Visual storytelling

Visual storytelling, in the sense used here, is the practice of turning a dense document (a research paper, a strategy, a design document) into a short sequence of visual panels, often about ten, that a reader goes through alone and without a speaker, and that prepares the reader for the full text rather than replacing it. The field it draws on is wide: narrative visualisation in information design, the learning-science literature on pictures, organizers and signalling, and practitioner writing on slides and memos. Its main disputes are how much a visual adds over well-written prose, whether the evidence for narrative framing carries over to dense scientific content, and whether the book of recommended structures has been tested on readers at all. Most of it has not: genre taxonomies and story structures are design analyses built from examples, not tested on readers (Segel & Heer, 2010).

History

Three works frame the field. Larkin and Simon (1987) asked why a diagram is “sometimes” worth ten thousand words, and answered that for the same information, the value of a representation depends on the operations a reader can perform on it, such as recognising a pattern, drawing a direct inference or searching, not on the information it holds (Larkin & Simon, 1987). Shneiderman (1996) classified visualisations by seven data types and seven tasks (Shneiderman, 1996), and Heer, Bostock and Ogievetsky (2010) toured techniques by data type: time series, statistical data, maps, hierarchies and networks (Heer et al., 2010). Segel and Heer (2010) then looked at 58 examples of stories told with data and named seven genres and three structures (Segel & Heer, 2010).

The history of the practice of introducing documents with visual panels, before narrative visualisation, is missing.

Core concepts

When a picture does something prose cannot

Meta-analyses find that graphics help, but modestly. Across 39 studies, adding graphics to text raised reading comprehension by a moderate amount (Hedges’s g = 0.39); pictures, pictorial diagrams and flow diagrams did not differ significantly from one another, and graphics helped more on open-ended assessments than on true/false ones (Guo et al., 2020). In problem solving, pictures gave a small-to-medium accuracy gain overall, but only organizational pictures helped; informational pictures and multiple pictures had no significant aggregate effect (Hu et al., 2021). The Larkin and Simon account explains why such results vary: a diagram is only sometimes better, because it depends on whether the diagram makes the needed recognition, inference or search direct (Larkin & Simon, 1987).

Detail can work against the reader. In a study summarised by a secondary source, a detailed, colourful diagram produced the same learning of the taught case as a plain one but weaker transfer to other species (Watson, 2020). Adults who saw a plainer diagram of metamorphosis transferred more accurately than adults who saw a rich one (Menendez et al., 2020). In three experiments, students taught a rule with simple abstract symbols transferred it better than those taught with concrete objects (reported from a press release; symbols, not diagrams) (Grabmeier, 2005). A meta-analysis of 58 studies (more than 7,500 students) found lower learning outcomes when interesting but irrelevant “seductive details” were added, naming prior knowledge as an open moderator (Zaske, 2020).

Prior knowledge also changes the sign of help. The expertise reversal effect is the finding that guidance that helps novices may have negative consequences for more experienced learners; the page describing it notes that many demonstrations rely on subjective load measures (Wikipedia contributors, 2026d). The harm of a rich diagram was reported to be smaller for learners who already knew the topic (Watson, 2020). Spatial ability evidence is thin: in 6-year-olds, high-spatial-ability children learned better from animation than from static pictures, while low-spatial children gained no particular benefit (Ben Chikha et al., 2021).

The range of representation forms

Lengler and Eppler’s taxonomy reaches beyond charts: they organise over a hundred visualisation methods into groups that include concept, metaphor, strategy and compound visualisation (Lengler & Eppler, 2007). Metaphor forms, such as the metro map and the story template, organise information and convey an insight through the characteristics of the metaphor; strategy forms, such as the strategy canvas and technology roadmap, are management-specific; compound forms combine several formats in one frame (Lengler & Eppler, 2007). The authors call their table an aid for choosing and combining methods, not an exhaustive list (Lengler & Eppler, 2007).

Evidence on how well readers understand each form is uneven, and is reported here by family.

Quantitative forms. Quantile dotplots of 50 outcomes gave near-optimal and more consistent decisions in an incentivised bus-timing experiment (Fernandes et al., 2018). Violin and gradient plots led lay viewers to judgements more consistent with statistical inference than bar charts with error bars, which showed a within-the-bar bias; participants followed the expected strategy 89.2% (violin) and 88.5% (gradient) of the time against 83.2% for bars (Correll & Gleicher, 2014). Reanalysis of a correlation-ranking study found scatterplots combine low variance between people with high precision (Kay & Heer, 2016). For dynamic graphs, small multiples were faster than animation, but animation was more accurate on some tasks (Archambault et al., 2011). Rescaling large numbers to smaller ones helped people discriminate and rank government expenditures (Boyce-Jacino et al., 2022). Experts read 2x2 interaction bar and line graphs equally well, unlike novices, which is only indirect evidence for slope-style line encodings (Peebles & Ali, 2015). The main source recommending diverging stacked bars for Likert data is a design argument, and no experiment on reader comprehension was seen on its page (Heiberger & Robbins, 2014).

Structural forms. For sets, a crowdsourced comparison found linear diagrams fastest and most accurate and Venn diagrams worst (Chapman et al., 2014); for showing set sizes, area-proportional Euler diagrams with numbers were most effective overall, while for plain cardinality tasks linear diagrams were faster without being less accurate (Stapleton et al., 2019). In a 12-participant study, treemaps were slower than icicle and sundown charts on hierarchy tasks and less preferred, with accuracy differing significantly only on one hierarchy-relationship question (91.4% vs 97.2%) (Woodburn et al., 2019). Accuracy on alluvial (Sankey-type) diagrams falls as timesteps, entities, flow arcs and crossings increase, and varies strongly by task (about 88% for the most active timestep against about 57% for the largest entity), on synthetic data (Arunkumar et al., 2021). A replication of the node-link versus matrix comparison found significant differences between them, so neither is universally better (Okoe et al., 2017).

Concrete and depictive forms. Instructional guidance holds that illustrations such as cutaways and exploded views reveal structure a photo cannot; this is advice, not tested evidence (Winkelman, n.d.). Seventh-graders who studied text with maps showed a consistent but not statistically significant advantage over text alone (Ramirez & Gilmartin, 1996). Data comics have some direct support. In two studies, participants preferred data comics for engagement and the authors report better understanding and recall than with infographics and illustrated text, with preference stronger than understanding in the wording (Wang et al., 2019). In a between-group study of 98 participants an interactive data comic about cancer data was rated more engaging than a text version, and participants recalled and understood the data better (Fogwill & Manataki, 2024).

Interactive forms. Interactivity raises engagement more reliably than understanding. A scrollytelling format for privacy policies matched other formats, including plain text, on comprehension accuracy while improving experience (Mendez & Such, 2026). In a controlled study of 605 participants, interactive simulations improved AI literacy, transfer and self-reported confidence but did not differ significantly from a static PDF tutorial, and engagement alone did not predict learning (Zhao et al., 2025).

Text-based diagrams. No controlled study compares comprehension of ASCII or Unicode diagrams with rendered graphics. The nearest evidence is a command-line study in which participants preferred an interactive ASCII dependency graph to the system’s two existing approaches (Isaacs & Gamblin, 2019); the RFC Editor’s rule that plain-text RFCs carry only ASCII art within 72-character lines, with other figures left to the HTML version (Flanagan, 2016); and mailing-list opinion, where one OpenStack developer judged ASCII poor for user-facing documentation while others use it in specs and code documentation (Gordon, 2015). A terminal diagram renderer shows the practical conventions: a column-width budget and an ASCII-only fallback because box-drawing glyphs are not universally displayed; its Unicode text has gaps (dotted junctions render as solid, and very narrow widths can overlap boxes), and it supports flowcharts, state, sequence diagrams and pie charts but not Gantt, journey, ER or class diagrams (leboiko, 2026). Language models read ASCII layouts much more reliably than they generate them: even a reasoning model failed to construct accurate layouts in over half of cases, and interpretation relies on local pattern recognition, which limits what large or global diagrams convey (Huang et al., 2026).

Building a sequence

Segel and Heer derived seven genres of narrative visualisation from 58 examples, varying by number of frames and ordering of visual elements; they say no genre is right for every case and that the taxonomy is subjective, built from examples rather than reader tests (Segel & Heer, 2010). They also describe three structures on an author-driven to reader-driven scale. In the martini glass, an author-driven path comes first and the visualisation then opens to free exploration. The interactive slideshow is more balanced, each slide often stating its message before inviting interaction. In the drill-down story the author gives a theme and the reader chooses instances; it is more reader-driven but still needs heavy authoring (Segel & Heer, 2010). Bach et al. propose narrative design patterns, each with a purpose such as engaging the audience or evoking empathy, usable alone or combined (Bach et al., 2018).

Sequence itself has some evidence. Hullman et al. found that sequences using parallelism as a structural device improved memory and subjective ratings, and that transition types can be ranked by audience cost; in 42 professional narrative visualisations, general-to-specific transitions were the most common (28 of 42), followed by dimension walks (20) and measure walks (19) (Hullman et al., 2013). Animated transitions between related charts were more accurate than static switches and staged animation usually did better still, though not for every chart type; readers preferred staged to plain animation to static transitions, but for one chart (a donut) plain animation beat staged (Heer & Robertson, 2007). The authors name keeping viewers oriented as the core problem, which matters most when viewers are not interacting (Heer & Robertson, 2007).

A slideshow layout beat a vertical scroll layout on comprehension in a 180-person crowdsourced study of text-plus-chart stories, and linking text to chart elements raised engagement and, with the slideshow layout, recall (Zhi et al., 2019). Obie et al. compared purely author-driven narratives with interactive visualisations that had no author narrative, and found that narration aided understanding of information but had no significant effect on long-term recall (Obie et al., 2020).

Preparing the reader

A sequence introduces a document, so the evidence on previews and overviews matters. Instructor-provided graphic organizers for learning from text improved comprehension (g = 0.53, 49 comparisons) and memory (g = 0.70, 37 comparisons) in a 2025 meta-analysis, which warns of possible publication-bias inflation and found no significant moderators, so it does not say which readers benefit more (Ponce-Arias & Mendez, 2025). A synthesis for students with learning disabilities found graphic organizers gave greater reading-comprehension gains than comparison conditions, with large effects in the two studies of framed outlines (Kim et al., 2004). A meta-analysis of 55 studies of viewing or building node-link concept maps found better knowledge retention that ranged from small to large with the comparison treatment (Nesbit & Adesope, 2006).

Signalling structure helps. Teaching fourth graders to attend to preview statements produced higher performance than structure-cueing or no instruction (Spires et al., 1992). For college students, headings gave better memory for the subtopics of a text than preview sentences carrying the same information, with no difference for facts, and texts with signals to topic structure were outlined better than texts without (Lorch et al., 2013). A meta-analysis of signalling in multimedia learning found a small-to-moderate positive effect on learning outcomes (d = .38), larger in higher-quality studies and those controlling for prior knowledge (Alpizar et al., 2020). Questions posed before material were reported to raise later test scores on questioned items, with the limits that the videos were short, the test immediate and the benefit gone after a month in a classroom study (secondary source) (Nebel, 2017).

What makes a single panel memorable is partly separate from what makes it understood. In a memory study of 2,070 single-panel visualisations, colour and a human-recognisable object raised memorability, common graph types were less memorable than unique ones, and observers agreed on which were memorable, so memorability is a stable property of the image (Borkin et al., 2013).

Questions, controversies, and new developments

Does narrative framing help dense content?

A meta-analysis of over 75 samples found stories understood and recalled better than essays, with a moderate overall effect (g = .55), larger for memory (g = .72) than comprehension (g = .48), and no significant moderators including whether difficulty or content was controlled across genres; prior knowledge was not tested as a moderator (Mar et al., 2021). A review of 36 studies found encouraging evidence for science narrative texts at various educational levels, especially in delayed assessments and longer interventions (Soares et al., 2023). In three experiments, scientific explanations preceded by narrative introductions were recalled better than ones with expository introductions, but took longer to read (Mensink et al., 2021). Whether the benefit holds for dense scientific or business content is unsettled.

Do visual abstracts and comics improve understanding?

The evidence on visual abstracts and comics is on reach and preference. In a randomised crossover trial of 50 publications, tweets with a visual abstract got more impressions, retweets and link clicks than tweets with a text-abstract screenshot, but understanding and memory were not measured (Hoffberg et al., 2020). In an online randomised crossover survey of 1,236 people, most preferred a comic over a lay summary or a scientific abstract; the outcome was preference, not comprehension (Kearns et al., 2022). Observationally, in one journal, papers without a graphical abstract got more PDF downloads, abstract views and citations than papers with one (Pferschy-Wenzig et al., 2016).

Slides for readers, not listeners

Sources propose structures for documents read without a presenter but test none on readers. Minto’s method organises ideas as a pyramid under one point and identifies the reader’s question through a Situation, Complication, Question framework (Minto, 2007). Amazon replaces slide presentations with narratively structured six-page memos read silently, with no measured outcome given (Amazon, 2017). A practitioner view is that a report read without a presenter must stand alone and can carry more text, while a live version should reveal one element at a time (Knaflic, 2013). How slides were authored matters when they are read without the speaker: detailed and big-points-only styles changed their usefulness as summaries (Grudin et al., 1999). Critics argue from opinion that standard slide templates weaken reasoning (Tufte, 2006), and the Columbia investigation board is quoted as warning that a senior manager could read a dense slide and miss that it concerned a life-threatening situation, an anecdote and not an experiment (Thompson, 2003). In randomised business-case experiments, PowerPoint and oral presentations were rated alike, with no evidence audiences remembered or comprehended the case differently; it concerns talks, not documents read alone (Moulton et al., 2017).

Generated panels and slides

Systems now generate such outputs. Paper2Poster summarises paper sections into an asset library, matches summaries to figures and refines layout with visual feedback, judging outputs by having vision-language models answer quiz questions from the poster alone; it reports that GPT-4o posters look good but communicate poorly on the quiz, with engagement the main bottleneck (Pang et al., 2025). PPTAgent outlines slides, pairs each with a reference slide and edits it by executable actions, scoring Content, Design and Coherence with an LLM judge that correlates with human ratings at 0.71 on average (PPTAgent authors, 2025). Quality is thus judged mostly by models, not by readers’ learning.

Author-driven or reader-driven?

Recommendations for author-driven stories rest mainly on design analysis, and the head-to-head evidence is a few studies: the slideshow-versus-scroll study (Zhi et al., 2019) and the narration study, in which the comparison was with interactive displays lacking narration (Obie et al., 2020).

Broader connections

The article overlaps three others. Visual presentations covers short slide sequences in general, including assertion-evidence headlines and generated slides. Visual representation of information covers how readers perceive chart encodings and when visuals beat text for learning. Writing for comprehension covers the prose side, where signalling and structure are also studied. The Larkin and Simon criterion, that a representation’s value lies in what it lets the reader do directly, is the link between them.

Further reading

• Segel and Heer (2010), the genre and structure taxonomy for narrative visualisation.
• Hullman et al. (2013), the main study of ordering and transitions in a sequence.
• Larkin and Simon (1987), the account of when a representation is worth more than another with the same information.
• Guo et al. (2020), the meta-analysis of graphics added to text.
• Lengler and Eppler (2007), a taxonomy that includes concept, metaphor and strategy forms.
• Mar et al. (2021), the meta-analysis of stories against essays.