How to write an article that AI assistants actually cite

An article gets cited when it has four things: a headline that repeats a real person's question word for word, a first paragraph that answers it in 2-3 sentences, at least one number of your own with a period and a source, and a block of customer questions at the bottom. Length, clever subheadings and keyword density barely matter.

The reason is arithmetic. A search results page holds ten or so links, and the eighth one still brings people in. An answer from ChatGPT, Claude, Perplexity or Alice names a handful of sources, and there is no eighth place. Out of everything you write, the model will take only what it does not already know: your case and your number. Reasoning about the importance of good service it will write without you, crediting no one.

What did the Princeton study say about getting cited?

Researchers from Princeton University and IIT Delhi published "GEO: Generative Engine Optimization", accepted at the KDD 2024 conference. They built GEO-bench: 10,000 queries across 25 domains, each paired with the sources a generative engine would use to answer. Then they took nine ways of editing a text and measured how each changed a source's visibility inside the answer.

Three won: adding statistics to the text, adding a quotation from a relevant source, and citing sources. In the authors' wording, those three produced a relative improvement of 30-40% on the Position-Adjusted Word Count metric (how many of your source's words made it into the answer, weighted by where in the answer they sit) and 15-30% on Subjective Impression (a separate judge model rating how relevant and authoritative the source looks). The abstract folds this into one figure: visibility goes up by as much as 40%.

Now what did not work. Keyword stuffing (padding a page with search terms) produced, in their words, "little to no improvement". The tactic half of old-school SEO rested on does nothing inside a generated answer.

The caveat without which those numbers lie. The measurements were run on the authors' own engine and separately on the live Perplexity, where visibility rose by up to 37%. The paper came out in 2024, and the engines have moved since. This is not a promise of "plus 40% customers", it is a direction: numbers, quotes and links raise your odds of landing in an answer, keyword padding does not. Primary source: arxiv.org/abs/2311.09735. Checked on 10 August 2026.

What does a headline that matches a real question look like?

The headline is the line a person types into search or dictates to an AI assistant. Not a category name, not a pun, not a phrase from your industry price list.

How not to:

The Cost of Accounting Services: Market Overview and Pricing Factors

How to:

How much does an outsourced accountant cost

This is not about elegance. People type the second line, nobody types the first. And you do not invent it: Google Trends and Yandex Wordstat show the wording people actually use, with monthly search volume next to it.

Check for two questions in one: if the word "and" sits in the middle of your headline ("how much an accountant costs and how to choose one"), that is two separate articles. Split them, or you answer neither properly.

What belongs in the first paragraph?

A complete answer to the headline question, in 2-3 sentences. It has to read on its own: cut it out, show it to someone who has not seen the rest, and they get an answer.

How not to:

The question of accounting service costs concerns many entrepreneurs. In this article we will break down what shapes the price and give recommendations on choosing a provider.

There is no answer here, only a promise of one. A model finds nothing quotable in a paragraph like this and moves on to whoever already wrote the answer down.

How to (the numbers are illustrative, they show the structure; yours come from your own contracts):

An outsourced accountant for a sole trader on the simplified tax scheme costs 4,000 to 18,000 roubles a month. The spread depends on transaction volume and on whether you have employees: with one person on payroll the price roughly doubles. Counted across 63 of our own contracts in 2025.

Answer, reason for the spread, source of the number. Any one of those sentences can be lifted and quoted on its own.

Never open with "today", "in our time", "many people face". That is a run-up, and it eats exactly the two sentences that could have got you named as a source.

Where do you find your own number if you have never run a study?

Your own number is a figure that does not appear in other articles on the topic. You do not need research for it, you need access to your own paperwork. Four places it lives:

• the price spread across your contracts over a year, plus the reason for the spread
• the calendar time from enquiry to result, counted from dates rather than from memory
• the share of enquiries that fall through, and the exact point where they fall through
• your own count: open ten competitor sites and tally how many have the thing you are writing about

The formula for the sentence: the number itself, the period, and what it was counted from. "Out of 214 enquiries in 2025, 138 got a reply the same day" works. "We reply fast" does not, because everyone writes that and there is nothing to quote.

Check against a document, not against memory, and do not round in your own favour. If a figure does not reconcile, leave it out: that is the line people will check.

Why keep a questions block if Google barely shows FAQ results anymore?

FAQPage markup (a piece of code telling the search engine the page holds questions and answers) is still sold as a way to take up more room in results. Since 8 August 2023 Google expands those blocks only for well-known, authoritative government and health sites; for everyone else, in Google's words, "this rich result will no longer be shown regularly" (Google announcement).

The questions block in the text is still worth having, just not for the snippet. People rarely ask an AI assistant in exactly the words sitting in your headline. They ask "what if I have two companies", "do you work with marketplaces", "what happens if the tax office sends a demand". Every one of those in your text is another phrasing that can surface you.

Where to get them: from your correspondence with the last ten customers, in their words. You want 5-7 questions, including the awkward ones about price, deadlines and what you do not handle. Answers of 2-4 sentences, with the answer itself in the first one.

Invented questions are obvious at a glance: they are convenient, and answering them feels pleasant.

Why do batches of AI-written articles drag down the entire site?

Because the penalty lands on the domain, not the page.

Yandex, in its webmaster guidance, lists "sites with automatically generated content that is useless to the user" and states in the same place that the restrictions "may affect the ranking of the entire site or its sections" (Yandex help). Not just the pages where the junk was found.

At Google this is a named violation, scaled content abuse: "many pages are generated for the primary purpose of manipulating search rankings and not helping users". The key clause in the policy text: unoriginal, low-value content breaks the rules "no matter how it's created", meaning it does not matter whether a model wrote it, a human wrote it, or both. The consequence is spelled out as "Sites that violate our policies may rank lower in results or not appear in results at all" (Google policy). Checked on 10 August 2026.

Then it is a question of proportion. You wrote three solid articles with your own figures and padded the site with forty generated pieces along the lines of "what is a CRM and why you need one". Those forty are what define your site to a crawler, and the three good ones dissolve in them. Same on third-party platforms: Habr, the Russian tech publisher, gives generated text its own clause in the rules. What each platform takes is covered separately in where to publish articles.

A workable pace on your own site is one article a week. Boring, but ten articles a year, each with a number of its own, do more than two hundred in a month.

How do you get a model to write from your material instead of the internet average?

The prompt "write an article about [topic]" returns the average words of the internet. Material appears when the roles swap: the model asks, you answer.

The setup. In a fresh chat, instruct it to interview you, ask one question at a time, and write nothing until you have answered them all. Questions about your case, not the topic: what the starting situation was, which numbers you can name, how long it took in calendar time, what went wrong, what customers ask and in what words. Add a separate ban on inventing: if you did not give a number, it does not appear in the article in any form, and a marker like [FACT NEEDED: which one] goes in its place.

The interview takes 15-20 minutes and works best with short replies, like texting. If numbers or names you never mentioned show up in the finished draft, do not fix them one by one: write "remove everything that was not in my answers" and reread the whole thing.

How long before this shows results?

Google's documentation says crawling a new page can take anywhere from a few days to a few weeks, and that requesting a recrawl of the same URL several times will not get it crawled any faster (Google help). That is only about getting indexed.

Nobody publishes official timelines for showing up in AI answers. From what we see at the Academy, the first mentions arrive roughly a month and a half to two months into regular publishing, faster in narrow niches. That figure is ours, from our own practice, not someone else's statistics.

If you are still working out where AI answers come from in the first place, start with the overview: GEO optimization: what it is.

The full walkthrough with ready-made prompts sits in the "Echo" block of the Academy, three lessons open free: academy.agineai.com/block/b12-eho.