An AI readability audit measures the things a tired reader feels.

Five properties predict whether someone finishes a paragraph: sentence length, how much it varies, paragraph length, jargon load and whether the structure lets them skip.

None of those requires a reading-level formula. Averages hide the problem, because readable text is not uniform text; it alternates.

The audit reports the distribution rather than a single grade, which is the difference between a diagnostic and a vanity metric.

A paragraph of twenty-word sentences scores as readable and reads as a wall. A paragraph alternating six and thirty reads faster despite a worse average.

So the useful signal is low variance: every sentence landing in the same band, which is what generated prose does by default and what a heavy edit often reintroduces.

The fix is not to shorten everything. It is to cut one sentence hard and let another carry the detail, which produces the rhythm a reader can follow.

Long blocks fail on a screen regardless of sentence quality, and generated documents drift toward 150 to 250-word paragraphs because that is what an essay-shaped output looks like.

The check flags a block past 220 words. The editorial response is to split at the point where the subject changes, which is usually visible once you look for it.

On mobile, the practical limit is much shorter. If most of your traffic is a phone, run the audit on the rendered page rather than on the source.

Technical readers are slowed by explanation, not by terminology. Consumer readers are stopped by any term they have to look up.

So jargon density only means something next to a stated audience. The same draft can be too dense for a landing page and too shallow for an engineering handbook.

A practical test: identify the three terms a reader must know to use the page, and check they are either defined or genuinely universal in that audience.

A heading set, a summary line and a short list do more for comprehension than any amount of sentence surgery, because they let a reader decide what to skip.

Generated documents often use structure decoratively: headings that restate the title, lists where the items are not parallel, a summary that repeats the opening paragraph.

Check that every heading delivers a section, every list has items of the same grammatical shape, and the first paragraph tells a reader whether they are in the right place.

Reading-level formulas were designed for school texts and behave badly on names, code, numbers and product vocabulary. They reward short words over clear ones.

No measurement knows whether a sentence is true, whether the tone suits the audience or whether the piece is worth reading. Those stay with an editor.

Treat the audit as a map of where attention will be lost, then make the editorial decisions the map cannot make.

On paper, a comfortable measure is around sixty-five characters. On a screen it depends on the container, the type size and how much of the viewport the column occupies, which is why a draft that reads well in an editor can be exhausting on a phone.

Measure the rendered page rather than the source: how many words fit on a line at 390 pixels wide, and how far the reader has to move their eye to find the next one. A full-width column on a large monitor produces lines nobody can track.

Most readers do not read; they scan headings and follow links. That makes heading text and link text the two highest-leverage pieces of copy on the page, and both are usually written last and edited least.

A heading should tell a scanning reader what the section answers. A link should make sense out of context, because a reader navigating by links hears only the words inside the anchor. Download the pricing sheet beats click here, every time.

A list speeds reading when its items are parallel in shape and length, and slows it when they are three sentences of different structure. A table helps only when the reader is genuinely comparing values across the same fields.

Both break on narrow screens. A four-column table on a phone becomes a horizontal scroll that most readers abandon, so check the rendering at the width your traffic actually uses before deciding the table was the clearer choice.

Every figure asks the reader to hold something in mind, compare it with something else and work out whether it matters. That is three operations where a sentence of prose is one, which is why number-dense paragraphs read slowly even when every number is correct.

Reduce the operations: one unit per comparison, a stated baseline for any percentage, and rounding that matches the precision the reader needs rather than the precision the source provided.

A readability number on its own says nothing, because the acceptable range depends on format and audience. A policy page and a landing page are not competing on the same measure, and scoring them against each other produces a decision nobody should act on.

Measure a page against the last version of itself, and against pages in the same channel. The finding worth acting on is a page that has drifted, not one that sits above an arbitrary threshold somebody read about.