Lookups now rival writing in ChatGPT: what OpenAI's Signals data means for product questions

There's a running argument in our corner of the industry about whether ChatGPT is a search engine. OpenAI just published data that doesn't settle it, but moves it along. In the third version of its Signals release, out on October 26, information seeking made up 20.7% of messages on consumer ChatGPT plans in August 2026. Writing was 20.6%. That was the first month in the series, which goes back to July 2024, when looking something up came out ahead.

It didn't last. In September, writing moved back in front at 21.4% against 20.2%, as Search Engine Journal's Matt G. Southern reported on October 27. So the honest headline is not "lookups beat writing". It's that the two are now about level, after writing spent two years falling.

For a CPG brand this matters for one reason. OpenAI's definition of seeking information includes questions about products and services people can buy. Those are your questions: which oat milk foams, which deodorant has no baking soda, which protein bar has the least sugar.

What the Signals data shows

Signals is a sample of 1 million messages a month from consumer accounts on the Free, Go, Plus and Pro plans, limited to users with a declared age of 18 or over. Enterprise and Codex are excluded. LLM-based classifiers sort each message into a topic automatically, according to the release's data dictionary.

Over the full series, writing shrank and seeking information grew. Practical guidance, which covers how-to advice, tutoring and similar help, has been the largest topic every month since March 2025 and was 28.1% in September.

WritingSeeking information

Show the data

Month	Writing	Seeking information
Jul 2024	34.1%	16.7%
Jun 2025	25.7%	18.3%
Aug 2026	20.6%	20.7%
Sep 2026	21.4%	20.2%

Source: Search Engine Journal, citing OpenAI Signals v3.0, October 27, 2026. Sample of 1 million messages a month from adult Free, Go, Plus and Pro users. Enterprise and Codex excluded. Later Signals versions may revise these months.

Two details in the chart are easy to miss. The first is how steady the decline in writing has been: from 34.1% to 21.4% over 26 months. The second is that seeking information has risen less steeply, from 16.7% to 20.2%. Most of the convergence came from writing falling, not from lookups taking off.

Lookups are growing at work too

SEJ also pulled out the work-related slice. Work messages were 32.2% of all consumer messages in September. Within them, seeking information sat between 10.7% and 12.1% every month from July 2024 to July 2026, then rose to 13.9% in August and 15.2% in September. Two months isn't a trend, but it's the sharpest move in that series.

Show the data

Month	Share of work-related messages
Range, Jul 2024 to Jul 2026 (high end)	12.1%
Aug 2026	13.9%
Sep 2026	15.2%

Source: Search Engine Journal, citing OpenAI Signals v3.0, October 27, 2026. The first bar shows the top of the reported range (10.7% to 12.1%) across July 2024 to July 2026.

For B2B-ish CPG questions, such as a buyer at a grocery chain checking a supplier's certifications or a café owner comparing oat milks for a wholesale order, that's the bucket where they'd land.

Where product questions sit in the data

The data dictionary lists the fine-grained labels behind each topic. Seeking information is made of three of them.

Label	OpenAI's definition	A CPG example (ours)
specific_info	Specific information typically found on websites, including facts about people, current events and history	"Is Brand X still making the unsweetened version?"
purchasable_products	Inquiries about products or services available for purchase	"Which electrolyte powder has no sugar and no stevia?"
cooking_and_recipes	Seeking recipes, cooking instructions, or culinary advice	"How much chickpea pasta per person, and how long do I boil it?"

The catch is the one SEJ points out. OpenAI publishes the share for the whole topic, not for each label, so nobody outside OpenAI can say from this release how many messages are product questions. It could be a small slice of the 20%. It could be a big one. We'd be inventing a number if we gave you one.

What we can say is that for a food, beverage or personal care brand, two of the three labels are close to home. Recipe and cooking questions are a natural entry point for pantry brands, and product questions are the obvious one. The 2025 research paper behind this classifier described seeking information as "a very close substitute for web search", which is exactly how a shopper treats a product question.

Health, fitness, beauty and self-care questions, by the way, are not in this bucket. The dictionary files them under practical guidance, the largest topic. A question like "what should I look for in a sunscreen for sensitive skin" would most likely count there, not as a product lookup. So product-adjacent asking is spread across at least two topics.

Signals vs the 2025 paper: cite the right one

If you've seen a higher figure, it probably came from "How People Use ChatGPT", the 2025 NBER working paper by OpenAI and Harvard researchers. It reported seeking information growing from 14% to 24% of usage between July 2024 and mid-2025. Signals, using the same classifier prompts, shows 16.7% for July 2024 and 18.3% for June 2025.

Show the data

Dataset	Seeking information
2025 OpenAI and Harvard paper, about mid-2025	24%
OpenAI Signals v3.0, June 2025	18.3%

Source: NBER Working Paper 34255 (2025); Search Engine Journal citing OpenAI Signals v3.0, October 27, 2026. Different samples: the paper took one message per conversation from about 1.1 million conversations and averaged over 28-day periods; Signals samples 1 million messages a month.

That's an almost six-point gap for roughly the same period, and neither OpenAI's blog post nor the data dictionary explains it. The writing series lines up better: the paper had writing falling from 36% to 24% over its year, while Signals has it falling from 34.1% to 25.7%.

Our advice for anyone putting these numbers in a deck: name the dataset, the month and the version. The changelog shows each Signals update adds a quarter and revises earlier months, so a v3.0 figure may not match the same month in v4.0.

What the data doesn't tell you

It's worth being clear about the limits, because this kind of number tends to get stretched.

• It isn't about search or citations. Signals classifies what people ask. It doesn't say how often ChatGPT searched the web to answer or which sites it cited. A product question answered from memory counts the same as one answered with five links.
• It isn't about traffic. A rise in lookups doesn't mean a rise in clicks to brand sites. Our earlier piece on why AI traffic is the wrong first question still holds.
• It's consumer only. Enterprise accounts are excluded, and the sample is adults who declared an age.
• It's a share, not a count. Total ChatGPT usage has grown, so a flat share can still mean more messages in absolute terms, but the release gives shares, not volumes.

What a CPG brand should do with this

The useful reading is simple. A fifth of what people send ChatGPT is a request for a fact, and some of those facts are about things for sale. When the assistant answers, it either recalls what it learned in training or reads pages to check. Both routes reward the same thing: facts about your product written plainly, in the same words, in places it can read.

Write product facts as answers

Most CPG product pages are written to persuade a person who is already on the page. A lookup needs a different kind of sentence. "Contains 10 g protein and 4 g sugar per 40 g bar" is quotable. "Fuel your day the delicious way" isn't. Our tip on describing a product so assistants can match it to a question goes through the wording.

Put the specs in text, not just images

Nutrition panels, ingredient lists and certifications often live only in package photos. A model reading the page may not see them. An HTML table with the key facts is the cheapest fix we know of, and the tip on HTML tables shows the layout.

Cover the recipe and usage questions

For pantry and beverage brands, the cooking label is an opening. A short, specific usage page ("how to cook our lentil pasta so it doesn't go mushy") answers a real lookup and names the product in the answer. Keep it honest and specific, which is what gets cited.

Check what ChatGPT says, the way shoppers ask

The Signals categories are a decent template for a test list: a few fact questions about your brand, a few product-comparison questions, a few recipe or usage questions. Ask them in ChatGPT and note where it's wrong. Tools can automate this. XENA Intelligence, for example, asks assistants shopper-style questions and scores how each product shows up, and its SearchPanel view lists the actual answers per SKU. A spreadsheet works too if you only have a handful of products.

The bottom line

Signals doesn't prove that ChatGPT has become a search engine, and it doesn't tell you how many people ask about your category. It does show that looking things up is now as common in consumer ChatGPT as writing, after two years of steady movement in that direction, and that product questions are part of it. That's enough reason to treat your product page as a reference entry, not only a sales page. We'll watch the next Signals version to see whether the August crossover holds or gets revised away.