What makes content hard to read aloud?

Anything the ear cannot resolve: symbols and abbreviations a synthesiser has to guess at, long sentences that exceed what someone can hold in memory without a visual anchor, and phrases like "see the table below" or "as shown above" that refer to something no listener has.

How long should a sentence be for voice?

Short enough to follow with no ability to re-read, which in practice means around 20 words and rarely past 25. A reader can jump back to the start of a long sentence; a listener cannot, and by the end of a 40-word sentence the beginning is gone.

Does writing for voice hurt the reading experience?

Not at the level that matters here. Shorter sentences, spelled-out units and passages that do not depend on layout are better prose for readers too. It only goes wrong if it is pushed to the point of choppiness, which is a writing failure rather than a constraint of the medium.

Writing a passage that survives being read aloud

Read your best paragraph out loud, in one breath, to somebody who cannot see it and cannot ask you to repeat it. Most of what survives on a screen does not survive that, and the reasons are specific enough to fix.

The short answer

Read aloud, a page loses everything it was doing visually: bold, links, headings, tables and layout all vanish. What survives is sentence order and sentence length. Content written for a voice assistant needs shorter sentences, spelled-out numbers and symbols, and no phrase that depends on something being visible.

RefinementAudit check · Speakable content

Everything the page was doing for you disappears

Written prose leans on the page constantly, and you stop noticing. A parenthesis tells the reader this part is an aside. A comma gives them a beat. A line break says a new idea starts here. Bold says look at this word. A reader who is not looking at anything receives none of that, and what they get instead is a flat stream of words at a constant pace.

Nested clauses are the first casualty. A sentence that suspends its subject through two subordinate clauses before resolving works on a page because the eye can go back. Spoken, the listener is holding an unfinished thought in memory while new words keep arriving, and if the resolution comes late enough they have lost the beginning. They cannot scroll back. The sentence simply did not happen for them.

Punctuation that carries meaning is the second. A dash used to pivot mid-sentence, a parenthetical qualification, a semicolon joining two related statements: all three are read as a pause of roughly the same length, so the distinction you were drawing is not conveyed. If a qualification matters, it belongs in its own sentence. If it does not matter, a passage being read aloud is not the place for it.

Formatting is the third, and the most damaging when it goes unnoticed. A passage whose meaning depends on a bolded word, a list layout or a "see below" is a passage that arrives incomplete, and it arrives with no indication that anything is missing. This is the same failure as a section that cannot be lifted out of its page, with the volume turned up: a listener has no page at all.

The characters a synthesiser has to guess at

Text-to-speech has to turn every symbol into a word, and it makes a decision whether or not the right decision exists. Most of these are invisible in review because the written form is perfectly clear.

spoken-forms.txt

written              what a listener may receive
-------------------  ----------------------------------------------
Dr. Osei             doctor Osei, or D R Osei
St. Mark's Road      saint, or street
3-4 days             three hyphen four, or three to four
1/2                  one slash two, or a half
~500                 tilde five hundred, or about five hundred
$1.2M                one dollar two M, or one point two million
2026-08-14           two thousand twenty six dash zero eight...
IT department        I T, or "it"
US                   U S, or "us"
e.g.                 e g, or "for example"
&                    and, usually
->                   read as nothing at all

The two rows to take seriously are the ones where the wrong reading is a real word. IT read as "it" and USread as "us" do not sound like errors, they sound like a different sentence, and the listener has no way to detect the substitution. There is no markup that fixes this in a passage nominated for reading aloud. The fix is to write "the IT department" as "the technology team", or to accept the risk.

Dates and ranges are the ones that come up most often in practice. Write "three to four days", not a hyphenated range. Write the month out. And write numbers the way you would say them: "about five hundred" rather than a tilde, "one point two million dollars" rather than a compressed currency form. It costs you a few characters in a passage that is deliberately short, which is the tension worth being aware of and the trade worth making.

Sentence length as a memory constraint

The usual advice about short sentences is a style preference. Here it is a mechanical requirement, and the mechanism is worth stating because it changes what you do with a long sentence.

A listener holds the current sentence in working memory until it resolves. Length alone is not the problem: a long sentence built as a sequence of complete clauses is followed without effort, because each clause closes before the next begins. What breaks is depth, where the sentence opens a structure and does not close it for another twenty words. So the fix for a long sentence is usually not to cut it in half. It is to reorder it so nothing is left hanging.

The passage as a whole has a budget too. The guidance for nominated passages is on the order of twenty to thirty seconds, which is roughly two or three sentences, or somewhere near sixty words spoken at an ordinary pace. That is short. It means the passage has to be a self-contained answer rather than an introduction to one, which is the same demand answering first makes, under a harder limit.

The test, and it takes ninety seconds.

Paste the passage into any text-to-speech tool, your operating system already has one, and listen to it once at normal speed without reading along. Reading along defeats the test, because your eye repairs everything your ear misses. Anything you would have had to rewind for is the finding.

Then listen for the three specific things: an abbreviation read as letters, a number read as digits, and a sentence whose subject you had lost by the time the verb arrived. Those three account for nearly every real defect in this check.

Judging a passage takes the passage, not a value

There is no field in a single-page fetch that says whether a paragraph works aloud. It is a judgement about sentences in context: how deeply each one nests, which symbols are in it, whether the abbreviations are the ambiguous kind, and whether the whole thing resolves inside the time a listener will give it.

This one needs the full crawl

Judging whether a passage works aloud means reading its sentences for length, abbreviations and symbols in context, which the page-level audit does across the whole passage rather than from any single value.

The instant search on this site audits a single page, so rather than show you a verdict it cannot support, this guide sends you to the place the check actually runs.

Run the full audit, free

The check that pairs with this one is the markup that nominates the passage. Marking a passage that does not survive being read is worse than marking nothing, because you have volunteered it.

Does this affect AI search?

Yes, and by a route that has nothing to do with voice. Everything this check asks for, resolved clauses, spelled-out numbers, no dependency on formatting, a passage that stands alone in sixty words, is also what makes a paragraph easy to extract and quote. A model lifting a sentence out of your page faces a version of the same problem a listener does: it has the words and not the layout.

The overlap is not complete, and the difference is instructive. A model can handle a nested sentence without difficulty, because it is not holding anything in working memory and can attend to the whole passage at once. So the sentence-depth advice is genuinely about human listeners. The symbol and abbreviation advice, though, transfers directly: an ambiguous US or a hyphenated range is ambiguous to anything reading the plain text, and a quoted fragment that depended on adjacent formatting is as broken in a chat answer as it is in a spoken one.

Which makes this a check with a better return than its rating suggests. Almost nobody is eligible for the voice feature this exists to serve, and the writing it asks for is the writing that gets quoted. Do it for the second reason.

Why prose written for the ear is a Refinement

Because the page works. Nothing here is a defect in the ordinary sense: the passage is accurate, published, indexable and readable by every human being who arrives with a screen, which is nearly all of them. The bottom of the scale is where improvements to a working page belong, and this is one.

The rating is also honest about scope. This is editorial work, done sentence by sentence by somebody who can write, and it cannot be applied by a template or a build step. On a site with nine thousand pages it is a thing you do to the twenty that matter. A check whose fix does not scale should not be asking for a place near the top of a plan.

What stops it being pointless is that the work is not spent on voice. If the only argument were the assistant reading you aloud, this would be a check for news publishers and nobody else. The argument is that a passage which survives being read to a stranger with no screen is a passage that survives being quoted, summarised, or cut out of your page entirely, and those things happen to everyone.

Questions this check raises

What makes content hard to read aloud?: Anything the ear cannot resolve: symbols and abbreviations a synthesiser has to guess at, long sentences that exceed what someone can hold in memory without a visual anchor, and phrases like "see the table below" or "as shown above" that refer to something no listener has.
How long should a sentence be for voice?: Short enough to follow with no ability to re-read, which in practice means around 20 words and rarely past 25. A reader can jump back to the start of a long sentence; a listener cannot, and by the end of a 40-word sentence the beginning is gone.
Does writing for voice hurt the reading experience?: Not at the level that matters here. Shorter sentences, spelled-out units and passages that do not depend on layout are better prose for readers too. It only goes wrong if it is pushed to the point of choppiness, which is a writing failure rather than a constraint of the medium.