25.3% of messages sent to ChatGPT in Spain were classified as information seeking in September 2026, up from 19.1% in July 2024. Writing moved in the opposite direction, falling from 35.5% to 17.9%. The OpenAI Signals data, analyzed by Carlos González, point to a shift in how people use the assistant: they are increasingly asking it to answer questions about facts, products, services, or recipes, rather than to write for them.
That does not mean that one in four Google searches has moved to ChatGPT. OpenAI’s category groups information-seeking queries, questions about products and services for sale, and recipe help; it does not count how often ChatGPT consults web pages, links to a source, or prompts a user to click. The distinction matters when interpreting the figure and, also, for companies that depend on appearing in search engines.
From writing to asking: the shift in Spain
The Spanish Signals v3.0 series shows sustained changes over the 27 months it covers. Information seeking rose from 19.1% to 25.3% of messages between July 2024 and September 2026, reaching a high of 28.1% in August 2026. Over the same period, writing fell from 35.5% to 17.9%.
Practical guidance also gained ground, rising from 24.2% to 30.0%. OpenAI includes requests for instructions, tutoring, advice, and ideas in this category. It is the broadest category in the dataset, ahead of both information seeking and writing. The shift therefore does not simply describe the replacement of a search engine by a chatbot: it also reflects the use of ChatGPT to get guidance tailored to a task.
The global comparison offers a different perspective. In the worldwide Signals data, information seeking grew from 16.7% in July 2024 to 20.2% in September 2026; writing fell from 34.1% to 21.4%. In August 2026, information seeking overtook writing for the first time in the monthly series, by a narrow margin: 20.7% versus 20.6%. In September, writing moved slightly ahead again.
Spain stands out in this evolution: the increase in information-seeking queries and the decline in writing are more pronounced than the global average. According to González’s analysis, the trend also differs from that observed in Mexico, Argentina, and Colombia, where the share classified as information seeking fell between July 2024 and September 2026. The result is a reminder not to treat Spanish as a single market for usage.
What Signals measures—and what it leaves out
OpenAI published Signals v3.0 on October 26, 2026. The release brings together a monthly sample of one million messages from consumer accounts with a declared age of at least 18, on the Free, Go, Plus, and Pro plans, from July 2024 through September 2026. It excludes Enterprise and Codex accounts, so it does not directly represent business activity in those products.
Messages are automatically classified using language models. The information-seeking label combines three types of request: facts typically found on web pages, queries about products or services for sale, and recipes or cooking help. By combining them into a single category, the dataset does not show how much belongs to each type.
The methodology provides a signal about the subject of messages, not the path taken by a response. It does not indicate whether ChatGPT browsed, which pages it consulted, whether it included citations, or whether the user visited a site. As Search Engine Journal notes, Signals also does not break down how many queries about products, facts, and recipes fall within that category.
We publish aggregated statistics, withhold results for small groups, and add statistical noise to protect individual messages
— OpenAI
OpenAI publishes aggregated figures, suppresses small groups, and adds statistical noise to protect individual messages. This protection has a trade-off for anyone looking to interpret small differences: figures by country and month are more uncertain than global aggregates, and historical values may change when the series is updated. Version 3.0 also resampled previous months after expanding the monthly sample from the 300,000 messages used in earlier versions.
There is another caveat when comparing results. Search Engine Journal notes that Signals puts information seeking at 18.3% in June 2025, while a 2025 paper by OpenAI and Harvard calculated 24% around the middle of that year. The writing series also differ, though less. The earlier study selected conversations and averaged 28-day periods; Signals samples messages by calendar month and applies differential privacy. OpenAI has not published a reconciliation of the two figures.
A signal of demand, not a traffic counter
For publishers, retailers, and other businesses, the growth in information-seeking questions shows that ChatGPT is being used as an interface for answering queries that could also be entered into a search engine. But Signals does not allow us to conclude that those queries generate visits to websites or to quantify how much traffic they displace. Demand for answers can grow without clicks doing the same: the dataset does not measure that step.
This changes what can be inferred from a usage metric. A brand might be the subject of a question about products without the dataset revealing whether ChatGPT mentioned it, what information it used, or whether it cited the brand’s own page. The 25.3% is not a search market share or a measure of brand visibility: it is the share of consumer messages that the classifier assigned to a broad topic category.
There is also a clear limit to extrapolating the result to organizations. Although the included accounts are individual rather than corporate, some are used for work; OpenAI Signals excludes Enterprise and Codex. In the September 2026 analysis cited by Search Engine Journal, work-related messages accounted for 32.2% of the worldwide total. That figure describes the work context of consumer messages, not the use of ChatGPT within business accounts.
The evolution of the content of those messages adds context, but does not close that gap. In Spain, information seeking is approaching a quarter of usage, while practical guidance reaches 30%. The assistant appears not only as a substitute for a search box, but also as a tool for requesting a personalized explanation or guide. For content and product teams, these two behaviors raise different needs, even though Signals does not measure outcomes or response performance.
The Spanish figure and its operational limits
The most robust interpretation is a narrow one: among consumer messages classified by OpenAI, information-seeking queries gained ground in Spain while writing lost ground. The Spanish trend is more pronounced than the global one and does not match that of several Latin American markets included in González’s analysis. But neither a shared language nor a common label is enough to assume that users in those countries are looking for the same things.
The series can be used to track how usage reflected in messages changes, with the caveats of automatic classification, statistical noise, and version updates. It does not answer the question that matters most for website traffic: how many of those questions lead to an external search, an answer with a source, or a visit. For now, the Spain data runs through September 2026; that path from question to click remains outside Signals.
Sources: Xataka, Search Engine Journal, Carlos González · SEO Consultant in Seville, NBER
