Write the paragraph your spec table never had

A shopper asks a model whether a pair of trail shoes runs true to size. The model goes looking for a sentence that answers it. Your page says Weight · 291g, Drop · 4mm, Upper · engineered mesh, and Add to cart. None of that is an answer, so it cites somebody who wrote one.

This is the most fixable problem in AI search, and it has almost nothing to do with technical optimization. It is a writing problem. Here is how we write these.

1. Start from the question, not the product

Do not open a document and describe the shoe. Open the question and answer it in the first sentence. They run about a half size small. Then explain. The models are extracting a claim; give them a claim to extract, not a paragraph that eventually implies one.

If you cannot state the answer in one sentence, you do not know it yet, and you should not be publishing on it.

2. Use the shopper’s words, not your spec sheet’s

Your catalog says forefoot volume. The shopper says wide feet. Your catalog says lug depth 5mm. The shopper says will these grip on wet rock. Both should be on the page — the plain phrasing is what gets matched, the specification is what makes the answer credible.

Write the sentence in the shopper’s language and let the number sit inside it: the 5mm lugs hold on wet rock and feel slappy on pavement. One sentence, both vocabularies, an actual opinion.

3. Say what it is not for

This is the rule people resist and it is the one that works hardest. A page that says a shoe is wrong for road running is more useful, more quotable, and more likely to be trusted than a page that says it is excellent at everything.

It also protects you. An answer that sends the wrong buyer to your product page produces a return, not a sale. Being named for the thing you are genuinely good at is worth more than being named often.

4. One question per block

Do not write a 900-word essay covering fit, grip, durability and care. Write four blocks, each opening with the question as a heading and answering it underneath. The models pull passages, not pages. A block that maps to exactly one question is a block that can be lifted cleanly.

This is also why FAQ markup keeps outperforming prose for this job. Not because of the schema — because the format forces you to write one answer at a time.

5. Never fill a gap you cannot source

If your catalog does not record whether the upper is waterproof, do not write that it is. Do not write that it is not. Leave the question open and go find out.

An answer that is confident and wrong is worse than no answer at all, because it will be quoted with your name attached to it. We build this into our own drafting: where the catalog is silent, the draft leaves a question for a human rather than inventing a specification. It is the one rule we do not let anything override.

6. Date it and revisit it

Fit changes between model years. Compatibility lists grow. A page that answered correctly in March and is wrong by September is actively costing you. Put a visible last-reviewed date on answer content and actually review it.

What good looks like

Here is the difference in practice, on one question.

Same product. Same facts. One of them can be quoted.

If you want to know which questions are worth writing for your own catalog before you write anything, run the free check — it shows the questions in your category, who is being named, and which pages are earning the citations.