There isn't one. Legibility is a relationship between letters and a particular reader under particular conditions, and among competently drawn text faces the differences are small — small enough that they are routinely swamped by size, line length and spacing. A well-set page in Arial beats a badly set page in anything.
That is an unsatisfying answer to a question people ask sincerely, usually because they are trying to do right by someone. So here is the useful version: what actually moves the needle, what demonstrably doesn't, and a test you can run in ten seconds.
The serif versus sans argument is mostly noise
You have heard that serifs guide the eye along the line, and that sans serifs are cleaner on screens. Both claims are repeated with total confidence and neither survives contact with the evidence. Studies comparing the two have been run for decades and have failed to produce a consistent winner on reading speed or comprehension.
What they do produce is a pattern worth noticing: whenever a study finds a difference, it is usually smaller than the difference you get from bumping the type up two points. The variable everyone argues about is not the variable doing the work.
Use a serif or a sans because of what the page should feel like. That is a real reason. Just don't tell yourself it is a legibility decision.
The uncomfortable part: dyslexia fonts
OpenDyslexic was released in 2011 by Abbie Gonzalez, with weighted bottoms on the letters intended to stop them flipping or rotating for readers with dyslexia. It spread fast, into schools and e-readers and accessibility toolbars, on an idea that sounds obviously right.
The research has not backed it up.
- In 2013, Luz Rello and Ricardo Baeza-Yates used eye tracking with Spanish readers aged 11 to 50 who had dyslexia. OpenDyslexic did not significantly improve reading time, and did not shorten fixations.
- In 2016, Kuster and colleagues asked readers with dyslexia which face they preferred among Arial, Times New Roman and dyslexia-specific designs. They preferred Arial.
- In 2017, Wery and Diliberto tested children with dyslexia across Times New Roman, Arial and OpenDyslexic in Annals of Dyslexia, and found no consistent gain for the dyslexia-specific face.
- In 2023, Broadbent compared Arial and OpenDyslexic with dyslexic and non-dyslexic students. Of those who expressed a preference, most chose OpenDyslexic — largely on how it looked. Test scores showed no difference by font.
That last finding is the one to sit with. People liked it, and it did not help them. Preference and performance came apart, which is exactly the trap that makes design decisions feel obvious and be wrong.
None of this means the fonts are fraudulent or that people who prefer them are mistaken about their own comfort. It means the specific claim — that these letterforms measurably improve reading for dyslexic readers — has been tested repeatedly and has not held up. And the same body of work found something more actionable: plain, unadorned faces did better than decorative or italic ones. The win was in the absence of noise, not in the presence of a special trick.
What actually moves the needle
In rough order of how much they matter:
- Size. Nothing else in this list comes close. Most text on the web is set smaller than it should be because designers evaluate it at arm's length on a good monitor while young.
- Line length. Somewhere around 45 to 75 characters. Too long and the eye loses the return; too short and it breaks the rhythm every few words.
- Line spacing. Near 1.5 for body text. Tight leading makes the eye pick up the wrong line on the way back.
- Contrast. Grey text on white looks refined in a mockup and fails a tired reader on a phone in daylight.
- Character distinction. The first genuinely typeface-level item, and the one people never check. More below.
- Which family you picked. Last. Genuinely last, among faces drawn for text.
Three settings that cost you more than any font choice
If the typeface is the last thing on the list, these are near the top, and all three are decisions people make for reasons that have nothing to do with reading.
All caps. Capitals throw away word shape — the pattern of ascenders and descenders that lets a reader identify a word before resolving its letters. It is the same reason American freeway signs moved to mixed case in 1958 and British ones followed, a change made on legibility grounds after testing rather than taste. A heading in caps is a stylistic choice with a real cost; a paragraph in caps is just harder to read.
Italics for long passages. Italic is a change of voice, not a body style. Set more than a few sentences in it and reading slows measurably. This is one of the few places the dyslexia research is unambiguous and positive: plain upright faces outperformed decorative and italic ones, which is a finding about restraint rather than about any particular font.
Justified text without hyphenation. Justification without a hyphenation engine opens rivers of white space through the paragraph, and the eye follows them downward instead of along the line. Left-aligned ragged-right is almost always the safer setting on the web, where hyphenation support is still uneven.
Why people ask this question
Almost nobody asks it idly. It usually arrives attached to a real obligation — a form that older patients have to fill in, a school handout, a set of instructions for people who will be tired and in a hurry. The question is a proxy for how do I not fail this reader.
Which is why the font-name answer is worse than unhelpful. It converts a solvable problem into a shopping decision, and it lets everyone stop before they have done the part that would have worked. A document set at 12px in a beautiful accessible typeface is still set at 12px.
If someone hands you that brief, the useful reply is a question back: how big, how wide, how far away, and who is reading it in what light? Every one of those has a bigger lever attached to it than the font menu does.
The ten-second test
Set this string in the face you are considering, at the size you will actually use:
If the first three are the same shape, you have a face that will turn Illinois into a picket fence and make every password, licence key and serial number a guessing game. If the O and the zero are indistinguishable, the same applies to anything numeric. If rn closes up into something like an m, you will get modern read as modem at small sizes.
This is the check that actually separates faces, and it takes ten seconds. It is also the one place where the choice of typeface genuinely beats a setting decision — no amount of extra leading will rescue a font whose l and 1 are the same drawing.
Where the real accessibility work is happening
Atkinson Hyperlegible, released by the Braille Institute in 2019 and named for its founder J. Robert Atkinson, is instructive precisely because it does not claim a trick. It was drawn for readers with low vision, and its method is to push confusable characters further apart — some letters exaggerated well beyond what you would draw in a normal text face, because recognition matters more than evenness. It won Fast Company's Innovation by Design award for graphic design that year.
Note what it optimises. Not flow, not weighting, not some theory about how the eye moves. Just: make sure this character cannot be mistaken for that one. That is the mechanism legibility research keeps returning to, and it is the same mechanism Germany's licence plate typeface arrived at from a completely different direction, trying to stop a forger with a roll of tape.
What a type designer takes from this
Two things, and the first is humbling if you draw type for a living.
The typeface is not the lever. If someone asks which font to use for a document their whole team has to read, the honest answer is to set it bigger, give it a sane measure, loosen the leading, and stop worrying about the family. Recommending a font feels like helping. Fixing the setting is helping.
The second is where a typeface does earn its keep: the confusable pairs. That is a drawing decision, it cannot be fixed downstream, and it is the reason Vigil keeps the tail on its l, the flag on its 1, and open counters at text sizes — the whole point of rationing the rounding rather than softening every corner. A face that blurs those is not a stylistic preference. It is a face that will cost somebody a wrong digit at the wrong moment.