How to write content AI assistants quote: answer first, then the evidence

To write content for AI search, put the answer in your first sentence, then back it with evidence an assistant can lift whole: a dated statistic, a quotation from a named person, a link to the source. In the research that coined "generative engine optimisation", adding quotations and statistics raised a page's visibility in AI answers more than any other change tested, and citing sources also helped.

Generative engines, in the GEO paper's description, summarise information from several sources into one answer. Your page competes to supply a sentence, a number or a row of a table to that answer. Content that is easy to quote has the answer up front and the proof attached. This guide shows what the research found, what Google says, and how to apply it, with before-and-after examples.

In short

• Lead with a one- or two-sentence answer that makes sense on its own.
• In the GEO study (KDD 2024), adding quotations raised a page's visibility score from 19.5 to 27.8, statistics to 25.9 and source citations to 24.9. Keyword stuffing lowered it to 17.8.
• Put every number with its date and source in the same sentence, and quote real, named people.
• Use tables for comparisons and finish with a short FAQ whose answers stand alone.
• Google says you don't need to write "in a specific way" for its AI features. Write for readers; these habits help both.

What does the research say about content AI assistants quote?

The best public evidence is "GEO: Generative Engine Optimization" by Pranjal Aggarwal and colleagues, published at KDD 2024. The authors built a test engine that fetched the top five Google results for each question and had GPT-3.5 Turbo write an answer from them. They then had a language model rewrite one source page in different ways and measured how much of the answer drew on it, weighted by position. Their benchmark covered 10,000 queries across 25 domains.

Change made to the page	Visibility score (unchanged page: 19.5)
Adding quotations from relevant sources	27.8
Adding statistics	25.9
Making the text more fluent	25.1
Citing sources	24.9
Adding technical terms	23.1
Making it easier to understand	22.2
Making the tone more authoritative	21.8
Keyword stuffing	17.8

Source: Aggarwal et al., Table 1, position-adjusted word count on GEO-bench.

Three findings matter for a marketing team:

1. Evidence beats a confident tone. Making the text sound more authoritative scored 21.8; adding a quotation scored 27.8. Plain fluency helped too (25.1), so clear writing still matters.
2. It depends on the subject. Statistics helped most for law and government topics and opinion questions; quotations for people and society, explanations and history; citations for factual questions.
3. Lower-ranked pages gained most. Citing sources raised visibility by 115.1% for pages ranked fifth in search, while the top-ranked page fell by 30.3% on average.

The abstract reports visibility gains of "up to 40%", and up to 37% on Perplexity.ai. Two caveats: the test engine was a 2023 research setup, not today's ChatGPT or Gemini, and the study measured visibility, not truth. On your site, only real, sourced facts belong.

Does Google want you to write differently for AI?

No, and that is consistent with the research. Google's guide to its generative AI features says "you don't need to write in a specific way just for generative AI search" and that ordinary SEO best practice still applies. Its emphasis is on "non-commodity content": first-hand experience and expert views, not a restatement of what is already online. It also recommends pages organised into sections with clear headings.

Google warns against one tactic directly: creating a separate page for every variation of a question in order to manipulate AI responses breaches its scaled content abuse spam policy. Write one strong page per topic.

The GEO habits are simply good writing for busy readers: the answer first, the proof attached, the source named.

How do you write content for AI search that gets quoted?

All examples below use fictional brands, and their figures and quotes are invented for illustration.

Answer the question in the first sentence

Before (Kestrel Insurance, fictional): "At Kestrel, we understand that choosing car insurance can be overwhelming. There are many factors to consider, and every driver is different."

After: "Kestrel's comprehensive car insurance pays for damage to your own car and to other people's property. Its third-party cover pays only for the damage you cause to others."

The first version gives an assistant nothing to quote. The second answers the question a buyer typed and stands alone.

Add statistics with their date and source

Before (Tandem Motors, fictional): "Our service plans save customers a lot of money."

After: "Tandem Motors customers on a five-year service plan paid an average of R14,200 less in workshop costs than customers without one, according to Tandem's own service records for 2021 to 2025."

A number without a date and source is a claim. With them, it is evidence an assistant can attribute.

Quote a named person

Before (Meridian Bank, fictional): "Experts agree that small businesses should separate personal and business banking."

After: "'Most of the cash-flow problems we see start with one account doing two jobs,' says Thandi Mokoena, head of small business banking at Meridian Bank."

Quote your own specialists, with their permission, or credible outside experts you have actually interviewed. Never invent a quote.

Cite your sources

Link to the primary source for every figure you did not produce yourself: the regulator, the official statistics, the original study. In the GEO study, citations helped most on factual questions.

Use tables for comparisons

When a buyer asks "which is better, X or Y?", a table puts each difference in its own row, ready to lift. Comparison pages, fee tables and specification tables are natural candidates.

Finish with an FAQ whose answers stand alone

Write four to six questions a buyer would type, each answered in one to three plain sentences that make sense without the rest of the page. FAQ markup is optional; the visible answer is what gets quoted. See structured data for AI search for what the markup does and does not do.

What should you avoid?

• Keyword stuffing. It scored below the unchanged page in the GEO study.
• Invented or undated numbers. They mislead readers and are easy to disprove.
• Burying the answer under a long introduction about your company.
• Answers that rely on context, such as "as mentioned above", which break when quoted alone.
• Text that only appears after JavaScript runs. Some AI crawlers will never see it; check with the test in is your website readable by ChatGPT?

A checklist before you publish

1. Does the first sentence answer the question in the page's title?
2. Does every number have a date and a source in the same sentence?
3. Is there at least one quotation from a real, named person?
4. Is every outside fact linked to its primary source?
5. Is any comparison set out as a table?
6. Do the FAQ answers make sense quoted alone?
7. Is the text in the HTML your server sends?

Then measure. Ask the same buyer questions on several assistants before and after you publish, and count whether you are named and cited; how to measure GEO explains how to do that without fooling yourself. For the wider strategy, start with generative engine optimisation for South African brands.