How do I write a page so AI can quote it?

Why forty to seventy words?

Because that is roughly the length of one extractable passage: long enough to carry a complete claim with its qualifier, short enough to be quoted whole rather than cut.

Retrieval systems do not work with pages. They work with chunks of text, and they quote the chunk that best matches the question. So the unit you are writing is the passage, and its job is to survive separation from everything around it.

• Under about forty words you usually cannot fit the subject, the claim and the condition, so the passage leans on the sentence before it and stops making sense alone.
• Over about seventy and something quoting you will cut it, at a point you did not choose. Better to make the cut yourself and put the second thought in the paragraph underneath.

How do I test whether a passage stands alone?

Cover the rest of the page and read only the first block. If a stranger could not tell who or what it is about, it fails.

Call it the lift-out test. It takes about a minute per page.

1. Copy the first block into an empty document. Nothing else. No heading, no title, no image caption.
2. Read it cold. Does it name the subject, or does it open with It, This, We or Our? Those four words are the most common reason a passage collapses when lifted.
3. Does it say where? An Australian business whose answer never says Australia, or the state, or the suburb, cannot be extracted for a location question, and almost every buyer question is a location question.
4. Does every number carry its denominator and its date? A percentage with neither is unverifiable, and unverifiable reads as invented.
5. Will it still be true in three months? If not, date it inside the passage so it ages honestly instead of silently.

This page is written to its own rule, which is the only demonstration worth offering. The block at the top names the subject, states the rule, gives the length, and stops. Nothing above it is load-bearing, and you can check that by scrolling back up and reading it on its own.

Which phrases make a passage unquotable?

Anything that points at context the reader cannot see.

1. As we mentioned above, and every other backward reference. Lifted out, it points at nothing.
2. It depends as an opening. It usually does depend. Say what it depends on in the same sentence and you have an answer instead of a shrug.
3. Bare pronouns for the subject. It is the fastest option for most businesses quotes as nonsense. Name the thing.
4. In this article we will look at. An announcement of an answer is not an answer, and it spends the one position that gets extracted.
5. Undated relative time. Recently, currently, this year. Write the date. A passage quoted eighteen months later still needs to be right or visibly stale.

There is a sixth that is easy to miss: opening with the question again. If your heading is How much does it cost?, the first sentence underneath should be a price, not So how much does it cost? A heading and its first sentence are frequently retrieved together, and two copies of the question is a wasted pair.

Where do the numbers go, and why does every number need its denominator?

Inside the passage, with the denominator, the date and the source, because a number lifted without them cannot be checked and reads as made up.

The difference is easiest to see side by side. Weak:

Strong:

The second is longer, and it is the one that survives being quoted. It names the source, the sample, the date and the exact test. The first has nothing in it for anyone to verify, which makes it worthless to a system that is trying to decide whether to repeat you.

The same discipline protects you later. If a figure gets restated wrongly somewhere else, a passage that carried its own denominator is far easier to correct than one that did not. And if you have the underlying data, publish it. Ours is at the study as CSV and JSON under CC BY 4.0, which means anyone can re-run our arithmetic and disagree with us in public.

How do I structure the depth underneath so a human still reads it?

Headings phrased as the next question a reader would ask, and a first sentence under each that answers it.

• Every section is its own small answer. Someone arriving at an anchor halfway down should not have to scroll up to work out what is being discussed.
• One question per page. A page trying to answer three questions will be extracted well for none of them. Split it and link the pieces.
• Let the URL read like the question with the filler words taken out. It is a small signal and it costs nothing.
• Evidence after the claim, not before it. Setting the scene first is a magazine habit, and it buries the sentence you wanted quoted.
• Name yourself properly in the text. Our own matcher, published on The Receipts Engine, is a regular expression over the stored answer text, and a brand name of six characters or fewer with no space only matches when it is capitalised, because short brand names are usually ordinary English words too. Leaning on we costs you the mention in any system that counts the way ours does.

See how a real answer gets counted.

Writing a quotable page is half of it. A free scan shows you whether the engines are quoting anyone in your category yet, and which pages they are pulling from.