What the Model Actually Sees

Your page is 2,400 words of carefully arranged HTML. An answer engine reduces it to a few hundred words of plain text and reads until its budget runs out. This chapter is about seeing exactly what survives that reduction, because almost nobody has.

The problem

Everything you know about writing for the web assumes a reader with eyes. Layout carries meaning: this is a hero, that is a sidebar, the important number is large and orange. A model receives none of that. It receives text, stripped of your navigation and your footer and your cookie banner, in document order, and it stops reading at a budget.

So the question that decides whether you are quotable is not “is my page good?” It is: what does my page become when it is flattened, and is the answer still in there when the reading stops? Until you can see that, you are writing blind.

The hard way

People approximate this in two ways, both bad. They read their own page and imagine what a model would take from it — which is impossible, because they already know the answer and cannot unsee it. Or they paste the URL into a chatbot and ask “what is this page about?” — which tests the chatbot's fetching, not your page, and gives a fluent answer either way.

Neither tells you the thing you need: where the reading stops. And that is usually where the problem is.

The screen itself

In the report, open What the model sees. It shows three things.

One: the extracted text. Your page with the boilerplate discarded — navigation, footer, cookie notice, the repeated calls to action — leaving what an extractor would treat as the content. Read it as a stranger. For Freddy, the extract opens with “Try free. No card. 14 days.” because that was the first real text after the nav, and the sentence explaining what Freddy is sat in an image caption forty lines further down.

Two: the reading budget line. A marker at roughly two thousand tokens, drawn through the text. This is an approximation of where a retrieval system's attention typically ends — not a rule any engine publishes, and Sjá says as much. But the shape of the lesson survives the approximation: text far below the line is competing for attention that has mostly been spent, and if your answer is down there, you are relying on a budget you do not control.

Three: the quotable passages. Sjá ranks the passages most likely to be lifted into an answer, with a score and the reason for each. Reasons are concrete: this passage is self-contained; this one begins with a pronoun whose antecedent is two paragraphs up; this one states a fact with a number and a unit.

Why self-contained wins

That last point deserves its own section, because it is the single most useful writing lesson in this course.

An answer engine quotes a passage, not a page. The passage is torn out of your document and dropped into a reply next to sentences from three other sites. A passage that only makes sense in place — because it starts with “It also supports…”, or “This means that…”, or “Unlike the above…” — cannot be used. Not because a filter rejects it, but because it would be incoherent, and the engine's own quality pressure selects against incoherence.

So Freddy's paragraph:

It stores them as plain files, which means you can
open them in anything.

becomes:

Freddy stores every note as a plain Markdown file in
a folder you choose, so any editor can open them and
nothing is locked in a database.

Same fact. The second one can be quoted standing up. Note what changed: the subject is named rather than referred to, the mechanism is stated rather than implied, and the consequence is in the same sentence as the cause. That is the whole technique, and it is worth applying to the first paragraph of every page you own.

The exercise

Do this once, on your most important page, and the habit will stick:

1. Note the top passage score in the report.
2. Rewrite the page's opening paragraph so it names the subject, states what the thing is, and stands alone. Move it above the fold in the HTML — document order, not visual order.
3. Re-run and compare. The passage list will have reordered.

You have now done something most site owners never do: optimised for the reader that has no eyes, and verified that it worked.

What you learned

• A model reads flattened text, in document order, with a budget. Layout carries none of your meaning.
• Look at the extract as a stranger. What comes first is often a call to action, not a definition.
• Engines quote passages, not pages. A passage that needs its neighbours cannot be used.
• Name the subject, state the mechanism, keep cause and consequence in one sentence. That is what makes a paragraph quotable.
• Put the answer early in the HTML, which is not the same as early on the screen.