Visual representation of information

A visual representation of information uses spatial position, shape, colour and connection to show what could otherwise be stated in words or numbers. The subject covers three overlapping practices: data visualisation (charts and plots of quantitative data), information design (diagrams, tables, flowcharts and timelines that show structure, process or relations), and the study of when pictures help people understand and remember compared with text. The main disputes are whether a single ranking of visual encodings by accuracy exists, whether embellished charts help or hurt, and how far the benefits of pictures depend on the task and the reader’s prior knowledge.

History

Terminology has never been fixed. The Nielsen Norman Group defines a data visualisation as a graphical representation of a set of data, and an infographic as a multimedia graphic presenting complex information, adding that a data visualisation alone is not an infographic (Kane, 2018). An overview in Library Technology Reports cites many definitions of information visualisation, among them the Card, Mackinlay and Shneiderman account of interactive representations that enhance cognition, and uses the term interchangeably with data visualisation (Chen, 2017).

The research tradition on how well people read charts is usually traced to Cleveland and McGill (1984), whose accuracy hierarchy is reproduced in course slides: position along a common scale, then position along non-aligned scales, then length, direction and angle, then area, then volume and curvature, and last shading and colour saturation (Soltoff, 2026). Heer and Bostock (2010) extended the tradition to area judgements such as bubble charts and treemaps (Soltoff, 2026), and showed that such perception experiments could be run on Mechanical Turk (Heer & Bostock, 2010).

On the diagram side, Larkin and Simon (1987) argued that diagrams help because they group information spatially, which reduces the search needed to find it (Larkin & Simon, 1987). Work on pictures in learning from text is older still: a review of 1990s research concluded that carefully constructed text illustrations generally improve learners’ performance on text-based tasks (Carney & Levin, 2002).

Core concepts

Choosing between prose, table, chart and diagram

Sources converge on a split by what the reader has to do. Tables suit comparing values among related items and looking up specific numbers, while figures suit summarising trends and patterns and explaining procedures (UBC Wiki, 2021). Canada.ca guidance says charts communicate data more efficiently than text or tables, but that a chart should not be used when the data are simple enough for a table, and recommends pairing a chart with an expandable table of the underlying data (Canada.ca Experience Office, 2018). The choice is field-dependent: Kastellec and Leoni argue that replacing tables with graphs, notably for regression results, improves clarity and accessibility for readers in political science (Kastellec & Leoni, 2007).

When there is little to show, a visual may add nothing. A data-visualisation practitioner advises that a few numbers can be stated in simple text, which can beat a table or graph (Nussbaumer, 2012). That is practitioner advice; no experiment testing when prose beats a visual was found.

How people decode a chart

A chart works by mapping data to visual features such as position, length, angle, area and colour, and readers recover values by judging those features. The perception tradition ranks features by how accurately they are judged (see History). See Perception for the underlying visual processes.

Several kinds of design failure are documented. Bar charts whose y-axis does not start at zero can visually exaggerate effect sizes, and the subjective impact persists even with visual cues that mark the truncation; truncation can also be beneficial, depending on communicative and analytic intent (Correll et al., 2020).

Diagrams: structure, process and relation

Diagrams help when they exploit spatial grouping of information (Larkin & Simon, 1987). Tversky and colleagues report that people use a small set of schematic figures (lines, blobs, crosses, arrows) to convey context-specific concepts, and that readers interpret them appropriately in context (Tversky et al., 2000).

Arrows are the documented weak point. In biology textbook figures, arrow styles were used with no consistent relation to meaning (Wright et al., 2018); few arrow styles had intrinsic meaning for study participants, and in 14 interviews all students experienced at least some confusion interpreting arrows (the paper’s Discussion states this more strongly than its Results) (Wright et al., 2018).

Practitioner conventions add to this but carry no experimental evidence. A diagramming vendor says flowcharts should run top to bottom or left to right and not mix the two, and that decision symbols need extra room for branch labels (Nishadha, 2024). A UK public body advises that an image should assist rather than be the only carrier of information, should not be added for decoration, and that text and clutter in charts and diagrams should be minimal (Acas (Advisory & Arbitration Service), 2026).

Accessibility

WCAG requires parts of graphics needed to understand content to have at least 3:1 contrast against adjacent colours (W3C Web Accessibility Initiative, 2026a), and requires that colour is not the only visual means of conveying information; a chart conveying information by colour alone is a listed failure (W3C Web Accessibility Initiative, 2025). The W3C advises that complex images such as charts need a two-part text alternative, a short and a long description (Eggert & Abou-Zahra, 2026). See Inclusive design for the wider picture.

Do visuals improve learning?

Meta-analytic evidence favours adding graphics to text. A meta-analysis of graphics in reading found a moderate positive overall effect on comprehension (Hedges’s g = 0.39) regardless of grade level; pictures, pictorial diagrams and flow diagrams did not differ significantly, and benefit varied by assessment format (Guo et al., 2020). The multimedia principle, that words plus pictures teach better than words alone, is the organising statement of this field, and the handbook chapter on it treats prior knowledge as a condition under which it is studied (Butcher, 2014).

Animation is not clearly better than still pictures. A meta-analysis found a small positive overall effect of animation over static graphics (g = 0.226), larger when there was no accompanying text (Berney & Betrancourt, 2016).

Placing words and pictures together

The split-attention effect concerns learners who must hold apart a diagram and its explanatory text. Ginns’s (2006) meta-analysis, as reported by Beege and colleagues, supported it with a weighted mean effect of d = 0.85 for placing related representations close together. The effects were large for complex material but small and non-significant (d = 0.28) for low-complexity material, and the design rule is to place related representations close together (Beege et al., 2019). Chandler and Sweller are reported to have found that integrating text and diagrams reduces cognitive load and aids learning (Wikipedia contributors, 2025). This links to the limits of working memory.

Redundancy, repeating the same content in several forms, is a separate question. A meta-analysis found spoken-plus-written presentations beat spoken-only ones, but that concerns redundant text, not pictures versus text (Adesope & Nesbit, 2012).

Questions, controversies, and new developments

Is there one ranking of encodings?

The Cleveland and McGill ordering is widely taught, but its universality is contested. Saket, Endert and Demiralp report that the effectiveness of basic chart types varies significantly by task and data attribute type, so no single ranking fits all tasks (Saket et al., 2018). Averages can also hide individual differences: some people show patterns of accuracy that credibly deviate from the canonical rankings (Davis et al., 2022). Reda and colleagues go further and argue that graphical-perception studies of this kind are fundamentally flawed, because small changes in data distribution or design details can shift perception even when the encoding stays the same; this is a recent position paper that has not yet been peer-reviewed (Reda et al., 2026).

Chart junk and embellishment

Minimalist advice treats decoration as harm. Against it, Kosara reports that in Bateman et al. (CHI 2010) participants were more likely to get the message of Holmes-style embellished charts and recalled them better after two to three weeks (Kosara, 2010). He also notes that the study did not test reading exact numbers, so it does not support embellishment for analysis tasks (Kosara, 2010). The EU data visualisation guide advises minimalism for readable values and comparisons, and the Holmes style for engagement and memorability (Publications Office of the European Union, 2023).

In multimedia learning, irrelevant material is called a seductive detail. A meta-analysis summarised by Désiron and Schneider confirmed an overall negative effect on learning, but only for seductive details in audio or text-and-picture format; the mere presence of a pictorial seductive detail had no overall significant effect, so harm depended on format (Désiron & Schneider, 2024).

Pie charts

Kosara argues, citing Spence and Lewandowsky, that pie charts perform about as well as bar charts and better on more complex tasks (Kosara, 2011). Few argues that bar graphs outperform pie charts on every task except comparing summed parts, and that the pro-pie study used an unrealistic task (Few, 2013).

Do novices and experts need different visuals?

A blog summary of Kalyuga, Ayres, Chandler and Sweller (2003) reports that novices benefited from text integrated into diagrams, while experienced trainees did better with a diagram alone, because information essential to novices can be redundant for experts (Cobb, 2021). A blog summary of a 2025 meta-analysis of the expertise reversal effect reports that high-support instruction helps low-prior-knowledge learners and hurts high-prior-knowledge learners; it concerns instruction in general, not visuals specifically (Watson, 2025b).

Visuals in AI-generated output

Language models produce charts mostly as code, and they sometimes fail. VisEval scores generated charts on validity, legality and readability across 2,524 queries (Chen et al., 2024b); invalid code often comes from incorrect or non-existent API calls, and a main cause of illegal data transformations is misunderstanding the data table (Chen et al., 2024b). Reported error rates depend on the metric: LIDA reports a 3.5% visualisation error rate defined only as code compilation errors, and including a data summary in the prompt reduced the error rate compared with field names alone (Dibia, 2023). LIDA’s chart quality is scored by GPT-4, and its author calls for more comprehensive benchmarks (Dibia, 2023). MermaidSeqBench scores generated Mermaid sequence diagrams with language-model judges whose strictness differs, so absolute scores depend on the judge, and larger models do better than smaller ones in the same family (Shbita et al., 2025).

On how readers respond, only small qualitative evidence exists. In a formative study of chat users, text responses were found overwhelming and lacking visual hierarchy (Jiang et al., 2023). In the Graphologue evaluation, participants said diagrams helped them see connections and get information faster, but also reported scrolling, intimidating whole diagrams, distracting animation and latency; the authors note node-link diagrams are not always the best format, and that tables can be clearer for comparisons (Jiang et al., 2023). In Chatbot Arena votes, answer length was the dominant style factor and the effects of markdown lists, headers and bold were smaller; the authors caution that the results are observational and may be confounded (Li et al., 2024a).

Where rendering is limited, a practitioner reports that an ASCII diagram cost about 55 tokens versus about 10 for the same simple diagram in Mermaid, and that Mermaid needs a renderer and becomes hard for humans to read as raw text when complex (Michelini, 2025). Codex CLI turns wide tables into stacked record layouts in narrow terminals because wrapped grids damage cells (Schmalbach, 2026).

Broader connections

Slide sequences and panels built from these elements are covered in Visual presentations, and the use of visual panels to introduce dense documents in Visual storytelling. Layout, hierarchy and colour choices for screens are treated in Interface design. Because diagrams reduce search and keep related items together, the learning findings above connect to the limits on how much a reader can hold in mind at once (Working memory).

Further reading

• Guo and colleagues (2020), a meta-analysis of graphics and reading comprehension: the most direct quantitative evidence that pictures help (Guo et al., 2020).
• Heer and Bostock (2010), which replicated the perception rankings by crowdsourcing (Heer & Bostock, 2010).
• Saket, Endert and Demiralp (2018), on why chart effectiveness depends on task (Saket et al., 2018).
• Wright and colleagues (2018), a study of the confusion arrows cause for learners (Wright et al., 2018).
• Jiang and colleagues (2023), on diagrams as a response format for language-model chat (Jiang et al., 2023).
• Kosara (2010), a critical account of the embellishment study (Kosara, 2010).