Writing content that AI assistants quote, without the hype

A new set of terms has appeared for making content visible in AI answers: generative engine optimisation, answer engine optimisation, LLM optimisation. Behind the labels, the underlying question is reasonable. When an AI assistant or an AI search feature answers a question by reading web pages, which passages does it use, and how do you write so that yours are among them?

Nobody outside the vendors knows the selection mechanisms in detail, and they differ between systems. But the way these systems work in general (retrieve pages, pick passages, compose an answer, cite sources) makes some writing habits clearly more useful than others. None of them require tricks, and all of them also make pages better for human readers.

How answer engines use a page

In broad terms, an AI search feature does three things with your page:

1. Retrieves it as relevant to the question, using a search index. This is ordinary search: the page must be crawlable, indexable and relevant.
2. Selects passages that address the question or a part of it. Long pages are often split into chunks, and the chunk is what gets considered, not the page as a whole.
3. Composes and attributes: the answer paraphrases or quotes the passages, and links to the sources that supported specific claims.

Step 1 is classic SEO. Steps 2 and 3 are where writing matters: a passage has to make sense when read on its own, and it has to contain something worth attributing.

Habit 1: write self-contained statements

A passage lifted out of context should still be true and understandable. Compare:

Depends on context: "It's usually around 2.5 seconds, but as mentioned above, that depends."

Self-contained: "Google considers a Largest Contentful Paint of 2.5 seconds or less good, measured at the 75th percentile of page loads."

Name the subject instead of using "it" and "this" at the start of key sentences. Include the condition in the sentence ("for homes built before 1980…") rather than in a paragraph three screens up. This feels slightly repetitive to write, and it reads perfectly well.

Habit 2: answer first, explain second

Under each heading, state the answer in the first one or two sentences, then add the reasoning, evidence and exceptions. Readers who want only the answer get it; readers who want depth keep going; systems looking for a passage that addresses the question find it at the top of the chunk. The featured snippets guide describes the same pattern for search snippets.

Habit 3: make headings match questions

Headings act as labels for the chunks beneath them. "Pricing" is vague; "How much does a boiler service cost in 2026?" or "What is included in the annual service" tells both reader and system what the section answers. Use the questions customers actually ask, from support emails, sales calls, People Also Ask and Search Console queries.

Habit 4: be specific

Specific statements are attributable; generic ones are interchangeable. If fifty pages say "prices vary depending on many factors", any of them can stand in for the others. If your page says "a standard annual service for a gas combi boiler typically takes 45 to 60 minutes; we charge a fixed fee and parts are extra", it offers something a summary can cite. Useful specifics include:

• Numbers with units and the date they apply to.
• Named conditions and thresholds.
• Steps in order, with the tools or inputs each needs.
• Real examples, with enough detail to be credible.
• Your own data or measurements, with the method stated.

Habit 5: use real structure

Use HTML lists for steps and options, and HTML tables for comparisons and specifications. A table with clear column headings and units is one of the easiest structures for any system to read accurately. Avoid putting key information only in images, PDFs, carousels or tabs that load on click.

Content	Best structure
A sequence of actions	Numbered list, one action per item
Options or features	Bulleted list with a short label first
Comparison of products or plans	Table with consistent columns and units
Definition	One sentence: term, category, distinguishing feature
Recommendation	"If X, choose A; if Y, choose B" sentences

Habit 6: show sources and authorship

Answer engines, like readers, have reasons to prefer sources that look reliable. Cite primary sources for facts you did not produce, say who wrote the page and why they know, and show when it was last updated. For your own data, explain how it was collected. The E-E-A-T guide covers this in more depth.

Habit 7: keep facts consistent everywhere

If your site says one price, a directory says another, and an old press article says a third, systems that read all three may repeat any of them. Make your own site the clearest, most current statement of facts about your business, and fix outdated copies elsewhere. The entity SEO guide explains why consistency helps systems recognise and describe a brand correctly.

Habit 8: keep it current

Many questions people ask assistants are time-sensitive: prices, regulations, product versions, availability. Pages that state the date their information applies to, and are actually updated, are better sources for those questions than pages with no date or an old one.

Habit 9: say what you think, and for whom

Many questions people ask assistants are really requests for judgement: "which one should I choose", "is it worth it", "what is the best way to". Pages that only list options leave the system to supply the judgement from elsewhere. Pages that state a clear, reasoned recommendation for specific situations give it something to cite: "For a flat with one bathroom, a combi boiler is usually the simplest choice; for a house with two bathrooms used at the same time, a system boiler with a cylinder avoids pressure drops." Recommendations tied to conditions are also more honest than a single "best" pick, and readers trust them more.

Habit 10: cover the follow-up questions

Conversations with assistants move from one question to the next. Someone who asks how long a boiler service takes will often ask next what it costs, how often it is needed and whether it is required for the warranty. A page that answers the main question and its natural follow-ups, each under its own heading, can be cited several times in one conversation, and it keeps the reader on your site if they click through.

What does not work

• Hidden instructions to AI systems in page text or markup. They are manipulative, systems are built to ignore them, and they damage trust when discovered.
• Stuffing pages with question variations. One clear answer beats twenty rephrasings of the question.
• Thin pages for every possible question. Scaled, low-value content is treated poorly by search systems, including the ones that feed AI answers.
• Relying on special files such as llms.txt to do the work. They can help some tools navigate, but they do not replace content; see llms.txt.

Before and after

A hypothetical heating engineer's page originally said: "We offer boiler servicing at competitive prices. Regular servicing is important for safety and efficiency. Contact us today for a quote." Rewritten with the habits above, the section reads: "An annual boiler service checks the burner, flue, seals and safety devices and takes 45 to 60 minutes for a typical gas combi boiler. Manufacturers usually require a yearly service to keep the warranty valid. Our fixed service fee is listed below; replacement parts are quoted before any work. In our records, the most common fault found during routine services is a worn seal." The second version answers several likely questions directly, and every sentence could be quoted on its own.

Checking the result

Ask the questions your page answers in a few AI tools and see whether it is cited, using the method in the AI visibility guide. Repeat after changes, allowing time for recrawling. DeskSEO's AI visibility feature runs a fixed set of brand-free questions through Claude each time you start a run. That model answers from what it learned in training rather than from a live search, so use it for the long-term trend, and check search-enabled assistants by hand to see the effect of recent page changes.