.es domains appeared in 47.21% of 430 observed Gemini responses with citations

A frequent suffix among the observed citations

In the Bee LLM Observatory sample, 203 of 430 Gemini responses with at least one valid citation included an address whose hostname ended in .es. That amounts to 47.21% of the responses with citations observed for this channel during the seven days examined. The segment covers 104 distinct prompts.

The finding establishes that .es addresses appeared frequently within this collection. Its scope is limited: it does not describe every Gemini response or every query from its users. Prompts and accounts carry unequal weights, and responses without valid citations are outside the denominator.

What counts as an .es appearance

The measure checks whether a response has at least one citation whose hostname ends in .es. This is a property of the web address. On its own, it cannot establish the language of the content, the nationality of its author or the ownership of the site.

Each response meeting that condition contributes once to the numerator. A response with several .es citations counts no more than one with a single such citation. Nor is 47.21% the share of links ending in .es: the same response can cite addresses with this suffix alongside addresses with other endings.

Prompt weighting changes the reading

The primary result gives every observed response equal weight. A prompt appearing more often therefore has more influence on the final percentage. This measure directly describes the responses collected, but it can reflect how frequently particular prompts were repeated.

As a sensitivity check, giving every prompt equal weight puts the .es rate at 42.94%. That is 4.27 percentage points lower. The suffix still appears frequently, although the description “almost half” depends on weighting by responses. The shift shows that weighting matters; it does not, by itself, identify which prompts or accounts explain the difference.

Evidence of presence, with clear boundaries

For readers interested in the sources accompanying AI answers, the finding is concrete: a substantial portion of the cited responses in this segment included at least one .es address. That presence was compatible with other domains also being cited within the same responses.

The measure does not assess source quality, answer accuracy or how much any page influenced the generated text. It also does not demonstrate a general preference for .es. These observations provide no basis for concluding that using this suffix or publishing on a particular site would increase the likelihood of being cited.

The wider sample and its channels

Across the channels covered, the dataset contains 2,196 responses with citations, 133 prompts, 20 projects and 17 accounts. Those figures describe the observatory's wider coverage, rather than the size of the Gemini segment. They should not be treated as representative of all AI conversations or as a measurement of market share.

Differences between channels are descriptive: they may reflect different prompts, account weights and collection methods. They are not a controlled comparison using matched prompts. The labels also identify collection channels; an API channel does not equate to a provider's consumer app. An absent or suppressed channel must not be interpreted as a zero.

How the count was assembled

Observation covers seven complete days in Europe/Madrid: from September 9, 2026, at 00:00 inclusive to September 16 at 00:00 exclusive. The publication date is September 17, 2026. Collection included completed executions, non-branded prompts and accounts whose email addresses did not contain “demo.” The latest response per execution and stored model was retained, requiring nonempty text without an error.

Of 2,885 valid responses with a known channel, 31 lacked citation telemetry, leaving 2,854 measured responses and 2,196 with citations. Only URLs recorded as citations with a known collection method were used, rather than addresses merely detected in text. Empty or missing citation records were excluded from the headline denominator. Each public domain counts at most once per response, including its subdomains.

What this result can support

Published segments must contain at least 100 responses, 10 prompts, three projects and three accounts; qualifying hits must come from at least three accounts. These minimums support aggregate reporting, but they do not remove unequal weights or turn the collection into an experiment.

The defensible conclusion is frequent .es presence within this observed segment, with a rate that is sensitive to prompt weighting. No personal data or details of individual customers, projects, prompts or responses are published.