The key to being cited by ChatGPT is not how high your page ranks or how high your domain weight is, but whether it can be "cleanly extracted and the risk of citation is low enough." We classified the 500 ChatGPT references we tracked for clients in the past six months one by one. The results were very concentrated: more than 80% fell into six source types. If your page does not fall into these six categories, even if you are ranked first on Google, the model will probably ignore you and instead refer to a page that is ranked eighth but is better to extract.
How do we count these 500 citations?
Only by explaining the sample clearly can we know how far the conclusion can be pushed. These 500 responses come from the actual ChatGPT responses we received when we were doing Brand Radar monitoring for B2B and SaaS customers—covering questions and answers to start a search, as well as situations where the model directly attaches source links, with topics ranging from product comparisons, terminology explanations, selection suggestions, and operational tutorials. We record the cited URL one by one, its top ten Google rankings for the question, the page structure type, and what the extracted text looks like. This is not a third-party academic study, but our own observations marked one by one on the execution side. The sample is biased towards B2B, and is biased towards queries in English and Traditional Chinese. Please look at the ratio below with this premise in mind.
Six of the easiest sources to choose from
Give the conclusion first, and then break it down one by one. Ranked from high to low by the number of citations, the six sources are roughly distributed like this -
- Official documentation and product pages (technical documentation, pricing pages, update logs): approximately 26%
- Original research and survey page with raw numbers: ~21%
- Third-party reviews, comparisons, and curated lists (not our own roundup): ~18%
- Forum and community discussions (Reddit, Stack Overflow, PTT, professional societies): about 14%
- Clear step-by-step instruction and how-to: about 12%
- High consensus reference pages (Wikipedia, industry terminology and standards pages): ~9%
It’s not surprising that the official document ranked first. When users ask "Does a certain tool support a certain function?" and "How much does a certain solution cost?", the safest way to model is to cite the manufacturer's own documents - it is first-hand, the responsibility is clear, and the chance of being corrected is low. This is why many SaaS blogs are written very frequently, but the only ones that are cited are the few pages of docs and pricing. The second place original research won by "numbers that no one else has": a survey that states the sample, method and time will be regarded by the model as a citable source of fact because it cannot be copied from elsewhere.
The third-place comparison list hits the "neutral requirements" of the model. When the question is "Which one is better, A or B?", the model tends to cite a source that does not belong to A or B, so that it does not read like a karma. So instead of writing ten self-proclaimed comparative articles, it is better to let yourself appear in the list compiled by others. The fourth-placed social discussion is the most easily underestimated: Reddit and various professional forums are often cited because it provides something that official documents cannot provide - pitfalls that real users have stepped on, and evaluations that have not been packaged in marketing language. The model needs this tone to balance the official statement.
The fifth place in operational teaching is the structure. A how-to that puts the questions in subscripts, the answers immediately after, and the steps are clearly numbered, the model can be moved almost entirely without rewriting. The sixth-ranked high-consensus reference page is the "safety card" of the model: when a fact is cross-corroborated by multiple sources, the wiki or standard page becomes the lowest-risk option for citing it. If you look at these six types, you will find a common thread.
Common points: easy to draw, checkable, and new enough
The domain weights of these six sources are very different. A Reddit reply and a certain manufacturer's docs are obviously not in the same order of magnitude, but they satisfy three things at the same time. Easy to draw means that the answer is a paragraph of text that can be established independently. There is no need for the model to read the entire page and then summarize it by itself. Verifiable means that it has a clear source, date or number, and the source can be identified by citing time stamps in the model without fear of refutation by users. Fresh enough means that the page has a clear update time, and the model prefers to reference content that does not appear to have expired. Ranking is Google’s logic, and these three things are the logic cited by AI. The two only partially overlap.
Ranking highly but not cited usually results in the following mistakes:
On the other hand, looking at those pages that rank very well but have not been cited once, the problem is almost not the quality of the content, but the difficulty of extraction. The most common ways of writing will directly exclude yourself from the citation list.
- Bury the answer in the middle of a long paragraph, with three sentences of background in front. The model cannot capture sentences that can be quoted independently.
- Without a clear release or update date, the model has no way of determining whether the content has expired.
- Key figures are only placed in pictures or table screenshots, and cannot be captured in the text layer.
- The first-person marketing tone of "our solution" throughout this article was judged to be a biased source and was omitted.
- There are ten topics crammed into one page, and none of the paragraphs can correspond to a clear question.
How to make yourself become those six sources
You don’t have to do all six, just pick the two or three that are closest to the assets you have. If you are a SaaS, first rewrite every function and every solution on the docs and pricing page into a structure of "the question is the title and the answer is the first sentence". This is the step with the highest return rate. Then squeeze out a number that only you have, even if it is just an anonymous summary of your own product backend, specify the sample and time, and it will have the opportunity to become a second source. Next, take the initiative to manage your presence in third-party lists and let others mention you in comparison articles. After completing these three steps, you will have covered the three with the highest hit rate among the six sources.
Instead of asking "How can I make ChatGPT mention me more times?", it is better to ask "Can the model on my page directly quote a sentence without changing a word?" The answer is no, which is why you haven't been cited yet.— Tenten GEO
Just do one thing first: take the question you most want to be quoted on and ask it on ChatGPT to see who it is citing now and why it is citing it instead of you. Most of the time the answer is jarring - the page being cited is not more authoritative than you, it's just easier to draw. If you want to know which level of the six sources your page is stuck on and which one you should start with, you can make an appointment for a 30-minute GEO diagnosis. We will use your own key questions and run it through for you on the spot.
