Why AI Answers Are Hard to Read — and What to Fix Before You Paste One Into Your Notes

A long model answer is tiring to read because nothing in it ranks anything. Every sentence arrives at the same weight; every bullet is the same size. That uniformity — not the bullets, not the length — is what turns reading into scanning. The repair happens after the text arrives, in a file you control.

The evidence is narrower than the folklore around it, so here is the record up front. In November 2017, Kara Pernice of Nielsen Norman Group published an article whose title contains the word Misunderstood.¹ It revisits the F-shaped scanning pattern the firm identified in 2006 — the one where a heatmap of readers' eyes glows in the shape of a capital F.

Pernice's contribution is not the pattern. It is the conditions. She names three, and the first is a page with "little or no formatting for the web."²

She studied web pages. She did not study chat output. The argument that a pasted model answer satisfies her three conditions is this post's, not hers — and it is the only claim here that a careful reader should push on. Everything downstream of it is her documented craft advice, applied to a document she never saw.

The record, and what it actually says

Nielsen Norman Group identified the F-shaped pattern in 2006, and its 2017 revisit exists mostly to correct people who quote it. The article's own introduction says many have referred to the research "sometimes correctly, and in many other instances misinterpreting it."³ The F is not a law of reading. It is a documented default.

It is thinner than twenty years of citation suggests. The famous early heatmap — the one reproduced in a thousand slide decks — aggregates "gazes from more than 45 people trying to learn what happened in Galveston Texas in 1900," and Pernice dates it to "our studies in the early 2000s."⁴ That is the base of a finding that hardened into a law.

The 2017 article re-tests it. Pernice reports the pattern is "alive and well in today's world — both on desktop and on mobile"; the heatmap she shows for that new study aggregates data from 47 people.⁵ Alive, then. Still not a law.

The correction that matters most is quantitative. The F is one of six patterns the article documents, in a section titled "The F-Shaped Pattern Is Not the Only Scanning Pattern."⁶ Among the others: the layer-cake pattern, which "occurs when the eyes scan headings and subheadings and skip the normal text below"; the spotted, marking and bypassing patterns; and the commitment pattern, "fixating on almost everything on the page."⁶

Eyes do many things. The F is what they do under specific conditions.

The idea that people skim rather than read is older than the F. Jakob Nielsen's 1997 article reports that "79 percent of our test users always scanned any new page they came across; only 16 percent read word-by-word."⁷ A 2008 follow-up put the ceiling at "at most 28% of the words during an average visit; 20% is more likely."⁸

Treat both as history: measurements of websites, from NN/g's own test users, complicated by the 2017 commitment pattern. They date the lineage of an idea. They measure nothing you will read today.

The three conditions that produce F-scanning

Pernice lists three conditions, and the F appears when all three are present: text with "little or no formatting for the web," meaning a wall with no bolding, bullets, or subheadings; a reader trying to be efficient; and a reader "not so committed or interested that he is willing to read every word."² Remove one and the pattern loosens.

The conditions are the useful part, because they are causal rather than descriptive. The article's thesis sentence is one line: "The F-pattern is the default pattern when there are no strong cues to attract the eyes towards meaningful information."⁹

A few sentences earlier in that same paragraph, separated by an intervening line about content ordering, Pernice writes: "It's not that people will always scan the page in the shape of an F."⁹

Two spans, one point. The F is a symptom of an absence.

Pernice is blunt about the damage. "Make no mistake, the F-shaped scanning pattern is bad for users and businesses: it means that users may skip important content simply because it appears on the right side of the page."¹⁰

And she is equally blunt about the cure, in two consecutive sentences: "Good web formatting reduces the impact of F-scanning. If your pages have big chunks of unformatted text, people will scan it in an F-shape."¹¹

The condition is the wall. The wall is fixable.

Why a pasted answer meets all three

This is the post's argument, not Pernice's finding: a long model answer meets all three conditions by default. It arrives as one undifferentiated block. You are skimming it to get on with the task. You did not write it, so you have no commitment to any sentence in it.

The third condition is the one people miss. Commitment is not a character trait — it is a relationship to a specific text. You are committed to a paragraph you fought to write. You are not committed to one that appeared in under a second.

The uniformity is the mechanism, and Pernice's taxonomy names the failure. Her bypassing pattern "occurs when people deliberately skip the first words of the line when multiple lines of text in a list start all with the same word(s)."⁶

Read a long generated answer and count how many bullets open with the same construction. That mapping is this post's observation, not her measurement. But the name for what eyes do to identically-shaped lines existed before machines produced them at volume.

Where a note sits inside a model's context window also changes the answer it gets back, but that is a machine's problem and a different post's. This one is about the text after it arrives. Keeping the original and treating the summary as a lossy view is the epistemic half.

This is the visual half.

The bulleted wall is a default, and the vendors are patching it

Markdown is the documented default. OpenAI's Model Spec says assistant output "should be formatted in Markdown with LaTeX extensions" unless told otherwise.¹² But the same spec asks for formatting used "judiciously to aid the user in scanning the response" — and rejects the list-of-facts answer outright.¹³ The wall is a default, not a design goal.

That word is worth stopping on. The spec's stated purpose for formatting is scanning — the exact behaviour Pernice's eye-tracking is about. Vendor and researcher describe one mechanism from opposite ends.

Two sentences later in that section, separated by a line about purple prose, the spec adds: "If the user asks a question, the response should be phrased as a direct answer rather than a list of facts."¹⁴ So the bulleted wall is not what the specification asks for. It is a gap between spec and shipped behaviour — and the Model Spec concedes the gap itself: "Our production models do not yet fully reflect the Model Spec."¹⁵

The other lab spends system-prompt tokens on the same defect. Anthropic's published system prompt contains a rule that reads like a design review:

"Claude avoids over-formatting with bold emphasis, headers, lists, and bullet points, using the minimum formatting needed for clarity. Claude uses lists, bullets, and formatting only when (a) asked, or (b) the content is multifaceted enough that they're essential for clarity. Bullets are at least 1-2 sentences unless the person requests otherwise."¹⁶

That passage is dated June 9, 2026. The version before it, dated May 28, 2026, carries the same opening sentence and its own paragraph against the same habit, telling the model to respond "in prose rather than lists or bullets unless asked."¹⁶ It is a standing instruction, not a one-off.

That is the honest shape of this problem. Nobody at either lab thinks the bulleted wall is good. It is a default two vendors fight at the most expensive layer they have: mutable, dated, and not a property of machines.

So everything here about how AI answers look describes July 2026. The argument outlives it, because the argument is not about bullets. It is about who can know what matters.

How to re-rank an answer before it becomes a note

Pernice's antidote section instructs designers: "Do the work for the users instead of forcing them to exert effort and take bad shortcuts."¹⁷ For a pasted answer, you are both parties — the one who formats and the one who reads. That inversion is the whole method. Her tips are the checklist; the table below applies each one.

Bullets are not the villain. Uniformity is. Nielsen Norman Group recommends bullets in as many words, and this post is not a case against them. A list of three genuinely parallel items is a ranking device. A list of eleven items of equal weight, each opening with the same verb, ranks nothing — it re-renders the wall with dots in front of it.

Every move above is structural: it changes the text, not the way the text is rendered. How wide you set the line is a separate lever.

Notice what the headings do. They do not stop the scanning; they make the scanning work. Pernice's taxonomy names the eyes that "scan headings and subheadings and skip the normal text below": the layer-cake pattern.⁶ She describes it neutrally, as one pattern among six.

Reading it as the goal is this post's interpretation, not her recommendation. But a reader who skips your body text and still finds the section they need has been served, not failed. Headings are also what make a long note navigable by its outline.

None of this is a prompt — and the reason is not that prompting is unsophisticated. It is that the model does not know which of its sentences matters to you. It has no access to your project, or to the phrase you will search for in November.

Ask for fewer bullets and you get fewer bullets. You do not get your hierarchy, because your hierarchy is not in the input.

Which is why the cutting and promoting is not overhead on the note. Deciding what matters is the work — a claim this post takes from active note-taking rather than re-argues here.

Paste-and-done produces an artifact you own and do not know.

The caveats: what this argument does not prove

The evidence base here is one practitioner eye-tracking record and two vendor specifications. It is enough for a craft argument and not enough for a scientific one. Nielsen Norman Group's research is not peer-reviewed literature, and this post does not dress it as such. Five limits are worth naming plainly.

• Pernice studied web pages, not chat output. Every mapping from her conditions to a pasted answer is this post's reasoning. No eye-tracking study cited here measured anyone reading a model's response.
• The F-pattern is not a law. It is one of six documented patterns, under three stated conditions, in an article whose own title calls it misunderstood. Any sentence of the form "people read in an F" is the folklore this post corrects.
• Bullets are recommended, not condemned. The defect is undifferentiated weight, not the marker.
• The old scanning numbers are lineage, not measurement. 79/16 is from 1997; 28%/20% from 2008. Both describe websites of their era.
• Model output form is a dated, mutable default. It is patched at the system-prompt layer, and system prompts ship monthly.

What survives all five is narrow and mechanical: text where nothing ranks anything defeats scanning, and only a reader knows what to rank. That is not a claim that generated prose is worthless. Both vendors publish specifications asking for better. That is a shared engineering observation, not a complaint.

The villain is undifferentiated text. It has no vendor.

Frequently Asked Questions

Six questions this post gets asked, answered from the record above rather than from folklore. Where an answer rests on interpretation rather than a cited finding, it says so. The short version: the defect is real, the mechanism is uniformity — and the fix belongs to whoever holds the file.

Why is AI output hard to read? Because nothing in it ranks anything. Every sentence arrives at the same weight and every bullet at the same size, which gives your eyes no cues to aim at. Nielsen Norman Group's Pernice calls a page with "little or no formatting for the web" the first condition for F-scanning.²

Do people actually read in an F-shape? Sometimes, under conditions. Nielsen Norman Group documents six scanning patterns, not one, and its 2017 revisit exists partly to correct people quoting the F as a universal law.⁶ Pernice's framing: it is "the default pattern when there are no strong cues to attract the eyes towards meaningful information."⁹

Are bullet points bad in AI answers? No. Nielsen Norman Group recommends them verbatim: "Use bullets and numbers to call out items in a list or process."¹⁸ The problem is a list where every item carries identical visual weight, which ranks nothing. Keep the bullets that are genuinely a list; turn the rest back into prose.

Can I just prompt the model to format it better? Partly, and it will not finish the job. The model does not know which of its sentences matters to you, because your project and the phrase you will search for later are not in its input. OpenAI's Model Spec already asks for a "direct answer rather than a list of facts."¹⁴ The gap persists.

How do I format an AI answer into a note? Promote the real answer above the preamble, add headings, rename each heading to the two words carrying its gist, bold the phrases you will search for later, keep only the bullets that are truly lists, and cut the rest. Each move maps to a documented Pernice antidote.¹⁸

Is the F-pattern still relevant in 2026? Nielsen Norman Group's 2017 re-test says the pattern is "alive and well in today's world — both on desktop and on mobile," illustrated by a heatmap aggregating data from 47 people.⁵ That is the most recent measurement cited here, and it is nine years old. Treat it as a durable observation, not a current benchmark.

F-scanning is not how people read. It is what reading decays into when nothing on the page ranks anything. No model can do that ranking for you, because no model knows what you came for.

If you want the answer you paste to become a file you actually control — open Markdown, on your own device, under headings you chose — that is what mnmnote.com is for.