typeface-selection

name	typeface-selection
description	Use when choosing or replacing a typeface — "which font should we use", evaluating candidates, building a type palette, or replacing a face that isn't working. Also use when a font must carry a brand tone, survive small sizes, support extra scripts, or fit a licensing and performance budget.

Typeface selection

Selection is a constraint-satisfaction problem, not a taste contest. Taste decides between the candidates that survive the constraints.

1. When to invoke

• Choosing a typeface for a product, brand, or document.
• Replacing a face that fails at some size, script, or medium.
• Building a type palette for a system.
• Auditing an inherited choice.

Do not invoke for combining faces (typeface-pairing) or for expressive critique (typography-critique).

2. Required context

Do not name a font before you have these:

• Roles required — body, UI, display, data, code. Different constraints each.
• Smallest and largest sizes it must work at.
• Scripts and languages required, now and planned.
• Weights and styles genuinely needed (including true italics).
• Figure requirements — tabular for data, oldstyle for prose. Confirm the file actually ships them (../type/references/opentype-features.md).
• Licence and budget — web, app, desktop, redistribution.
• Performance budget — see font-loading-and-performance.
• Brand tone, and any existing system constraint.
• Medium — screen, print, low-DPI, e-ink, embedded.

3. Invariant principles

• Constraints first, taste last. Script coverage, licensing, weights, and figures eliminate most candidates before aesthetics is consulted.
• Judge at the size of use, in the medium of use, with real content.
• Serif vs. sans is not a legibility decision. Research has failed to prove serifs improve reading speed or comprehension; "bias favors familiarity." (Lupton) Choose on tone and rendering behavior.
• A typeface must be judged in running text, not from its logotype-friendly showcase glyphs. Santa Maria: flourishes that look charming in a specimen "stick out in running text."
• The face must survive its worst case, not flatter its best — the longest German compound, the smallest label, the lowest-DPI screen.
• Familiar is not a failure. Familiarity aids reading.

4. Context-dependent heuristics

By role.

Role	Prioritise
Body / long-form	Low-moderate contrast, open apertures, real italic, comfortable at 15–18px, wide language coverage
UI / dense product	Open apertures, distinguishable 1 l I / 0 O, tabular figures, tight vertical metrics, several weights
Display / brand	Character, distinctive forms, optical sizes if available; small-size behavior irrelevant
Data / tables	Tabular figures mandatory; unambiguous digits
Code	Monospace, disambiguated 0 O l 1 I, clear punctuation

Reading tone from form. Coles' value is explaining why a face reads a certain way, from evidence:

Wanted tone	Visible signals to look for
Unobtrusive, systematic	Neo-Grotesque; regular proportions, low contrast, closed-ish apertures
Technical, precise	Geometric or Grotesque sans; uniform stroke, circular rounds
Humanist, warm	Angled axis, calligraphic trace, open apertures
Editorial, authoritative	Transitional/Rational serif; vertical axis, higher contrast
Contemporary, sturdy	Large x-height, low contrast, chunky serifs, very open apertures
Expressive, characterful	Distinctive terminals, unusual proportion

Say the evidence out loud. "Feels modern" isn't a reason; "geometric construction, near-uniform stroke weight, circular bowls" is.

Do not say "neutral." Rutter: "There's no such thing as a neutral typeface… You cannot wear neutral clothes. Boring, predictable, unadventurous and inoffensive, perhaps; but never neutral." Every choice carries a voice, and calling one neutral hides the decision. When a stakeholder asks for "something neutral," translate it: they want a face that gets out of the way and lets the ideas shine through — which Rutter notes is a demanding brief, since "any successful boring typeface will have been carefully designed to be highly readable." Say "unobtrusive" and mean it.

Classification analyses; genre finds. There is no universal classification system, and no vendor lists faces under the descriptive terms designers use — you won't find "linear dynamic sans serif" as a browsable category. Use classification (typeface-anatomy) to describe and defend a face from evidence; use genre, description, and specimen browsing to discover candidates.

Make a specimen to reduce the shortlist. Set the same real content in every candidate, at the real sizes, and compare them side by side in one document. This is the step people skip, and it is the one that actually eliminates candidates.

Building a palette (Santa Maria). Keep a small personal set you know deeply rather than searching fresh each time. Knowing how a face behaves at 12px, in italic, in long German words, and under a fallback swap is worth more than novelty.

Superfamilies — a serif and sans sharing skeleton and metrics — solve pairing and metric-matching at once. Strong default when a project needs both.

Variable fonts collapse many weights into one file. See variable-fonts.

How the face renders on screen is part of the choice. Rasterisation is lossy and varies by engine, only Windows uses hinting, and handhelds use greyscale smoothing rather than subpixel anti-aliasing — so check your type across platforms rather than judging on your own machine. Delicate high-contrast faces and hairline weights are the first casualties on low-DPI displays. Full detail in typeface-anatomy ("How letterforms become pixels").

⚠️ One piece of the 2017 advice is stale. Rutter recommends disabling subpixel anti-aliasing (-webkit-font-smoothing: antialiased) only for reversed-out text, because light-on-dark rendered "an entire weight bolder" on low-resolution macOS. macOS removed subpixel anti-aliasing system-wide in Mojave (2018), so the specific condition no longer holds; the property is now used for consistency with the OS, not as a reversed-text fix. Verify current behavior before applying either. The underlying design point survives: reversed text reads heavier, so test it.

System font stacks cost zero bytes and render natively. Legitimate for performance-critical and utility interfaces; they forfeit brand distinctiveness and render differently per platform.

5. Failure patterns

Pattern	Cause	Fix
Beautiful in the specimen, muddy at 14px	Judged at display size	Re-evaluate at the real size
Lovely until German/Polish shipped	Coverage unchecked	Verify glyph coverage per language
Numbers jitter in the table	Proportional figures	Require tabular figures
Faux italic renders	No true italic in family	Check styles, not just weights
Licence blocks app embedding	Web-only licence	Confirm every medium up front
Face "feels off" in body, fine in headings	Display face used for text	Separate display and text roles
Distinctive flourishes distract in paragraphs	Judged on isolated glyphs	Judge in running text
Chosen for readability because it's a serif	Folklore	Choose on tone and rendering
Nine weights shipped for a 4-weight design	No inventory	Ship only what renders

6. Evaluation procedure

1. Write the constraint list (§2). Eliminate anything failing a hard constraint — script coverage, licence, required styles, tabular figures.
2. For survivors, render real content — the actual longest string, the actual smallest label — at the real sizes in the real medium.
3. Apply the typeface-anatomy evaluation procedure to each: x-height, contrast, aperture, counters, 1 l I, 0 O.
4. Set a full paragraph in each and read it. Do any glyphs distract?
5. Test the worst case: longest locale word, densest table, smallest size, lowest DPI, reversed on dark.
6. Check the fallback: how far are its metrics from your candidate? (font-loading-and-performance)
7. Check the payload for the weights you'll actually ship.
8. Only now compare on tone — and state the visible evidence for the tone claim.

7. Output format

Roles: <body|ui|display|data|code>
Hard constraints: scripts <…> · styles <…> · figures <…> · licence <…> · budget <…>
Candidates eliminated: <face> — <which constraint>
Recommended: <face>
  Evidence: x-height <…>, contrast <…>, aperture <…>, digits <…>
  Tone claim: <tone> — because <visible signals>
  Worst case tested: <what>
  Payload: <n> files, <n> KB
Runner-up: <face> — <why not>
Unverified: <what wasn't rendered>

8. Examples

Fintech dashboard, dense tables, EN + DE + PL, 4 weights, dark mode.

Hard constraints eliminate most candidates immediately: Polish requires ą ć ę ł ń ó ś ź ż, tables require tabular figures, and dark mode means the lightest weight must survive reversal. That leaves a Neo-Humanist or Grotesque sans with a full Latin Extended set. Recommending Inter: open apertures keep c/e/s distinct at 13px, 1 l I and 0 O are disambiguated, tabular figures ship, and one variable file covers all four weights. Verified with the longest German compound in the narrowest column and with the 300 weight reversed on the dark surface — 300 washed out, so the minimum shipping weight is 400.

9. Counterexamples

• ❌ "Use a serif for body text, it's more readable." — Unevidenced (Lupton).
• ❌ "This font is trending, let's use it." — Trend is rule 9, the bottom of the hierarchy.
• ❌ "It looked great in the specimen." — Specimens are set to flatter.
• ❌ "We'll add language support later." — Retrofitting script coverage usually means changing the face.
• ❌ "Ship all nine weights so we have options." — Every unused weight is pure cost.
• ❌ "Helvetica is neutral, so it's the safe choice." — Neo-Grotesque closed apertures actively hurt small-size UI legibility.
• ❌ "The brand deck says 'modern and human', so use a geometric sans." — Geometric construction reads technical; humanist forms read human. Match evidence to the claim.

10. Source citations

• Coles, The Anatomy of Type — selecting type on visible evidence; explaining why a face reads technical, humanist, editorial, neutral, or expressive; comparing similar typefaces; ~100 analysed faces with designer, foundry, and classification.
• Santa Maria, On Web Typography — evaluating typefaces; building a personal type palette; matching type to content and tone; flourishes that read fine in isolation and "stick out in running text"; balancing aesthetic and technical constraints.
• Lupton, Thinking with Type 3e — serif readability research is inconclusive; familiarity bias; typeface classification.
• Rutter, Web Typography — choosing typefaces for screens; apparent size and x-height compensation; OpenType and figure styles; "there's no such thing as a neutral typeface"; "type genres are more useful than classifications"; make a specimen to reduce your shortlist; judge typefaces in context; choose robust faces for readability on screens; understand how fonts render on screens and check type across platforms; clearly distinguish functional text from body text.
• Latin, Better Web Typography — choosing typefaces; language-range considerations.
• Stocks, Universal Principles of Typography — classification's uses and limits; width as a palette tool; "type should be legible" as a first principle.