181k visual atoms, nine subtypes. Logos and photos tie at 45k each. Headshots cluster in pitch decks. Screenshots argue harder than they ship. Each subtype carries its own editorial job — call them by name.
Logo45k · 13%endorsement — proof someone trusted you
Photo45k · 11%mood / story — when the human matters more than the data
Illustration22k · 6%concept made visual — abstract ideas given a shape
Headshot21k · 3%speaker / expert — voice given a face
Screenshot20k · 4%product / proof — "this is what it looks like"
Logo grid10k · 3%trust strip — pluralised endorsement
Icon grid9k · 2%feature list — visual bullet points
Map6k · 2%geography — when location IS the variable
Infographic2k · 1%visual data — chart and diagram had a child
§ 6.c.1 · subtype contrast
Logos endorse. Headshots persuade. Photos seduce.
The frequency rank tells one story; the editorial rank tells another. Three editorial roles, nine forms — pick the role first, the form follows.
3:2:1argue · narrate · decoratethree editorial roles, nine forms
Arguelogo-grid · screenshot · headshot"someone trusts us. someone tested it. someone vouches for it." Imports outside credibility into the slide.
Narratephoto · illustration · mapsets the scene the words sit on. Mood, place, era. Earns its slot when the words alone undersell the situation.
Decoratelogo · icon-grid · infographicfills space the layout could not solve. Most corpus images live here. Cheap to add, expensive to ignore.
03subkind face-off
Same claim. Three image choices. Three different reads.
"Our customers love us" sits above a photo, a screenshot, and a
logo grid. The words are identical; the slide is not. Each image
earns a different kind of trust — and skips a different kind. Pick
by what the audience needs to leave with, not what is closest to
hand.
The shared claim
"Our customers love us."
PhotoMarketing☻A smiling stock-photo customer
Earns
mood — sets the emotional register.
Fails at
verifiability. Could be anyone's customer, including no one.
ScreenshotProof▦The actual product UI dashboard
Earns
specificity. The reader sees what the customer sees.
Fails at
mood. The dashboard does not feel — it informs.
Logo gridEndorsement▢▢▢12 customer logos at scale
Earns
social proof at scale. Twelve names beats one.
Fails at
depth. No customer's story is told — only their badge is shown.
04subtype geography
Each subkind has its own canvas.
Logos are small and centred (16% × 9%). Photos take the hero slot (48% × 62%). Headshots cluster slightly upper-left (17% × 23%). Each subkind owns a different geometry — and the editorial role follows the geometry the corpus already chose.
Images get the lowest source-note coverage of any visual family. Photography is the corpus's silent uncited claim — the audience accepts what the chrome will not specify.
14ppmissing attribution38% of slides have an image · 24% pair it with a source
Images sit at the bottom of the corpus trust hierarchy. Of every 100 slides that ship an image, only 24 ship a source-note. The other 76 import their claim without provenance — logos, headshots, screenshots, maps — all uncited.
Charts get sourced 56% of the time, tables 50%, metrics 47%. The pattern: numbers force attribution; pictures escape it. The reader is trained to demand a source on a number and forgive its absence on a face.
An undated screenshot is a wireframe. An unsourced logo is name-dropping. The image needs the same chrome the chart already earned.
Source-note coverage by family · % of slides citing
Charts
56.5%
Tables
49.5%
Metric
46.6%
Text
37.7%
Diagrams
29.2%
Images
24.3%
06position
Images sit centre-mass on the slide.
Across 181k image atoms with measured bounding boxes, the visual takes the middle. Logos and headshots cluster wider; screenshots ride the upper-half rail.
modal image · 73×69%
leftcenterrighttopmiddlebottom
i
How to read this
Heatmap
Surface coverage. Each bbox in the corpus contributes to every 40×22 cell it overlaps; opacity is power-scaled (γ=0.6) so subtle bands stay visible. Darker = more atoms cover that point.
Sample box
The modal cluster. Atoms grouped by centre into 8×8 bins; the densest bin is plotted at its centroid, sized by the average width × height of atoms in it. Reads as: "the typical atom of this family looks roughly this big, here."
Source
corpus.components.bbox — atoms with measured bounding boxes only. Inline mentions extracted from body text are excluded from this analysis.
07slide-type cross
Images open the deck. They almost never close it.
Image presence per slide-type peaks at the start (cover 89%, team bio 91%, section divider 55%) and collapses at the end (data table 5%, appendix data 7%, disclaimer 29%). Images are ceremonial — they greet, divide, introduce — not argumentative. The corpus uses them to set tone, not to prove.
91%peakTeam bio
29%medianDisclaimer
5%floorData table
slide-type18 most-frequent · n = 189,105
Text
Images
Charts
Diagrams
Tables
Metric
Chrome
Team bio5,245
78%
91%
3%
2%
4%
5%
14%
Cover15,202
60%
89%
0%
0%
0%
0%
15%
Solution8,470
83%
63%
11%
38%
4%
6%
23%
Section divider14,194
21%
55%
0%
0%
0%
0%
16%
Propose solution5,667
87%
51%
14%
33%
6%
9%
53%
Key messages13,445
91%
50%
8%
13%
3%
8%
29%
Traction10,066
65%
36%
57%
6%
11%
28%
54%
Executive summary6,818
90%
35%
12%
6%
10%
17%
45%
Key takeaways10,052
89%
34%
24%
6%
4%
17%
44%
Disclaimer6,523
30%
29%
1%
0%
0%
0%
89%
Expose contradiction11,726
96%
24%
21%
4%
33%
1%
77%
Market sizing5,135
70%
24%
57%
14%
12%
33%
78%
08co-occurrence
Image-bearing slides under-source.
Images travel with text 2:1 — paragraph (40%) and list (34%) are common companions. But only 24% cite a source — even though logos, headshots, screenshots, and maps all import claims that need attribution. Images get a free pass on traceability the corpus does not extend to charts.
01Title
93%
almost universal — image without title is rare
02Paragraph
40.4%
narrative scaffolding — image illustrates the story
03List
34.4%
image as bullet anchor — feature lists, capability grids
04Source-note
24.3%
— only one in four. Logos, screenshots, maps cite NOTHING.
05Callout
9.2%
image + callout = "look at THIS detail"
06Metric
8.7%
image + number — the marketing default
07Disclaimer
7.9%
usually a stock-photo notice or copyright line
08Chart
7.6%
rare — image and chart compete for the visual focus
09anti-patterns
Five image failures the corpus repeats.
Images get less editorial scrutiny than charts because they feel cheaper. The corpus shows the cost of that assumption — five recurring failure modes, three severe.
◇
The stock decoration
A photo of nobody, doing nothing, related to the topic by mood only.
~30%
Bad
Stock photo of "diverse team smiling at laptop" next to a slide titled "Our culture".
Fix
Cut. Or replace with a real photo of the actual team, captioned by name.
◯
The logo without context
A client logo with no relationship disclosed — partnership? customer? vendor?
13% of logos
Bad
Wall of 12 logos under "Our partners" — no dates, no scope, no testimonial.
Fix
Disclose the relationship. "Customers since 2020 · serving 50k seats" beats a logo grid.
⊟
The undated screenshot
A product UI shot with no version, no date, no source — could be a mockup.
76% of screenshots
Bad
Screenshot of dashboard captioned "Our product".
Fix
Add date + version + source. "v2.3 · captured 2024-09-12 · production".
⊙
The headshot soup
Six smiling faces in a row, none of which the audience can place.
12% of headshots
Bad
Pitch-deck team slide with 8 headshots, no titles, no context.
Fix
Three headshots maximum, each with title and one credibility line.
◢
The oversized illustration
A whimsical drawing taking 60% of the canvas, the argument hidden in 30%.
8% of illustrations
Bad
Cute hand-drawn metaphor dominating a slide where the chart is the point.
Fix
Shrink the illustration to 20%. Let the chart breathe.
10safe zone
Three rules. One photo per slide.
Images are the cheapest editorial vehicle in the deck — and the most often misused. These four rules catch the failure modes the corpus repeats by default.
01
One image per slide. Maximum.
Two images compete for the focal point and split attention. The audience cannot read both. Pick the stronger one and cut the rest.
25%slides ship 2+ images
02
If it has a UI, date it
Screenshots are evidence — but only when versioned. An undated screenshot is a wireframe pretending to be a product.
76%screenshots undated
03
If it has a logo, disclose the relationship
"Customer", "investor", "partner" — three different things. A logo without disclosure is name-dropping, not endorsement.
13%logos undisclosed
04
If it has a face, name it
Headshots without a name are decoration. Headshots with a name and a credibility line (title + tenure + relevance) become endorsement.
