JSON-LD, llms.txt and Q&A pairs: the three formats AI cites
2026.09.24
AI search does not read for keyword density. It reads for sentences it can move straight into an answer. The same point written as prose and written as a question with a direct answer are not equally likely to get cited.
Q&A pairs: the smallest citable unit
If an answer fits in two sentences, an AI has a good chance of quoting the whole thing. Rewrite the main points of a page as question plus direct answer, and keep a number, a range or a condition inside the answer.
- Phrase the question the way people actually ask it, not in internal terms
- Lead with the conclusion, then the detail
- One question per answer. Do not merge them
JSON-LD: entities and answers a machine can read
People read page copy fine; a machine does not always parse what it means. JSON-LD uses schema.org vocabulary to state plainly what the product is and who answers which question. FAQPage, Product and SoftwareApplication cover most of it.
Declare the same Q&A pair once as a FAQPage and the extractor has a ready-made structure instead of a guess to make.
llms.txt: a site index written for AI
llms.txt is a plain text file at the site root that lists the brand, the core pages and the key facts. A crawler gets a human reviewed table of contents instead of inferring one from your navigation.
It pays off most on multilingual sites, SaaS product pages and documentation. Those are exactly the sites where one answer is spread across several levels of the tree.
Bottom line
Q&A pairs get you cited. JSON-LD gets you understood. llms.txt does a different job: it gets AI to find the site in the first place. The three do not compete, so fill the gap you actually have. If that is content, start with the pairs; if it is structure, start with JSON-LD; if AI never reaches you, start with the index.