What content should I create to rank in AI-generated answers?

Content strategy for AI answers is narrower than content strategy generally, because most of what a marketing team publishes cannot be cited under any circumstances. Knowing which types can changes what you commission far more than any optimisation tactic applied afterwards.

The organising principle

A model assembling an answer needs a passage it can reproduce and attribute. That single requirement filters the entire content universe.

Content gets cited when it contains a specific claim that can be stated without inventing anything, credited to a source with a reason to be trusted. Content that contains no specific claim — because it persuades, describes, or hedges — has nothing to extract, regardless of how well written it is.

The question to ask of any brief is not “is this good?” but “what sentence in this could a model quote, and why would it credit us?”

1. Original data

The highest-value type by a wide margin, and the most under-produced.

If a claim exists only on your page, any answer using it must cite you. There is nothing else to attribute it to. This is a structurally different position from competing to be the best-written version of a widely-known fact.

Most companies are sitting on publishable data they have never thought of as content. Aggregate patterns from your own product, anonymised and aggregated. A survey of fifty customers, which takes a fortnight. A benchmark you already ran internally for a purchasing decision. Support ticket categories, which describe what your market actually struggles with better than any industry report.

The research backs the instinct. The Princeton and IIT Delhi GEO study tested optimisation methods across roughly 10,000 queries and found Statistics Addition, Quotation Addition and Cite Sources were the three strongest, delivering relative visibility improvements of roughly 30–40 percent. Original data is the most defensible form of that.

It also compounds in a way nothing else does: other writers cite it, which builds the third-party corroboration that drives assistant mentions, from a single piece of work.

2. Single-question pages

Unglamorous, cheap, and the backbone of a citation surface.

One buyer question, answered completely, with the answer in the first two sentences and nuance below. These map exactly onto how a prompt is phrased, which is why they are extracted so reliably.

This inverts the content-hub instinct. A comprehensive 4,000-word guide frequently earns fewer citations than five focused pages, because a model needs a clean passage on a narrow point and the comprehensive guide buries each point among five others. Thirty single-question pages outperform three hundred general ones.

Where to get the questions: your support inbox, internal site search, People Also Ask, forum threads, and — most directly — the prompts where your scans show competitors being cited and you are not.

3. Honest comparisons

“X vs Y” questions are enormously common in assistant conversations, and models heavily prefer sources that read as balanced.

The counterintuitive rule: a comparison naming situations where your product is the wrong choice is far more likely to be cited than one that does not. It reads as analysis rather than marketing, which is exactly the distinction models are sensitive to. It also converts better, because readers apply the same test.

Structure them with explicit criteria, a table, and a clear verdict per use case rather than a general winner. “Choose X if you need A; choose Y if B matters more” is quotable in a way “X is the best option” is not.

4. Definitions and explainers

Precise definitions are extracted constantly, because “what is X” is one of the most common prompt shapes.

The format that works: a heading asking what X is, followed immediately by one sentence of the form “X is [category] that [does what] for [whom]”, then the nuance. No history, no preamble, no “before we can understand X, we must first consider…”.

These have a low ceiling individually — a definition is rarely commercially valuable on its own — and they build topical authority efficiently and cheaply. They are best treated as connective tissue rather than as the main investment.

5. Procedures

Step-by-step instructions are lifted close to verbatim for how-to prompts.

Numbered steps, each beginning with a verb, each self-contained enough to make sense out of order. Include the prerequisites, the common failure at each step, and how to verify it worked — those details are what separate a procedure that gets cited from one that gets paraphrased.

HowTo structured data helps here, provided it mirrors the visible steps exactly. Where it does not, it is a liability rather than a signal.

6. Criteria and decision frameworks

“How do I choose an X” is a common prompt and an under-served content shape.

The format: “Three things decide whether X works for you”, followed by a paragraph on each with an explicit test the reader can apply. This is highly extractable because the criteria list is itself a quotable unit, and it is genuinely useful, which is why it earns links as well.

It also positions well. A company that publishes the honest criteria for choosing in its category — including criteria on which it does not win — is describing the market rather than selling into it, and both models and buyers respond to that.

Three types that are never cited

Homepages and landing pages. There is no extractable factual claim in a page designed to persuade. When a model does reference company marketing copy, it typically frames it as a claim rather than a fact. This is expected rather than a failure — do not try to fix it by rewriting your homepage as a FAQ.

Gated content. A white paper behind a form is invisible. If it contains data worth citing, publish the findings openly and gate something else.

Client-side-rendered anything. AI crawlers do not execute JavaScript — Vercel and MERJ recorded zero JavaScript execution across more than 500 million GPTBot fetches, with the same behaviour from ClaudeBot and PerplexityBot. The best content in the world does not survive being unreadable.

Briefing content that can be cited

Two fields added to an existing content brief do most of the work, and they take fifteen minutes rather than a new process.

“The single question this page answers.” One sentence, no “and”. If the brief cannot state it, the page will cover several things partially and be cited for none of them. This field alone prevents the most common structural failure.

“The one thing this page will contain that nobody else has.” A number, a named source, a real example, a stated position. If the honest answer is nothing, the page is a restatement — and restatements are neither ranked nor cited, because there is always an older, better-linked version of the same claim.

A brief that cannot fill the second field is worth questioning before it is worth writing. That is an uncomfortable editorial discipline and it is the single highest-leverage change available to a content team, because it eliminates the pages that were always going to underperform before anyone spends a day on them.

Where writers get this wrong

Three habits work against citation and all three are taught as good writing.

Setting the scene. Establishing context before the answer is good essay technique and fatal here. The answer goes first; the context follows.

Varying the phrasing. Using synonyms and pronouns to avoid repetition makes passages unextractable, because lifted out of context they refer to nothing. Repeat the subject noun.

Hedging responsibly. “It may depend on circumstances” is professionally cautious and contains no claim. State the answer, then qualify it in the next sentence — the claim survives extraction and the qualification is there for readers.

Deciding what to commission

Work from evidence rather than from a content calendar.

1. Run twenty buyer questions through ChatGPT, Gemini and Perplexity. Record who is named and which URLs are cited.
2. List the prompts where competitors are cited and you are not. Each one is a brief.
3. For each, check whether a page on your site should already have been the answer. If yes, it is a restructure rather than a commission — cheaper and faster.
4. For the genuine gaps, pick the type from the six above based on the prompt shape. A “how do I” prompt needs a procedure; a “which should I” prompt needs criteria or a comparison.
5. Commission one original data piece per year regardless of what the gaps say. It is the only investment that produces citations you cannot lose to a better-written competitor page.

That process produces a shorter list than a content calendar and a considerably higher hit rate, because every item on it is a question you have watched an assistant answer without you.

Sources

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735); Statistics Addition, Quotation Addition and Cite Sources as top-performing methods, ~30–40% relative visibility improvement across ~10,000 queries.
• Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; same behaviour for ClaudeBot and PerplexityBot.
• Google Search Central — structured data guidance for HowTo and FAQPage.

Frequently asked questions