.es domains appeared in 155 of 319 sampled OpenAI API responses with citations

An .es reference in almost half the sample

In the Bee LLM Observatory sample, 155 of 319 OpenAI API responses with citations included at least one hostname ending in .es, a rate of 48.59%. That is a finding about references emitted in those responses, not about all OpenAI answers or the ChatGPT app. Prompts and accounts carried unequal weight, so the percentage describes this particular collection.

The observation period covers seven complete days, October 5–11, 2026, in the Europe/Madrid time zone. The publication date is October 13. The main finding is that this domain ending appeared frequently among the citations from this channel during the observed week.

What counts as an .es appearance

The measure asks a straightforward question: does the response include at least one citation with a hostname ending in .es? One qualifying reference is enough to place a response among the 155 counted. Several qualifying references do not make that response count several times. The 48.59% therefore measures responses containing an .es citation, not the share of all cited links that ended in .es.

The ending alone does not establish the language of the content, the nationality of its authors or the ownership of the website. The result cannot be described as a share of Spanish-language sources or Spanish companies. What is observed here is a feature of the cited address.

Giving each prompt equal weight barely changes the result

The primary percentage gives every response equal weight. A prompt repeated more often contributes more observations to that calculation. To check how much the result depends on this distribution, the evidence includes a second measure that gives each prompt equal weight. The OpenAI API segment contains 123 distinct prompts.

Under that adjustment, the rate is 48.51%, compared with the response-weighted 48.59%, a difference of 0.08 percentage points. The description “almost half” still holds. Their closeness indicates that balancing prompts changes little in this particular result. It does not show that accounts are balanced, or make the sample representative of general use.

A citation does not establish verification

For readers, the finding helps separate two questions: which references appear alongside an answer, and what we know about the process that produced it. This measurement addresses the first. A cited URL is an emitted reference; these aggregates do not establish that the model accessed the page, reviewed it or verified its contents.

The measurement also does not assess whether each reference correctly supports the accompanying claim, or how much it influences the answer. Interpretation needs to stay within that boundary: .es addresses appeared frequently in the sampled responses with citations. The percentage is not a score for quality, accuracy or trustworthiness.

The channels do not form a controlled comparison

The evidence also includes segments for ChatGPT web, Gemini and Perplexity API. Their observations come from different combinations of prompts and collection channels. Differences between segments are therefore descriptive. They cannot rank systems by an assumed preference for .es domains when answering the same questions.

The distinction between channels matters particularly here. OpenAI API identifies the collection channel and is not equivalent to the consumer ChatGPT app. The number of observations cannot be read as market share, either. If an engine is absent from the published segments, that absence does not mean it produced zero .es citations.

How the denominator was built

The measured collection contained 1,394 valid responses from known engines; 1,171 had at least one valid citation. Those cited responses covered 125 prompts, 27 projects and 25 accounts. These are coverage figures for the full collection, not for the OpenAI API segment. The denominator behind the main finding consists solely of that segment’s 319 responses with citations.

The collection included completed executions with nonempty, non-error text, retaining the latest response for each execution and stored model. Branded prompts and accounts with “demo” in their email address were excluded. Only citation records with a known collection method were used, rather than merely detected URLs. Responses without valid citations were excluded from the headline denominator; no missing citation telemetry was recorded.

What this observation supports

Published segments had to contain at least 100 responses, 10 prompts, three projects and three accounts; positive cases had to span at least three accounts. These thresholds provide coverage, but do not remove unequal account weights, repetition or differences in collection. No personal data, individual prompts or individual responses are published.

The evidence supports a narrow conclusion: in this weekly OpenAI API sample, almost half of the responses that included citations named at least one .es hostname, and giving prompts equal weight barely changed the percentage. It does not establish a trend over time or show that publishing content under this domain ending causes more citations.