare the alternatives, the position they prefer most is actually straight above the point. Automatic label placement is important in cartography, data visualization, and geographic information systems, because labels help users interpret visualized data.

Labels convey essential information about distinct data points, including cities and landmarks, as well as pivotal elements in diagrams and charts. The challenge of optimally positioning textual annotations remains an active field of study, especially concerning point-feature label placement.

Point-feature label placement primarily deals with the maximization problem: positioning labels for the maximum number of point features, also called anchors, without causing conflicts. A conflict occurs when a label overlaps or interferes with another label, an anchor, or other important map features, reducing readability or visual clarity.

The fixed-position model restricts labels to predefined positions, so a Position Priority Order ranks those positions according to predetermined priorities. Existing literature reveals a lack of consensus in Position Priority Orders, with different authors assigning different priorities to the same label positions.

Those priorities have historically relied on typographic and cartographic conventions or printer capabilities, with varying justification and consistency. This inconsistency creates a need for an empirically grounded methodology that reflects user perceptions and preferences, while recent approaches continue to rely on traditional orders.

With the exception of Zoraster, the top-right position emerges as the highest priority across multiple bodies of work. The priorities assigned to other positions show even more variability, highlighting a lack of consensus. All reported priorities are based on the authors’ experience and were not empirically verified by users, apart from a limited study covering only top-left, left, and bottom-left.

The disparities and gap in evaluations set the stage for the investigation. The study asks whether specific Position Priority Orders influence perceived label placement quality and aims to establish priorities validated by user-centered research.

The intended result is more intuitive and accessible mapping solutions grounded in empirical evidence. The study randomly selected thirty locations worldwide and excluded locations on the sea or ocean, as well as latitudes greater than minus sixty degrees to exclude Antarctica.

Each location became the center of an area defined by a location and a zoom level from five to ten, rendered as a vector SVG image at one thousand three hundred five by one thousand twenty-five pixels. Settlements with more than five hundred inhabitants from GeoNames were used as anchors, and anchors were filtered so all eight positions were available without conflict.

Areas with fewer than twenty anchors were discarded, producing thirty areas at zoom levels five to eight, with twenty to fifty-four anchors each. Each area was rendered eight times, placing every label in one of eight positions: top-right, top, top-left, left, bottom-left, bottom, bottom-right, and right.

This produced two hundred forty blind maps with a white background, red anchors, and corresponding labels. Blind maps were used to eliminate other factors that could influence judgment, such as patterns and vivid colors in the map background. The design separated position from other cartographic factors, including overlaps, label alignment, font sizes, and colors.

With eight label positions, each area required twenty-eight pairwise comparisons. Across thirty areas, the evaluation contained eight hundred forty pairwise comparisons to cover every pair once. To reduce participant fatigue, each participant received three areas, producing eighty-four comparisons in a batch.

Ten participants were required to cover all pairwise comparisons once. The pairwise comparison data were transformed into an n by n preference matrix for each participant. Each matrix element indicates how many times method i was selected over method j by that participant.

This conversion enabled preference analysis and identification of Position Priority Order patterns across participants. Preference matrices were transformed into quality scores using Thurstone’s statistical judgment model. The model represents each object’s quality score as a Normal random variable, with the mean treated as the true quality score and the variance representing uncertainty.

Thurstone’s Case V model accounts for inter-participant and intra-participant variance using a Normal distribution. The resulting quality differences can be interpreted as z-scores, distances from the mean measured in standard deviations. Participants who exhibited inconsistencies in their responses were eliminated.

After data refinement, two hundred twenty-five participants remained, with a dropout of twenty-seven percent. Those participants generated eighteen thousand nine hundred pairwise comparisons, so each comparison pair was evaluated by approximately twenty-three participants on average.

The final results showed a statistically significant difference between all pairs of label positions. The study therefore rejected the null hypothesis and identified a clear preference order for label positions based on z-scores. Figure three summarizes the PerceptPPO user study in two linked views.

Chart a reports quality z-scores with ninety-five percent confidence intervals for the label positions, while chart b lays out pairwise comparisons: red circles mark positions, and solid blue lines mark statistically significant differences, with absolute z-score gaps and p-values shown on the edges.

Together, the charts show significant differences between every pair, supporting the authors’ conclusion that label-position preferences follow the z-score ordering. Map renders were needed to evaluate and compare PerceptPPO with existing Position Priority Orders, while blind maps again isolated label position relative to the anchor.

That design raised a practical question: how many labels should be presented in a map area? Only a handful of cartographic works had studied label density, so a dedicated experiment examined users’ preferences on label density. The areas were repeatedly rendered across label density thresholds from two point five to forty percent in steps of two point five percent, followed by forty-five, fifty, seventy-five, and one hundred percent.

Higher density levels used larger steps because the renders became perceptually similar as label density increased. When renders for a given area were identical, only unique renders were retained. The overall median preferred label density threshold was twelve point five percent, with a mean of fourteen point eighty-five and a standard deviation of seven point twenty-two percent.

The preferred threshold was determined using the median because it provides a robust measure that minimizes the impact of potential outliers. The twelve point five percent threshold corresponded to a median of one hundred fifty labeled anchors and a median global label density of twelve point five percent.

For local label density, the preferred threshold corresponded to a median of sixteen point three percent. Figure six checks whether the different placement-priority orders are being tested on comparable numbers of labels. The box plots show broadly similar distributions of placed labels across all examined PPOs, both when aggregating label-density thresholds at thirteen-oh-five by ten-twenty-five and at the preferred threshold of twelve point five percent.

Panel c repeats this check at the smaller six-fifty-two by five-twelve resolution, supporting the authors’ claim that anchor configurations do not systematically favor a particular PPO. The comparison study found a statistically significant difference between the group containing PerceptPPO and Zoraster and the group containing Brewer, YoeliB, Christensen, Slocum, and Imhof.

There was no significant difference within either group, but there was a significant difference between the groups. PerceptPPO and Zoraster consistently outperformed the other Position Priority Orders, and participants perceived their group as significantly better.

The result supports placing the top position first, while the following label positions do not seem essential for perceiving Position Priority Order quality. Figure nine compares the quality scores of the preference-ordering approaches, with chart A showing ninety-five percent confidence intervals and chart B testing pairwise differences.

The triangle plot reports no significant differences within the groups containing PerceptPPO and Zoraster, or Brewer, YoeliB, Christensen, Slocum, and Imhof, but significant differences between these groups. This matters because it supports the authors’ grouping of the PPOs and shows that aggregated preferences contain a clear structural separation.

Although the study presents a statistically significant ordering of label positions, it also reveals distinct clusters of user preferences. The generic blind maps enabled precise evaluation by minimizing degrees of freedom, but other map elements could also influence user perception.

The study did not account for semantic considerations such as city labels across state borders or coastal city labels toward water bodies, and preferences might vary by map use and audience. It also did not examine effects related to script direction, familiarity with the presented area, ease of information search, or overall readability.

Despite participant screening and data-integrity measures, the uncontrolled online environment could introduce unseen biases. The central lesson is that user-tested label priorities can outperform inherited conventions: labels above point features were preferred, and users favored twelve point five percent overall label coverage on blind maps.