How to Get Cited by AI: Which Factors Matter Most?

If you want ChatGPT, Claude or Google AI Mode to cite your website, there is no shortage of advice. Add FAQs. Use schema. Include statistics. Write detailed articles and keep them up to date.

But which of these recommendations are supported by what AI platforms actually cite?

To investigate, I analysed 1,680 answers from ChatGPT, Claude, Gemini and Google AI Mode. I compared their sources with Google's top 10 and examined thousands of pages the systems found, including pages they did not cite.

I quickly realised that not all common recommendations held up. Higher-ranking websites were cited more often, but only 24% of citations were the exact page that ranked. And several apparent advantages disappeared when I compared pages on the same website.

Below, I explain how I conducted the research, what the findings mean and how I would apply them to your content. Where I refer to GEO or AEO, I mean work intended to improve your visibility in AI-generated answers.

Which factors matter? A quick overview

These findings describe patterns, not guaranteed results from changing a page. All page comparisons involved content the AI systems had already found.

How I conducted the research

The research started with 120 questions asked in Dutch and English. I collected 1,680 answers across four platforms in October 2026: ChatGPT, Claude, Gemini and Google AI Mode. The full method and the first results are in my study of the sources AI cites in the Netherlands.

A citation is a link or reference identifying a source in the answer. It is different from an AI platform simply mentioning your company, as I explain in how to get mentioned by ChatGPT.

The follow-up research examined three questions.

1. Do AI platforms cite the same websites that rank in Google?

I turned the questions into search keywords and collected Google's organic top 10 through Semrush's Netherlands database. Organic results are the regular search listings, excluding ads.

I checked whether each AI citation pointed to:

• A website appearing in the top 10.
• The exact page appearing in the top 10.

Those are different measurements. AI might cite a buying guide from a website whose homepage ranks in Google. That would count as the same website, but not the same page.

The main Dutch comparison covered 71 keywords with exact matches in Semrush and 1,703 citations. The broader study contained more answers, but not every question could be included in this particular comparison.

2. What distinguishes cited pages from other pages the systems found?

For ChatGPT, Claude and Google AI Mode, the records included pages returned during their research that did not appear as citations. This gave me a useful comparison: pages that were available as potential sources but were not visibly cited.

Gemini was excluded from this part because those additional pages were unavailable.

I crawled 7,545 pages in November, collecting information such as word count, tables, lists, dates, author information and structured data. I then compared cited and non-cited pages in three ways:

The last comparison needs a little explanation.

Imagine that frequently cited publishers put an author biography on every article. Comparing those publishers with other websites might make biographies look like the reason for their success. But the AI platform could be selecting those publishers for other reasons.

Comparing cited and non-cited pages on the same website helps investigate that possibility. It is not a perfect test: pages can still differ in purpose and other ways. It also tells us little about a feature when every page on a website uses it.

3. Did the cited pages match the question more closely?

I examined 1,568 pages from the same-website comparisons. Software compared the meaning of the questions with passages of about 250 words from each page. The best-matching passage contributed to the page's score.

I then checked whether question match was related to citation selection while also considering page length, numbers, tables and FAQ headings. I wrote down how I would read the results before calculating any scores.

A separate AI assessment of 186 page-question pairs helped check whether the score reflected how well a page answered. It was a useful but imperfect measure, particularly when the question and page were in different languages. It did not verify that every answer was factually correct.

What this research can tell us

It shows which characteristics accompanied citations in this dataset. It does not prove that adding a table or rewriting a page will cause more citations.

When I describe a result as clear, it means the statistical analysis provided reasonably consistent evidence of a difference. When a result is inconclusive, it means I could not reliably distinguish an advantage from no difference. It does not mean the feature has been proven useless.

Most importantly, all page comparisons concern content the systems had already found. A feature might help a page get discovered without determining whether it becomes a citation.

1. How well your page answers the question

Of all the factors tested together, how closely a page matched the question had the strongest relationship with being cited.

That relationship remained when I compared pages on the same website and when I excluded government, public, consumer, forum and similar sites. It was especially clear for ChatGPT. Claude's result was less certain, and there was not enough suitable data to reach separate conclusions for Gemini or Google AI Mode in this part of the research.

One caution: AI systems partly retrieve pages because they resemble the question. Some of this relationship may therefore reflect how the systems search, rather than something you can simply optimise.

It still supports a familiar recommendation: understand the question and provide the information needed to answer it.

But a page can discuss the right subject without answering the right question.

For example, someone asking whether they should use separate domains or subfolders for an international website needs more than definitions. They need the trade-offs, likely costs and circumstances that would change the recommendation.

What I would do: Review your pages against the questions customers actually ask. Check whether the answer is present, easy to find and specific enough to be useful. Add examples, conditions and explanations where they are missing.

Making an answer clear and direct is practical editorial advice. The research tested how closely the text matched the question; it did not prove a particular answer length, opening sentence or writing formula works best.

2. The website where the page is published

Several apparent page-level advantages became less convincing when I compared pages within the same website.

Author information, Article schema and dates illustrate this. In a comparison excluding public and similar sites, ChatGPT's cited pages had less of these features, while Claude's had more. Those differences were no longer clear when I compared pages on the same website.

The findings are consistent with the platforms selecting different websites whose pages happen to look different. They do not establish that one platform rewards a feature while another penalises it.

That is why the publishing website belongs in the discussion. An official government page, a comparison publisher and a specialist business offer different kinds of information. A source's suitability for the question could help explain its selection, although this study did not isolate that effect.

Do bigger websites have an advantage?

This research did not measure website size, traffic or an authority score as independent factors. It also did not calculate comparative citation rates for every website category. It therefore cannot establish that bigger websites are always more likely to be cited.

Reddit is a useful reminder. ChatGPT retrieved Reddit 2,293 times in the recorded data but cited it only four times. Google AI Mode cited it in 11 of its 26 recorded occurrences. It is the same large website for both platforms, yet they treated it very differently.

What I would do: Examine the sources cited for your important questions. Are they official websites, comparison publishers, specialist businesses or community discussions? Consider what makes their information useful and what your website can contribute.

This is a reason to look beyond page formatting. It is not proof that buying brand mentions or increasing a third-party authority score will increase citations. Backlinks, reviews and brand reputation were not directly tested here.

3. Your Google rankings

Higher-ranking websites were cited more often, but Google rankings did not provide a complete picture of AI visibility.

Across the Dutch questions with exact keyword matches, 40% of citations came from a website in the measured Google top 10. Only 24% were the exact ranking page.

The platforms differed substantially:

Higher positions were also associated with more citations. Websites at position one were cited by at least one platform in 61% of cases, compared with 22% at position ten. That measures whether a ranking website received a citation, rather than its share of all citations.

The findings give no reason to abandon SEO; AI SEO builds on it. They also show why a good Google ranking does not guarantee an AI citation. I describe this comparison in more detail in a Dutch article on whether SEO is enough for AI.

What I would do: Continue improving your important pages and technical setup, and measure AI citations separately from rankings.

Treat "outside the top 10" carefully. It means outside the results measured for that keyword. A page could rank for another related search. Google explains that AI Mode can run several related searches while developing an answer. The Semrush rankings also came from different dates than the AI answers.

4. Does it matter which language people ask in?

For businesses in the Netherlands, yes. In my earlier research, English questions asked from the Netherlands produced far fewer Dutch sources than the same questions in Dutch.

The overlap with Google followed the same direction. For 14 questions that could be compared exactly in both languages, 41% of citations for Dutch questions came from a top-10 website, against 26% for English questions. That is a small sample, so I treat it as an indication.

Question match mattered in both languages. The table result, discussed below, was clearer for Dutch questions than for English ones.

What I would do: If your customers in the Netherlands also search in English, check your AI visibility for English questions separately. The free Multilingual SEO Checker shows whether your language versions are reachable and properly translated. I explain why in why you need Dutch pages to reach customers in the Netherlands.

5. Tables

Tables were one of the more promising findings. Within the same website, 37% of cited pages had a table, compared with 30% of non-cited pages.

The relationship remained after accounting for question match, page length, numbers and FAQ headings. It also remained after excluding public, consumer, forum and similar sites.

The result was clearer for Claude and Dutch questions. It was inconclusive for ChatGPT and English questions. This does not establish that tables definitely work differently across those groups.

A table can bring useful facts together: specifications, prices, requirements or the differences between options. However, the study cannot tell us whether the table format itself helped or whether tables accompanied particularly useful information.

What I would do: Add a table when readers need to compare things. Use clear labels, units and conditions. Make sure it contains information worth comparing rather than simply repeating the article in a different format.

6. Keeping information up to date

Recently updated pages were more common among cited pages in several comparisons. Within the same website, the shares updated in the previous 12 months were 95% for cited pages and 91% for non-cited pages.

This provides some support for keeping content current. But most pages in both groups were already recent, and freshness was not included in the later analysis alongside question match. We therefore do not know how much of its relationship would remain in that analysis.

The result also does not show that changing a timestamp improves citation performance. A recent date is not proof that the information was meaningfully revised.

What I would do: Review information that changes, such as prices, availability, software instructions and requirements. Correct outdated details and use accurate update dates. My article on content freshness and SEO explains how to decide what to update first.

7. Article length

Longer pages initially appeared to have an advantage. In the simple same-website comparison, cited pages averaged about 1,728 words, compared with 1,572 words for non-cited pages.

But length did not show a clear advantage when tested alongside other page features, even before question match was included. With question match included, the estimate was close to no relationship. The separate ChatGPT and Claude results were also inconclusive.

This challenges the idea that reaching a particular word count makes an article more likely to be cited. It does not challenge explaining a subject thoroughly when the reader needs that detail.

What I would do: Expand a page when an important explanation is missing. Shorten it when repetition makes the answer harder to find. Let the question determine how much detail you need.

8. Numbers, statistics and original research

Cited pages contained more numbers in the simple same-website comparison. But longer pages naturally have more opportunities to include numbers.

When I looked at numbers relative to page length alongside the other factors, there was no reliable general advantage. One analysis limited to ChatGPT and excluding public and similar sites suggested a positive relationship, but the estimate was too uncertain to turn into dependable advice.

There is an important distinction here: I counted numbers. I did not measure whether a page contained original research, strong evidence or a useful statistic. A model number and an original survey finding could both contribute to the count.

These results cannot establish whether original research itself increases citations.

What I would do: Include measurements, prices and research findings when they answer the question. Explain their source, date and meaning. Do not add unrelated statistics simply to make the page look more authoritative.

9. FAQ sections

FAQs did not show a clear overall advantage in the final analysis, although there were positive results in some comparisons.

One finding is worth further investigation: in the ChatGPT analysis excluding public, consumer, forum and similar sites, FAQ headings were associated with citations even after question match was considered. That result was not clear in the broader ChatGPT analysis or the corresponding Claude analyses.

This is more nuanced than saying every page needs an FAQ section. The research detected FAQ headings; it did not assess how useful the answers were.

What I would do: Add FAQs for genuine follow-up questions, exceptions and practical concerns. Explain the central answer in the main content. An FAQ that supplies missing information is useful; repeating the introduction as several questions adds little for readers.

10. Factors that showed no consistent advantage

The following features are often recommended for AI visibility. In my comparisons, none of them consistently distinguished cited pages from other retrieved pages on the same website.

Headings written as questions

Question headings appeared on 67% of cited pages and 63% of non-cited pages in the broad comparison, but the difference was no longer clear within the same website. Make headings specific and descriptive. "Does this lock work during an internet outage?" helps readers, but the text beneath it still needs to answer it.

Bullet points and numbered lists

Cited pages initially had fewer lists, but that difference disappeared within the same website. Use numbered lists for steps, bullets for points that are easier to scan together and paragraphs for explanations that need connected reasoning.

Schema and structured data

Structured data, often called schema markup, is code that labels information on a page.

Most pages already used some structured data, and I measured its presence rather than its accuracy. The Product schema result does not mean you should remove it: a product page and a guide may serve different questions, and some findings can occur by chance when many features are tested. Keep appropriate markup accurate, and be cautious of promises that schema reliably increases AI citations. Google states that there is no special schema.org structured data you need to add for its AI features. My structured data service focuses on accuracy and trust signals for that reason.

Author names, biographies and experience

Author information appeared on about 51% of both cited and non-cited pages within the same website. The initially opposite ChatGPT and Claude patterns were no longer clear there. A name or biography does not tell us how much the author knows, so this does not make expertise irrelevant. Identify authors honestly and demonstrate experience through examples and well-supported explanations.

Publication and update dates

Having a date is different from having current information. Dates appeared on 64% of cited pages and 65% of non-cited pages in the same-website comparison. Show accurate dates when timing helps readers, and review the substance before changing an update date.

Internal links

Cited pages initially had fewer internal links, but that difference was no longer clear in the more specific comparisons. The pages had already been found, so this cannot tell us how links helped them become discoverable. Google lists making content easily findable through internal links among the SEO fundamentals for its AI features, so keep linking important pages from relevant places.

External links and references

External links also showed no clear advantage within the same website, despite initially appearing less often on cited pages. Counting links does not assess the quality of the evidence, and backlinks from other websites were not measured. Link to reliable evidence where it supports your claims.

Meta descriptions

In the broad comparison, 92% of cited pages had a meta description, compared with 88% of non-cited pages. After excluding public and similar sites, both groups were at about 92%, and there was no clear same-website advantage. Write an accurate summary, but fix a missing answer first.

Images

Images showed no clear difference in the broad comparison (16 images on cited pages against 17 on non-cited pages) or within the same website (18 against 16). Detecting an image does not distinguish an original diagram from a stock photograph, so add visuals that explain something and make essential information available in text as well.

11. PDFs, videos and community websites

Being found during research did not necessarily lead to a visible citation.

ChatGPT's records illustrate the difference:

These are recorded occurrences, not necessarily unique pages. They do not show whether uncited material influenced the final answer.

PDFs represented about 16% of ChatGPT's retrieved-only pages, compared with 2% of cited pages. That difference could involve the subject matter, source type or technical access, rather than the file format alone.

What I would do: If useful information exists only in a PDF or video, consider publishing an accompanying webpage that explains the main points. Keep access to the original material so readers can examine it in full.

This is a practical way to make information accessible. It is not proof that converting a PDF into a webpage will produce more citations, or that community and video marketing lack value.

Where I would focus your time

The most useful starting point is to examine what your pages actually help people understand or decide.

Can someone find the answer? Does it address their circumstances? Are the examples and evidence useful? Is the information accurate today?

Then examine the sources the platforms cite. If a question consistently brings up official guidance or comparisons across several providers, consider why those sources fit. Adding bullet points to a service page may not address that difference.

My priorities would be:

1. Check access. Make sure the relevant systems can reach your important content. For Google's AI search features, supporting pages must be indexed and eligible for a search snippet. OpenAI uses OAI-SearchBot for search, separately from its GPTBot training crawler. Access allows consideration; it does not guarantee a citation.
2. Choose useful questions. Focus on what your audience needs and where your business has something relevant to contribute.
3. Improve the answer. Add missing explanations, conditions, examples and evidence. Make the information easy to follow.
4. Choose the right presentation. Use tables, steps, headings and FAQs when they help communicate the answer.
5. Maintain accuracy and credibility. Review changing information, explain relevant experience and keep technical details accurate.
6. Measure the platforms separately. Record citations over repeated checks, and distinguish a mention from a link, a visit or an enquiry.

The results are a reason to question universal GEO checklists. They give stronger support to matching the information people need than to treating individual formatting or metadata features as reliable ways to get cited. If you want to know how AEO and GEO differ, I explain that separately.

Important limits to the findings

The research used Dutch and English questions with a Netherlands-focused Google comparison. It should not automatically be generalised to every language, industry or later model version.

Pages were crawled weeks after the AI answers, and the Google rankings came from Semrush's own collection dates. Some pages could not be fetched or assessed. Many features were tested, so some apparent differences could have occurred by chance. The website filter was partly created after seeing the initial results; the later question-match test was planned before its scores were calculated.

The question-match result may be partly circular, because AI systems retrieve pages partly by their similarity to the question.

Gemini was absent from the cited-versus-retrieved page comparisons. Google AI Mode contributed to the combined page results but had too few same-website groups for its own final model. Google AI Overviews was not tested.

To find out whether a particular change increases citations, the next step is to change pages deliberately and measure the results against suitable unchanged pages over repeated observations. This research helps identify what is worth investigating; it does not replace that test.