The idea that longer content is always easier for AI systems to cite is wrong. Brand Radar tracked 500 Traditional Chinese articles across ChatGPT, Perplexity, and Google AI Overviews. Articles above 3,500 words were cited less than half as often as medium-length pieces. What mattered was not total length but whether a passage could be extracted cleanly and whether its freshness signals matched the query. Here are the results from all 500 articles.
Method: 500 Articles, Three Engines, Eight Weeks
We selected 500 Traditional Chinese articles covering four topics: B2B SaaS, marketing, legal compliance, and developer tools. Their lengths ranged from 600 to 6,000 words. Each article was mapped to three to five natural queries and checked weekly for eight weeks in ChatGPT, Perplexity, and Google AI Overviews. We recorded whether the article appeared as a cited or linked source. These are Tenten's own tracking results, not figures from an outside report. The sample leans toward Traditional Chinese B2B content, which is an important boundary when interpreting the findings.
Length: The Peak Sits Between 1,500 and 2,500 Words
We divided all 500 articles into five length bands and calculated the share cited at least once. The result formed a clear inverted-U curve rather than a steady climb.
- Fewer than 800 words: 8% were cited. These pages often lacked enough information for an answer that could stand on its own, so AI systems usually skipped them.
- 800 to 1,500 words: 19% were cited. Articles began to appear as supporting sources, but rarely became the primary citation.
- 1,500 to 2,500 words: 34% were cited, the highest rate in the study. This range held a complete argument and supporting evidence without diluting the answer.
- 2,500 to 3,500 words: 27% were cited. Performance remained strong, although citations per unit of content began to fall.
- More than 3,500 words: 16% were cited. Long articles often buried the useful answer in the middle, leaving an engine to extract setup instead of the conclusion.
The lesson is not to write less. It is to stop padding an answer to reach a word count. Among articles above 3,500 words, the cited examples tended to lead each section with its conclusion before expanding. In effect, they turned one long article into several medium-length passages that could each stand alone. Length itself was not the penalty; loose structure was.
Freshness Works Differently for Time-Sensitive and Evergreen Queries
AI systems do not always prefer newer content. The effect depends on intent. We split queries mapped to the 500 articles into two groups: time-sensitive terms such as "2026," "latest," "price," "ranking," and "which tool"; and evergreen terms such as "what is," "how it works," and "how to." Freshness affected the two groups very differently.
- Time-sensitive queries: Articles updated within 90 days were cited 3.1 times as often as articles left unchanged for more than a year. AI systems were clearly filtering stale information for users.
- Evergreen queries: Publication date had almost no effect. A two-year-old explanation of GEO could still be cited when it remained accurate and well structured.
The goal is therefore to identify pages tied to time-sensitive intent, not to update the entire site constantly. Pricing pages, comparisons, annual lists, and feature directories lose citations when they go stale. Definitions, methods, and principles usually gain little from a cosmetic date change.
What Does a Citable Passage Look Like?
We examined the 10 most-cited articles passage by passage to see what AI systems actually extracted. The selected passages usually made sense without surrounding context, ran between 60 and 120 words, and stated the conclusion in the opening sentence before adding a number or concrete example. Passages that required readers to go back three paragraphs were almost never selected.
Why the Updated Date Matters More Than the Published Date
For time-sensitive queries, AI systems appear to read the visible and structured update date, not only the original publication date. Several articles had been revised slightly, but their pages still showed only a publication date from two years earlier and were treated as stale. After we added dateModified to the schema and displayed the updated date clearly, citation rates for those pages rose over the following weeks.
Most Traditional Chinese sites do not lose because their content is poor. They lose because an AI system cannot tell that a page is current or cannot extract a clean answer. Both problems can often be fixed through structure instead of a complete rewrite.
Three Changes You Can Make Now
Based on the results across these 500 articles, the following three changes offer the clearest place to start.
- Identify pages that answer time-sensitive queries, give them an update schedule, and show the updated date both on the page and in schema.
- Restructure articles above 3,500 words so every section opens with its conclusion, creating several passages that can be cited independently.
- Check every paragraph for self-containment. Rewrite sentences that make no sense without earlier context, especially the first sentence under each subheading.
Most of these fixes are structural and do not require deleting the page or retraining a model. Book a 30-minute GEO diagnostic session if you need to know whether length, freshness, or extractability is holding a page back. Brand Radar can show whether the major AI engines cite the article and where the gap begins.
