When a brand's website is cited, the brand is also mentioned in 72% of AI responses.

AI models do not always rely on the same sources when answering the same question. Some pages appear in a single measurement run, while others continue to be cited in subsequent responses. Understanding content performance requires examining which page types and characteristics are associated with this consistency.

In our previous study, we examined how the sources cited by different models diverge.

This study examines repeatedly cited sources in ChatGPT, Gemini, and Perplexity through their page types and the functions they serve for users. It also explores the relationship between content length and citation frequency, as well as the connection between a brand's website being cited and its name appearing in the response.

19,769 AI responses analyzed

1,901 Unique prompts

255,143 Source URL records

39,034 Page follow-up cases

HOW WE DID THIS

Methodology

Responses generated by the same model for the same brand and prompt were compared alongside their source lists. Duplicate records were consolidated, and three areas were examined.

Brand visibility: Across eight brands, we measured whether the brand's name appeared in the response text when its official website was cited. Link URLs and reference lists were excluded from this measurement.

Repeated citations: Pages were tracked over the five runs following their first citation. Pages cited again in at least three of those runs were classified as recurring sources. The analysis covered 39,034 follow-up cases involving 23,216 pages across eight brands; the same page was evaluated separately for different prompts and models.

Content type and length: Pages were ranked by the share of runs in which they were cited. We read 30 pages from the frequently cited end of the ranking and 50 from the infrequently cited end. Inaccessible pages were replaced with the next candidates. Additional comparisons brought the total to 94 URLs, evaluated by page type, function, and the word count of their main content.

FINDING 1

Nearly half of initial citations do not recur in the next five runs

In 48.1% of the follow-up cases, a page cited for the first time did not appear again in any of the next five runs.

Appearances of the same page in the next five runs

7.9%

14.5%

29.5%

48.1%

Recurring sources (3+ of 5 runs): 22.4%

All five runs
7.9% · 3,073 cases

Three or four times
14.5% · 5,669 cases

Once or twice
29.5% · 11,511 cases

Did not appear again
48.1% · 18,781 cases

Based on 39,034 follow-up cases.

The share appearing again in at least three of the five runs was 22.4%. The share appearing in all five was 7.9%.

A single citation is not enough to establish that a page is a consistently preferred source. Consistency becomes apparent when responses to the same prompt are compared across repeated runs.

FINDING 2

AI can return to the same website while citing a different page

Tracking whether a specific page is selected again produces different results from tracking whether other content from the same website is cited.

Appeared in at least three of the next five runs

22.4%

The originally cited page itself

The same URL is cited again

about

2×

→

44.5%

Any page from the same website

The site returns, not necessarily the page

Both rates were calculated using the same 39,034 page follow-up cases.

8,611 Follow-up cases where the original page did not reach the three-run threshold, while pages from the same website address did.

Continued use of a website does not mean that an individual piece of content is being reused to the same extent. This difference shows how performance measured at the website level can obscure changes at the page level.

FINDING 3

Half of the frequently cited pages list or compare alternatives

Of the 30 frequently cited pages reviewed, 15 presented options and helped users compare them.

What the page does for users: 30 frequently cited pages, one square each

Lists and compares options 15 · 50%

Explains a specific problem or process 7 · 23.3%

Presents a service and its provider 4 · 13.3%

Offers a selection tool 1

Enables users to start an action 1

Discusses user experiences 1

Presents a brand and its product range 1

Each page was classified once according to its primary function.

The pages shared characteristics beyond their subject matter: they also served identifiable functions for users. Helping users choose, explaining a problem, and enabling an action appeared across different page formats in the frequently cited group.

FINDING 4

Category and service pages are frequently cited alongside guides

When the same 30 pages were classified by format, articles and guides accounted for 40% of the group. Other page types made up the remaining 60%.

Page type, frequently cited pages (n = 30)

Article or guide

Longer content does not necessarily get cited more often

The reviewed pages showed no consistent pattern of longer content being cited more frequently. The result changed depending on which page types were included in the comparison.

Median word count: frequently vs. infrequently cited pages

Frequently cited Infrequently cited

All page types (30 and 50 pages)

678

858

Frequently cited pages are longer

Articles & guides only (12 and 15 pages)

845

1,174

Reversed: infrequently cited articles are longer

0 300 600 900 1,200

Median word counts were used to reduce the influence of a few unusually long pieces of content.

Across all page types, the frequently cited group had longer text. However, when the comparison was limited to articles and guides, the infrequently cited content was longer.

This reversal shows how page types affect the length comparison. Grouping category, service, or form pages together with guides can create a misleading impression of the role word count plays.

To reduce differences in subject matter and model, 15 additional comparisons were made within the same brand, prompt, and model. These showed a similar pattern.

15 article pairs with the same brand, prompt, and model

10 of 15

pairs where the less-cited article was longer

Teal: the less frequently cited article was longer (10). Grey: it was shorter (5).

The findings support the conclusion that increasing word count alone does not explain more frequent citation.

FINDING 6

The brand is also mentioned in 72.2% of responses that cite its website

Across eight brands, the brand's name also appeared in the response text in 2,645 of the 3,661 responses citing its official website.

Responses citing the brand's official website (n = 3,661)

72.2% brand mentioned

Brand mentioned in the response
2,645 responses · 72.2%

Website used as a source, brand not mentioned
1,016 responses · 27.8%

Across eight brands. Link URLs and reference lists were excluded from the measurement.

This finding points to a connection between a brand's website being cited and the brand appearing in AI responses. However, using a brand's website as a source does not require the model to mention that brand in its answer.

AI may use brand content specifically to obtain information. When the prompt seeks information rather than a brand, the brand's name may therefore be absent from the response.