When AEO and GEO became the new buzzwords, many marketers started looking for a special way of writing for artificial intelligence. The data and Google itself point in a different direction: there is no secret format for AI. What does exist is a short list of things that help both a human reader and a language model understand your page, plus a few popular beliefs that the numbers simply do not support.
If you want to know how to get cited by AI in ChatGPT, Perplexity, Gemini or Google AI Overviews, this article walks through the key research, the context missing from the headline numbers and what I see working in client projects.
The short version: what the data says
- Word count has almost nothing to do with being cited in Google AI Overviews. More than half of cited pages have fewer than 1,000 words.
- Freshness helps, especially in ChatGPT. In Google AI Overviews its effect is the smallest.
- Formats such as best-of lists, comparisons and original data get cited often, particularly for commercial queries.
- Four writing principles make the biggest difference: answer first, self-contained sections, specific names instead of vague phrases and simple sentences.
- Older pages that are losing traffic are often the fastest win, as long as you update them honestly.
Finding 1: article length does not decide citations
Ahrefs analyzed 174,048 pages cited in Google AI Overviews. The correlation between word count and being cited came out at 0.04, which is practically zero (Ahrefs study on content length in AI Overviews). This is how the cited pages split by length:
| Page length | Share of cited pages |
|---|---|
| under 1,000 words | 53.4% |
| 1,000 to 2,000 words | 30.6% |
| over 2,000 words | 16% |
This does not mean short content wins. It means length alone is not why AI picks a source. Google has said the same for years in its guidance on helpful content, where one of the self-assessment questions asks whether you are writing to a particular word count because you heard Google prefers it. The answer in brackets is a flat no (Google Search Central, creating helpful content).
In practice this leads to one rule: write as long as the topic needs and not a sentence longer. A guide to resetting a password does not need 3,000 words. A thorough comparison of accounting software will not fit into 400.
Finding 2: freshness helps, but not the way people say
A second Ahrefs study compared almost 17 million citations from AI assistants, including ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews, with regular organic search results. Pages cited by AI assistants were on average 1,064 days old, while pages in organic results averaged 1,432 days. That makes AI citations 25.7% fresher (Ahrefs study on AI assistants and fresh content).
The headline number hides three details worth knowing:
- ChatGPT showed the strongest preference for newer content.
- Google AI Overviews cited content almost as old as the organic results. For Google, freshness plays the smallest role.
- The average cited page was still almost three years old. AI does not read only the latest news.
You will also see a claim that 89.7% of the most cited pages in ChatGPT were updated in 2025 and about 76% within the last 30 days. That figure comes from an analysis of the 1,000 most cited URLs. Update dates could be detected for only about half of them, and more than half of those pages (53%) were Wikipedia articles, which people edit constantly (Ahrefs analysis of the most cited ChatGPT pages). The trend held even without Wikipedia, but it is a narrow sample of top cited pages, not the average website.
Changing the date without changing the content is not enough, and Google explicitly lists it as a sign of content made for search engines rather than people. The same guidance warns that adding or removing large amounts of content just to make a site look fresh will not help.
Finding 3: the formats AI likes to cite
The third finding is about content type. AI answers frequently cite:
- best-of lists of tools, services or products,
- comparisons of one product against another,
- original data such as surveys, measurements and statistics that do not exist anywhere else,
- reviews based on concrete, first-hand experience.
The often quoted number that best-of lists make up 43.8% of pages cited in ChatGPT comes from prompts in three commercial categories: software, products and agencies. The study did not look at informational questions (Ahrefs research on best-of lists in ChatGPT). If you want to understand why AI leans on these lists and how brands get included, I cover it in detail in the article on brand mentions and AI citations.
Comparisons work because they mirror how people prompt AI: which is better, A or B. Original data works because a model answering a question needs specific numbers and tends to cite the source that published them.
Four writing principles that help people and AI alike
The good news is that you do not need a separate writing style for AI. Google says this directly in its guide to optimizing for generative AI features: you do not need to write in a special way or break your content into tiny chunks to appear in AI results (Google Search Central, AI optimization guide). The principles below are therefore not AI hacks. They are the rules of clear writing.
1. Put the answer first
In military communication this is called BLUF, short for bottom line up front (Wikipedia, BLUF). Every section should open with the answer, not with history.
Instead of opening with a sentence like over the past few years, link building has changed dramatically, write the point straight away: the most effective way to earn backlinks today is original research with data nobody else has. Context and examples come after that.
People scan web pages and give most of their attention to the beginning, which Nielsen Norman Group describes as the F-shaped reading pattern (Nielsen Norman Group research). Language models have a similar weakness. With long inputs they use information at the beginning and the end better than information in the middle, as shown by researchers from Stanford and other institutions in the paper known as Lost in the Middle (Liu et al., 2023).
2. Write sections that stand on their own
Every H2 section should make sense even if someone reads it without the rest of the article. AI systems split pages into passages, and we do not know exactly where those boundaries fall. So does a reader who lands mid-article from a jump link.
A simple test: take any section and read it in isolation. If it contains phrases like the tool mentioned above or as we said earlier, spell out what you mean. Each part should not depend on something the reader cannot see.
3. Use specific names instead of vague phrases
AI understands text through entities, meaning specific brands, products, people, places and concepts, and the relationships between them. The sentence this tool helps with SEO says nothing. The sentence Google Search Console shows which queries your site appears for, how many clicks it gets and what the click-through rate is contains a tool, a vendor, a function and metrics. That version helps the reader and the model alike.
4. Keep sentences short and clear
One idea per sentence. Subject, verb, object. If a sentence needs a second read, it is too complex. Expert content can be deep and readable at the same time.
Name your ideas
Language models tend to absorb an original idea and repeat it as common knowledge, without credit. If you create your own framework, model or metric, give it a name tied to your brand, define it clearly and use it consistently across your website, presentations, social media and podcasts.
One caution. Only name what is genuinely yours. Putting your own label on a well-known practice is marketing cosmetics, and both experts and readers see through it. Proprietary methods make sense when real experience stands behind them, which is the core of E-E-A-T. The same logic runs through a broader AEO strategy for getting recommended by AI: AI recommends sources it can attribute clearly.
How to revive decaying pages step by step
Decaying pages are articles that used to bring traffic but are slowly fading. They have backlinks and history, only the content has aged. Updating them properly is often one of the fastest ways to improve visibility in both classic search and AI answers. Ahrefs finds them with its Site Explorer tool, but the basics can be done for free in Google Search Console.
1. Find pages that lost clicks
In Search Console, open the Performance report, use the Compare option to compare the last three months with the same period a year earlier and switch to the Pages tab. Sort by the difference in clicks and look for the pages that lost the most.
2. Check that they have backlinks
In the Links report, check whether other websites link to the page. A page with backlinks and falling traffic is an ideal candidate. If it never had links, a text refresh on its own will probably not do much.
3. Find the cause of the decline
Look at which queries the page lost positions for and what now ranks above it. The information may be outdated, competitors may answer better or the search intent may have shifted. Rule out technical causes too, such as indexing problems, crawl issues or broken rendering, which I cover in the guide to technical SEO for AI search.
4. Update the content meaningfully
Fix outdated facts, prices and screenshots. Add new data, examples and sections that answer what people ask today. Remove what no longer applies. Move the answer to the top and check that every section stands on its own.
5. Only then change the date
Changing the last updated date makes sense only when the content has actually changed. Ideally, note what you updated. Then request recrawling with the URL Inspection tool so Google sees the new version sooner. If you are unsure which options really speed this up, see my overview of the best Google indexing tools.
6. Measure the result
After a few weeks, compare the page performance in Search Console. If your property shows data for AI features, keep an eye on impressions in AI Overviews and AI Mode as well.
What the popular advice leaves out
Most advice on writing content AI cites is reasonable. From my own practice, though, I would add a few points.
Freshness numbers need context. The figure of 76% of pages updated within the last 30 days sounds dramatic, but it comes from a narrow sample of the most cited URLs dominated by Wikipedia. The authors of the freshness study themselves write that most brands will get better results from creating new quality content than from updating old content very frequently. If you start rewriting every article every month, you burn time and budget.
Freshness is a weaker signal in Google than in ChatGPT. If your customers mainly use Google, evergreen topics such as how VAT works do not need a monthly refresh. For topics that change, such as pricing, regulation or software, regular updates are essential.
Google does not ask for special AI writing. Its official guide to generative AI explicitly says you do not need a special style or tiny content chunks. Good AI advice is good because it is good writing advice, not because there is a secret code for machines.
Google stresses content that is not a commodity. The same guide warns against content built on widely known information and recommends expert insight or experience that goes beyond what everyone already knows. A top ten tips list that exists in hundreds of versions may never get cited. A list backed by your own testing has a real chance.
Correlation is not a recipe. The studies on length and freshness describe associations, not causes. Treat them as hints, not guarantees.
My pre-publishing checklist
This is what I check before publishing any article that should compete in classic search and in AI answers:
- Do the intro and every section answer the question right away? If not, move the conclusion to the top.
- Can every section stand on its own? No references to things the reader cannot see.
- Does the text use specific names, numbers and terms? Replace generic phrases with facts.
- Does the page offer something unavailable elsewhere? Original data, experience, a test or a comparison.
- Does the length fit the topic? Not to hit a word count, but to answer completely.
- Are the facts current and the sources verified? For fast-changing topics, schedule regular reviews.
- Is it clear who wrote it and why they should be trusted? Author, experience, sources.
If you want an outside view on why your content is not showing up in AI answers, WooAcademy (Milan Fraňo) works remotely with companies across the EU and the US. I can run an AI search visibility audit of your existing content, point out which pages are worth reviving and help write SEO articles designed for both classic search and AI assistants.
FAQ
Does article length matter for AI citations?
Almost not at all. In an Ahrefs analysis of 174,048 pages cited in Google AI Overviews, the correlation between word count and being cited was 0.04, and more than half of the cited pages had fewer than 1,000 words. What decides is whether the page answers the question.
How often should I update content to get cited by AI?
It depends on how fast the topic changes. Pricing, regulation and software pages need regular updates, while evergreen topics can be refreshed as needed. AI assistants cite content that is on average about 25.7% fresher than organic search results, but in Google AI Overviews this difference is the smallest.
Is it enough to change the publication date?
No. Google lists changing the date without substantial content changes as a sign of content created for search engines rather than people. A real update changes the text itself, meaning the facts, examples, data or answers.
Do I need to write differently for AI than for people?
No. Google states in its guidance on generative AI features that you do not need a special writing style or tiny content chunks to appear in AI results. The same principles that help people help AI: the answer first, a clear structure, specific facts and readable sentences.
Which content formats does AI cite most often?
For commercial queries in software, products and agencies, Ahrefs data shows best-of lists are cited most, with comparisons and reviews also performing well. Pages with original data do well across topics, because AI needs concrete numbers and tends to cite the source that published them.
What are decaying pages and why update them?
Decaying pages are older articles that used to bring traffic and have backlinks but are gradually losing visibility. Because they already carry authority, an honest content update often helps them faster than writing a new article from scratch.
Does naming my method after my brand help with AI citations?
It can, if the method is genuinely yours and useful. A clearly named concept used consistently across your channels is easier for AI to connect with your brand when it appears in answers. Relabeling a widely known practice does not help.
