SparkToro on November 17 published an analysis of 118,163 prompts that 12,321 people in the United States sent to ChatGPT between January 1 and October 29, 2026, and found that only 6.9% of the conversations showed someone researching something to buy. Most of the traffic went elsewhere: to image editing, trivia, household chores, health questions and help with emails and texts.

The research was written by Rand Fishkin, co-founder and chief executive of SparkToro, and rests on a panel run by Verso, a company that, according to the post, "makes opt-in, anonymized connections to OpenAI." Search engines publish query volumes through their advertising tools, and Google offers Google Trends for comparing interest in keywords. ChatGPT, Claude, Perplexity and Copilot publish nothing comparable, according to Fishkin, and that gap has left marketers "forced to make wild, impossible-to-verify guesses about the volume of demand for products or services in their sector that happen via an AI conversation."

The panel gives that demand a shape. Images and visual design led the 12 categories SparkToro used. Shopping, brands and local services accounted for between 5.3% and 5.7%, depending on which of the post's figures is used. ChatGPT's replies named at least one brand, product or company in 12.1% of conversations. Fishkin's verdict is blunt: ChatGPT is "a heavily-consumer app," and "the competition with Google has been massively overblown by marketing press and influencers."

In Short

SparkToro studied about 42,000 ChatGPT conversations from a US panel and found that only 6.9% involved someone researching a purchase, while 12.1% of ChatGPT's answers named a brand, product or company. For brands paying to be named in AI answers, or buying ChatGPT ads, that makes the commercial slice of ChatGPT much smaller than the one SparkToro previously measured on Google, where more than 15% of queries met the same test. The panel leaves out API traffic, likely under-counts desktop and work use, and some of the post's own figures disagree with each other, so the study describes what consumers ask far better than how much they ask.

Where the data comes from

Verso's dataset comes from an opt-in panel of US ChatGPT users who consented to share their chat history. According to the post, Verso co-founder Aaron Aknin approached Fishkin and supplied 118,163 prompts across 41,961 complete, anonymized conversations. Verso collected rich demographic data - gender, age, household income in bands and region - for most of the users.

The heaviest users were taken out first. "Before analyzing anything, we removed the 131 heaviest users, anyone who averaged more than 33 prompts per active month," the methodology states. "They were only 1% of panelists but sent 21% of all prompts, so leaving them in would have made the results describe a small group of power users rather than typical people." The post is not fully consistent about whether the 12,321 figure includes those 131 people. Its methodology says the dataset covers 12,321 people and that the removal happened afterwards, while its statistics section describes the 12,321-user group as "the whole entity (minus the ones I removed)."

Most of the charts use a subset weighted to US Census figures rather than the full panel. "Like any opt-in panel, this one doesn't perfectly match the US population. It skews younger, female, and lower-income," according to the methodology. SparkToro weighted the 9,675 panelists who had demographic records to 2024 US Census targets for age, gender, region, education and household income, and found that "No category moved by more than 1.4 percentage points." The captions on the charts list age, gender, region and household income, without education.

Two limits shape every figure. The panel cannot see requests made through OpenAI's API, which developers use to build OpenAI's models into their own software. And "Most panelists were recorded on phones, so desktop and work-account use is likely underrepresented (though we controlled for this using estimation)," the post states. The panelists were also light users. The full panel averaged 3.4 conversations and 9.6 prompts per person over ten months - by PPC Land's calculation, roughly one conversation every three months. Fishkin acknowledges that the panel "uses ChatGPT far less often" than ChatGPT's own claim of 330 million prompts a day in the US would suggest, and the methodology concludes that "this panel tells us what people ask ChatGPT far better than how much they use it."

What Americans ask ChatGPT

SparkToro sorted every conversation into one of 12 categories. Images and visual design covers editing uploaded photos, creating pictures from descriptions and designing flyers, logos, invitations and tattoos - including, in Fishkin's words, "the much-hated AI-generated posters and sandwich boards every mediocre bar in America seems to be using." Information and entertainment gathers curiosity questions, films, music, sports scores, news, betting picks, astrology and tarot. Everyday life and admin runs from car trouble and recipes to court dates, immigration paperwork and tenant rights.

Health and wellness spans symptoms, medications, lab results, diet and sleep. Money, work and business covers banking, loans, taxes, investing, job hunting and running a small business. Shopping, brands and local services is the category closest to search advertising: deciding what to buy or where to go, comparing products, finding a nearby business and its opening hours. Relationships and emotional support includes venting, loneliness and "using ChatGPT as a sounding board or stand-in therapist."

Category Share in the post's text Census-weighted chart
Images & visual design 16.1% 17.6%
Information & entertainment 14.0% 14.4%
Everyday life & admin 12.2% 10.3%
Writing & communication 11.9% 12.7%
Health & wellness 10.3% 8.5%
Money, work & business 8.5% 7.0%
Schoolwork & learning 7.6% 8.0%
Shopping, brands & local services 5.7% 5.3%
Technology: support & coding 4.4% 4.7%
Relationships & emotional support 3.4% 3.4%
Chit-chat & other 3.0% 4.3%
Creative writing & roleplay 2.9% 3.6%

The two columns do not match. Everyday life and admin is 12.2% in the text and 10.3% on the chart, and health 10.3% against 8.5%. The order changes too: the text puts everyday life third, the chart puts writing third. The text describes the data as per-user, while the chart caption describes "the percent of AI conversations." According to the methodology, SparkToro calculated results both ways and "The two views tell essentially the same story," but the post does not say which set of figures comes from which view. The broad picture survives either reading: images first, shopping in the bottom third, coding near 4% to 5%.

Tech workers stand apart

SparkToro also isolated a group it labeled tech workers, identified by combining their conversations with their demographic data. Only 70 people met that definition, about 0.6% of the panel by PPC Land's calculation, and they generated 638 conversations. On PPC Land's reading of the bar chart, technology support and coding took about 32% of their conversations, against 4.7% for the census-weighted group. They asked far less about schoolwork (under 3%) and relationships (under 1%), and made fewer image requests, about 9.5%. Fishkin's explanation for the low relationships figure is speculative: "perhaps they realize the AI firms may not always keep their data private."

The coding share for ordinary users is a long way from earlier measurements. A 2023 SparkToro clickstream study, which the post links, was titled "We Analyzed Millions of ChatGPT User Sessions: Visits are Down 29% since May, Programming Assistance is 30% of Use." OpenAI's own data shows a similar slide: technical help fell from 13.5% to 4.6% of consumer ChatGPT requests between July 2024 and September 2026, with OpenAI's coding agent Codex excluded from those figures.

Brands in prompts and in replies

SparkToro's classifier tagged whether each prompt asked for brand, product or company information, and whether ChatGPT's answer contained any.

Census-weighted, Jan-Oct 2026 Share of conversations
Prompt contained no request for brand, product or company information 86.7%
Prompt requested information on multiple brands, products or companies 6.6%
Prompt asked about a specific brand, product or company 6.7%
Response contained no brand, product or company name 87.9%
Response contained multiple brands, products or companies 7.0%
Response contained one brand, product or company name 5.1%

"No surprise, the responses contained a brand recommendation a little more often than the user requested it," Fishkin wrote. That holds for multiple brands - 7.0% of replies against 6.6% of prompts. Across both categories, prompts seeking brand information totaled 13.3% by PPC Land's calculation, slightly more than the 12.1% of replies that named at least one. The methodology adds a caveat on the prompt side: "The second rater flagged brand questions more often than the first, so our brand-question figure is probably conservative."

Fishkin set those numbers against OpenAI's published category of "seeking information," which he considers far broader than genuine brand demand. "I think they intentionally make their categories very broad to promote a belief that ChatGPT has more brand-related demand than it does," he wrote - his interpretation, which the post does not support with evidence about OpenAI's intent. He still treats the share as meaningful: "Still, 12.1% of responses containing one or more brands/products/companies is nothing to shake a stick at, especially given ChatGPT claims to have 330M prompts per day in the US." The post does not say where the 330 million figure comes from.

Purchase intent and the Google comparison

The last test was the one that matters most to advertisers. SparkToro classified each conversation for purchase intent in three groups. Conversations with no purchase intent included "make an illustration of my gnome wizard for my D&D game." Intent to buy while researching options covered prompts such as "best running shoe options for 50 year old men with plantar fasciitis." The third group analyzed a single brand, product or company with purchase-relevant questions.

Census-weighted, Jan-Oct 2026 Share of conversations
No purchase intent 93.1%
Intent to buy, still figuring out options 4.9%
Intent to transact with a specific brand, product or service 2.1%

Fishkin framed the question in terms of Google's business. "If ChatGPT has a lot of buyer-journey research, maybe it is displacing Google, and maybe its nascent advertising business has real potential," he wrote. His answer: "I suspect this is bad news for ChatGPT's advertising opportunities."

The comparison rests on SparkToro's earlier analysis of 332 million Google queries, conducted about two years ago, in which more than 15% matched the same definition. Fishkin then says Google is 5.5 times bigger than ChatGPT in total searches against prompts and about 2.8 times bigger in the share of searches with purchase intent. "Conservatively, I'd guess that people ask Google purchase-journey-related queries at 16X+ the rate of ChatGPT," he wrote, adding: "That market share number isn't dissimilar to the Google:Bing ratio."

The arithmetic deserves a closer look. Dividing 15% by 6.9% gives about 2.2, not 2.8, by PPC Land's calculation; the post does not show how the higher ratio was derived, though it describes the 15% as a floor. Nor does it give a source for the 5.5 times figure. Multiplying 5.5 by 2.8 produces 15.4, which Fishkin rounds to "16X+." Earlier estimates PPC Land has reported vary widely: Ahrefs said in February that ChatGPT handled the equivalent of 12% of Google's search volume while sending 190 times less traffic to websites, and SparkToro's own research with Datos, published in March, put Google at 73.7% of US desktop search activity in the fourth quarter of 2025, with Amazon, Bing and YouTube each ahead of ChatGPT.

OpenAI has described commercial demand differently. A working paper by Kaiser and Schulze, which PPC Land covered on November 2, relied on OpenAI's comments at Cannes, as reported by Marketing Dive, that a fifth of prompts carry commercial intent. "Commercial intent" and SparkToro's purchase classification are not the same measure, and neither party has published the two side by side. Fishkin's own conclusions add a third number: "That said, 12% of ChatGPT usage is part of the buyer journey." The post does not explain how that 12% relates to the 6.9% purchase-intent share; it is close to the 12.1% of replies that named a brand.

How the panel compares with OpenAI's own data

SparkToro also mapped its census-weighted results onto the seven categories OpenAI uses in its Signals research and compared them with OpenAI's global figures for September 2026. PPC Land has added OpenAI's own US figures for the third quarter of 2026, which PPC Land reported on October 27.

OpenAI category Verso panel, US, Jan-Oct 2026 OpenAI Signals, global, Sep 2026 OpenAI Signals, US, Q3 2026
Practical guidance 41.8% 28.1% 30.6%
Seeking information 18.3% 20.2% 21.0%
Writing 14.5% 21.4% 21.1%
Multimedia 15.1% 8.8% 7.4%
Self-expression 8.8% 8.9% -
Technical help 0.3% 4.6% 4.1%
Other/Unknown 1.3% 8.1% -

The first three columns come from SparkToro's chart. The US column comes from OpenAI's Signals files as reported by PPC Land and averages three months rather than ten.

Fishkin judged the match close enough. "Given Verso isn't capturing API requests, but OpenAI is almost certainly including them, the comparison appears quite reasonable," he wrote, and "Nothing here makes me think Verso's panel is massively skewed in some problematic direction." OpenAI's documentation describes a narrower sample, however. Version 3.0 of Signals is built from 1 million consumer ChatGPT requests a month sent by adults, with enterprise accounts and Codex excluded, according to OpenAI's README. That points to the ChatGPT app, not developer API traffic, so API requests are unlikely to explain the gaps.

Those gaps persist when the US figures are used. Practical guidance is 11.2 points higher in the Verso panel than in OpenAI's US data, multimedia is roughly double, and writing about seven points lower, by PPC Land's calculation. The 0.3% technical help share also sits oddly beside the 4.7% that SparkToro's own 12-category scheme gives technology support and coding; the post does not describe how its categories were mapped onto OpenAI's. Fishkin's caution to anyone reusing the data is that it is "likely to be more consumer-heavy and less Professional/API-heavy than ChatGPT's overall American usage."

How the classification worked

The methodology describes the process in this order.

  1. Panel. US ChatGPT users opted in through Verso and consented to share their chat history. The window ran from January 1 to October 29, 2026.
  2. Exclusions. The 131 heaviest users, those averaging more than 33 prompts per active month, were removed. They were about 1% of panelists and 21% of prompts.
  3. Codebook. A written codebook defined 34 fine-grained intents, grouped into 12 categories. "Each conversation got exactly one label, based on its main goal." When a user asked ChatGPT to produce something, "the output type won over the topic": "Write a refund email to PayPal" counted as writing and communication, not shopping, and "answer these quiz questions about the heart" as schoolwork, not health.
  4. Labeling. "Every conversation was read and labeled by an AI model (Anthropic's Claude) working from that written codebook. Labels were not assigned by keyword rules." The model saw the auto-generated conversation title, up to six user prompts, the opening of ChatGPT's first few replies, and any product or place cards and bolded names in the replies.
  5. Brand tags. The same pass recorded whether ChatGPT recommended two or more specific brands and whether the user asked about a specific brand "in a way that could steer them toward or away from it."
  6. Checks. A second, independent labeler re-coded a random sample without seeing the first labels. The two agreed on 90% of category labels (kappa of about 0.88) and on 97% to 99% of brand tags; most disagreements were between neighboring categories, such as general knowledge and schoolwork. Fishkin also hand-checked the numbers and found accuracy above 98 in 100.
  7. Weighting and views. The 9,675 panelists with demographic records were weighted to 2024 Census targets, and results were calculated both as a share of conversations and per person, "so frequent chatters don't dominate."

The post does not say how large the re-coded sample or the hand check was, or whether the second labeler was a person or another model.

Conversation statistics

Metric Full panel Census-weighted Tech workers
Users 12,321 9,675 70
Conversations 41,961 32,980 638
Prompts (user messages) 118,163 92,377 2,032
Average conversations per user (Jan-Oct) 3.4 3.5 9.1
Average prompts per user (Jan-Oct) 9.6 9.9 29.0
Average prompts per conversation 2.82 2.82 3.18
Median prompts per conversation 2 2 2
Share of single-prompt conversations 44.2% 45.3% 39.8%
Average words per prompt 26.1 24.6 26.9
Average words per ChatGPT reply 221.0 221.6 273.3
Average total words per conversation 693 693 970
Median total words per conversation 345 336 445
Reply words per prompt word 8.4x 9.0x 10.3x

The typical exchange is short. Close to half of conversations end after a single prompt, and the median runs to two. Users write about 25 words; ChatGPT answers with about 220, roughly nine words for every one it receives. Tech workers hold slightly longer conversations and get longer answers - 273 words on average - but their median conversation is still two prompts.

Fishkin's conclusions

Fishkin presents the data as imperfect but better than anything available since 2023, "when we were still able to see ChatGPT conversations via the clickstream." He draws four conclusions.

The first concerns who uses ChatGPT. "ChatGPT is a heavily-consumer app, and anyone who tells you otherwise is wrong," he wrote. "We can't see the API requests, but I'd guess that makes up 50-75% of all usage." That share is his estimate; the post gives no basis for it.

The second is about search. "It's not a search engine. People don't treat it that way. It's something else: a confidant, a creator tool, an entertainment product, an assistant." The third concedes the commercial slice, which he puts at 12% and describes as "close to the size of Bing than Google."

The fourth is aimed at the market for AI visibility, the tools and services that track whether brands appear in chatbot answers. "AI visibility matters a huge amount!" he wrote. "But NOT BECAUSE OF CHATGPT. It's Google and the Google AI Overviews and AI Mode that makes AI visibility so incredibly important and influential." Google's AI Overviews sit at the top of its search results, and Google said in May that AI Mode had crossed one billion monthly users. Fishkin's test for anyone tracking ChatGPT visibility is whether they also track rankings and AI visibility on Bing; treating ChatGPT as worth the budget while dismissing Bing, he argues, is inconsistent.

The post closes on Verso. Fishkin wrote that he is "hopeful that Verso will continue to grow," looks forward to more research with the company, and links to its consumer tool, "Your ChatGPT, in numbers."

Why it matters for advertisers

The study lands while OpenAI is selling ChatGPT as an ad channel. ChatGPT ads were extended to 89 countries on October 21, and OpenAI said more than 100,000 advertisers had run campaigns since launch, a cumulative figure that gives no count of active advertisers or spend. Adthena's tracking showed weekly unique ChatGPT advertisers rising from 7,306 to 12,075 between July 9 and August 31.

What those ads return is already contested. The Kaiser and Schulze paper found that, in US traffic, a visit from a paid ChatGPT ad generated $0.89 in revenue for shops, against $2.45 for a visit from a paid Google search ad. A survey of 1,295 US residents by Smarty Marketing, published in May, found that ChatGPT was the first choice for starting 7.26% of shopping searches, against 56.68% for Google and 28.96% for Amazon. Datos, in a report produced with Fishkin, put AI tools at 1.65% of US desktop events in the first quarter of 2026.

SparkToro's numbers add a measure of what users ask, not only where they go. How much of ChatGPT is advertising inventory at all? If 93.1% of consumer conversations carry no purchase intent, the inventory where a product ad is relevant is small, even if the overall audience is large. Ad prices and click-through rates on that slice are a separate question the study does not touch.

For brands working on visibility in AI answers, the 12.1% figure cuts both ways. Most conversations will never name a brand, yet ChatGPT now links out far more often than it did: Peec AI found that the share of ChatGPT conversations containing a clickable link to a website rose from about 10% to nearly half after early October. The study also fits a broader argument Fishkin has been making about the open web. SparkToro's analysis with Similarweb, published on November 10, found that visits to the top million websites fell 12% from the first quarter of 2022, and his book with Amanda Natividad, Zero Click Marketing, deals with marketing on a web that sends fewer visitors to websites.

What the study does not show

The post gives no figures broken down by age, income or region, although Verso collected them, and no category-level brand or purchase-intent rates. It does not describe how Verso recruits or rewards panelists, how its categories were mapped to OpenAI's, the size of the validation samples, or the sources of the 330 million and 5.5 times figures. The tech-worker group is 70 people. And because API traffic is absent and desktop use under-represented, the study cannot say how many purchase-related conversations ChatGPT handles in total.

Key facts

  • Publication: "We Analyzed 118,000 ChatGPT Prompts from 12,000+ Users to Answer: How Do Americans Use AI Tools in 2026?", SparkToro blog, by Rand Fishkin, November 17, 2026
  • Data provider: Verso (tryverso.ai), opt-in panel of US ChatGPT users; co-founder Aaron Aknin
  • Period: January 1 to October 29, 2026
  • Sample: 118,163 prompts, 41,961 conversations, 12,321 users; 131 heavy users (more than 33 prompts per active month, 21% of prompts) removed
  • Census-weighted subset: 9,675 users, 32,980 conversations, 92,377 prompts, weighted to 2024 US Census targets
  • Tech-worker subset: 70 users, 638 conversations, 2,032 prompts
  • Classification: 34 intents in 12 categories, one label per conversation, labeled by Anthropic's Claude; 90% agreement with a second labeler on categories (kappa about 0.88), 97% to 99% on brand tags
  • Headline findings (census-weighted): 6.9% of conversations with purchase intent (4.9% researching options, 2.1% about a specific brand); 12.1% of replies naming at least one brand, product or company
  • OpenAI comparison source: OpenAI Signals data (openai.com/signals/data/), global, September 2026

Timeline

Summary

Who: SparkToro and its co-founder and chief executive Rand Fishkin, using data from Verso, a company that runs an opt-in panel of US ChatGPT users; OpenAI's Signals data serves as the comparison.

What: An analysis of 118,163 prompts in 41,961 ChatGPT conversations from 12,321 US panelists, classified by Anthropic's Claude into 12 categories. Census-weighted, 6.9% of conversations showed purchase intent and 12.1% of replies named at least one brand, product or company; images and visual design was the largest category, and shopping took about 5%. Several figures in the post disagree with its charts.

When: Published November 17, 2026, covering conversations from January 1 to October 29, 2026.

Where: The United States, on SparkToro's blog; the panel was recorded mostly on phones and excludes API traffic.

Why: Chatbots publish no query volumes, leaving marketers to guess how much demand flows through AI conversations. The findings matter for advertisers now buying ChatGPT ads in 89 countries and for brands spending on AI visibility, because they suggest the commercial share of ChatGPT use is a fraction of Google's.