The references in AI answers depend on the content writing style or whether they are retrieved by search hits.
The reference in AI answers depends on content writing or search hits - according to the research criteria available on the day, the answer is that 'search hits come first, and writing has meaning only after that', and the two are not parallel relationships.This study's sample spans two ninety-day windows, with a total of over one hundred and fifty thousand responses and over four hundred and sixty thousand citations. The conclusion is written very directly: the search steps of the engine itself explain the main differences in citations, about thirty-two percent;And the page attributes only explain less than three points, and they are only noticeable when comparing across sites. That is to say, making the paragraph easier to摘 requires that this page has entered the engine's round of search results.
What did the study actually measure
The study compared two types of objects: pages within the same site and between different sites. The results are clearly layered. Within the same site, pages with titles and body paragraphs that match the user's real questions are cited more, while common question blocks, abstract boxes, and page speed do not show differences within the site;When compared across sites, they only show their relevance. What's more critical is the replay experiment: going through the same batch of questions through the engine interface again, each time ChatGPT references pages that are retrieved from its own search retrieval, and these search sub-queries are rewritten by the model temporarily each time the answer is given, rarely overlapping with the results of traditional search engines.
| Observation layer | Explanatory power in the study | The actions that the site owner can perform |
|---|---|---|
| Retrieval hit on the engine side | Explain the differences in the main body, significantly higher than the page properties | Allow the page to be indexed, understood, and the address to be stable |
| Ask whether the title matches the paragraph | Effective when comparing within the site | One question, one paragraph, conclusion first |
| Page structure technology (Q&A blocks, speed, etc.) | Cross-site related, almost no difference within the site | Do a good job as a basic engineering project, don't expect it to turn the table alone |
| Knowledge of the model | Affects sorting and brand naming | Ensure brand name consistency throughout the site |
Why does the search hit weight matter more
The engine does not read the entire site content when answering but first performs a lightweight search, and then organizes the answer within the few pages retrieved. If not retrieved, even if the content is neatly written, it will not appear in the reference list. The search depends on very simple inputs: whether the page has been indexed, what kind of description is left in the index, whether the address is stable, and whether the site itself is 'familiar' in the model knowledge. The research also mentions that the model will directly write in the known brand names into the search terms, and even name the brand without any supporting pages, indicating that consistent brand expression outside the site also affects the probability of being searched.
What three things can be done on the page side
The first thing is to write the 'question' in visible text. The title, description, and keywords (TDK) should be written according to the user's real questions, with the first sentence of the paragraph giving the conclusion, so that a complete and citable sentence can be extracted at one time, rather than having the answers scattered in the mixed text and image on the screen.
The second is to complete the discoverability. The site map is kept automatically generated and submitted, new pages go through the active push channel; at the same time, let the machine understand the site structure, for example, generate llms.txt according to the actual situation of the site, explaining the content organization and main entries, which belongs to the 'being understood' stage, not replacing crawling and inclusion.
The third is a stable entry. Frequent changes in addresses will make the indexed pages ineffective, the cost of remapping is very high, the revision needs to be equipped with 301 mapping, do not let the old address directly become a dead link.
The position of llms.txt in this chain
Its position is in the layer of understanding: telling the model what blocks the site has and where to start reading. It does not participate in the sorting of 'whether it is retrieved back' nor does it change the temporary sub-queries of the engine. Treat it as a reasonable configuration that adds luster, but treating it as a core method will deviate—there is no such file listed as a variable for explaining the differences in citation in the study. What really determines the citation result is the intersection of the first two things: it can be caught, and the text matches the question.
Common questions
So should we not care about the style of the content?It should be controlled, but the point of action is after the search hit. When comparing within the site to the page, whether the title and paragraph match the question is indeed a valid source of difference, the writing determines whether it can be selected and cited after the hit.
Can the fast page speed be cited more?In cross-site comparisons, relevance is visible, but almost no difference can be found within the same site. Treat it as the foundation of experience and engineering, rather than as a lever for reference growth.
What should be done first?Check for discoverability and address stability: site map, push, inclusion volume, and whether historical links are still accessible; then rewrite the title and first paragraph based on real questions; finally, consider auxiliary configurations like llms.txt.