
A 6.9% share of ChatGPT prompts in a US opt-in panel involved researching a brand to buy from, and that figure is now the reference point in the ChatGPT advertising and AI-visibility pitches landing on Singapore marketing desks this quarter. As of November 2026, the practical reading is narrower than the headline: the panel records what people typed into one assistant, not what they bought afterwards, and the same research puts search far ahead on expressed buying intent. A sem agency planning the next budget cycle should treat ChatGPT visibility as a small, measured line beside search until its own enquiry data says otherwise.
The first decision a marketing lead faces is whether a figure describes the market or one sample. These are the numbers as published, with the dataset and period attached to each.
| Measure | Figure | Dataset and period |
|---|---|---|
| Prompts involving purchase research | 6.9% | Verso panel analysed by SparkToro, Jan 1–Oct 29, 2026 |
| Prompts analysed | 118,163 | same panel |
| Conversations | 41,961 | same panel |
| Panel users | 12,321 | same panel |
| Responses naming one or more brands | 12.1% | same panel |
| Google queries matching the same definition | 15%+ | SparkToro Google study, January 2023–September 2024 |
| Google searches analysed | 331,697,810 | same study, about 130K US devices |

Verso runs an opt-in connector: a US ChatGPT user links an account and agrees that the conversation history can be shared, de-identified, with research. Out of that supply SparkToro analysed 118,163 prompts inside 41,961 complete conversations, written by 12,321 panel users between Jan 1 and Oct 29, 2026. Trade coverage then reduced the study to a single line, and the single line is what most budget conversations now quote.
For a buyer, the chain matters more than the headline. Each hop narrows the evidence. Only linked accounts appear, only conversations inside that window count, and only the categories one codebook defines are counted at all. None of that makes the figure wrong; it tells you which question it can answer.
Conversations were labelled from a written codebook of fine-grained intents grouped into 12 broad categories, one label per conversation, chosen by its main goal. When someone asked for an output, such as an image or an email, the output type beat the topic. The purchase-intent question used a separate three-way split: no intent to buy, intent to buy while comparing options, and questions about one specific brand, product or company that carry a purchase decision.
Two slices make up the 6.9%. The larger slice is people comparing options; the smaller is people probing one named brand. Both sit at the questioning stage. Neither records a transaction, and neither says the person bought anything afterwards.
Before analysis, 131 heavy users were removed. They averaged more than 33 prompts per month and sent 21% of all prompts, so leaving them in would have described power users rather than typical people. Most published figures are per user, which stops a vocal minority from setting every average. For demographics, 9,675 panelists with records were weighted to 2024 US Census targets for age, gender, region, education and household income, and no category moved by more than 1.4 percentage points after weighting.
Six or seven prompts in every hundred involved a brand decision of some kind: comparing options, or asking about one company's product with buying in mind. The rest of the panel's activity was something else, which matters for anyone who has been told that an assistant is the new search box.
The same report also puts 12% of ChatGPT usage inside the buyer journey, which is a wider net than the 6.9%. Early research, comparison and a question about one named brand can all belong to that journey, while only the last two read as purchase intent. Treat the two figures as a range rather than a contradiction, and quote the one that matches the decision you are defending.
The panel checked both sides of a conversation: whether the user asked about a brand, and whether the answer named one. Responses came back with at least one brand, product or company in 12.1% of cases. The rest named none. That is the surface an AI-visibility programme can influence, and per conversation it is small; ChatGPT's own claim of 330M prompts per day in the US is what turns a small share into volume. What the panel cannot tell you is whether a named brand was chosen later, or whether the mention helped.
Images and visual design led with 16.1% of conversations, information and entertainment followed at 14.0%, and shopping, brands and local services took 5.7%.
| Category | Share of conversations |
|---|---|
| Images and visual design | 16.1% |
| Information and entertainment | 14.0% |
| Shopping, brands and local services | 5.7% |
A shopping query is not the same as a purchase-intent query. Asking for opening hours or a phone number lands in that category without any decision to buy, while the 6.9% is the stricter cut. That difference is why the panel's shopping share is a weaker signal than the purchase-intent share for anyone weighing a sem agency retainer against an assistant-layer test. The gap between the two explains why two reasonable people can quote different ChatGPT numbers in the same meeting and both be right about a different thing.

SparkToro asked a comparable question of Google two years earlier. That study covered 331,697,810 searches for 320,775 unique query terms, from a panel of about 130K US devices, over 21 consecutive months from January 2023 to September 2024. Within it, 14.5% of searches were commercial, 0.69% were transactional, and 44% were for branded terms. On the purchase-intent definition later applied to ChatGPT, 15%+ of Google queries matched.
Three comparisons follow from that pairing. Google is 5.5X bigger than ChatGPT in total searches to prompts, about 2.8X bigger in the share of queries carrying purchase intent, and, on the researcher's conservative estimate, asked purchase-journey questions at 16X+ the rate of ChatGPT.
The denominators differ, and that is where most budget decks go wrong. 6.9% is a share of prompts inside a conversation. 0.69% is a share of individual searches. A conversation can hold several prompts, and a search is one event, so placing both numbers in one chart compares a stack of questions with a single query. What the two studies support is a direction: expressed buying intent is much denser in search than in an assistant's conversation mix. They do not support a claim that one channel sells more than the other, because neither study followed a purchase to its end.
The panel is evidence about a US consumer sample with light ChatGPT use. It is not local demand data, and no Singapore benchmark should be built from it directly. What a google ads singapore plan can take from it is the split between the two stages, not a local percentage. What travels is the shape of the finding: an assistant holds a lot of questions before intent forms, and search holds the moment intent is declared. The decision this quarter is not ChatGPT or search; it is how much of the budget the assistant layer deserves before your own numbers justify more.
Topkee Media has sold search advertising on one argument since it was founded in Hong Kong in 2010: the audience in a keyword ad already has a declared need. The company became a Google Partner in 2012 and a Google Premier Partner in 2016, now works with more than three thousand business clients, and opened its Singapore entity, Topkee Pte. Ltd., at Raffles Place in 2024. Its Performance Google Ads service audits impressions and clicks, researches keywords against competitors, builds landing pages that match the ad and keyword intent, and reports on the account in cycles. ROAS is measured one way only: closed-deal value divided by ad spend.
The AI layer is measured differently. Topkee runs prompt monitoring on a fixed sample, and each report states its monitoring period, prompt version, engine, sample size and failure count, alongside brand mentions, description tendency, competitor presence and cited sources. Plan capacity counts concurrently active prompts rather than monthly query volume, and prompt versions are retained so a before-and-after comparison stays honest. Google Search Console's generative-AI report covers AI Overviews and AI Mode impressions only, and cannot be extended to full prompts, internal fan-out or conversions.
Read together, the two services answer different halves of the 6.9% finding. Paid search is billed against intent that is already expressed, which is why the Google comparison looks the way it does. Prompt monitoring tells you whether an assistant describes the brand accurately when someone asks at the earlier stage. Topkee's position is layered rather than either-or: be the preferred answer inside the assistant, and be present when the prospect types a buying query, with ROAS as the single accountable number. Where a conversation does carry intent, the YME Chat Agent captures it on website chat, WhatsApp and Messenger, identifies the enquiry, guides contact capture, escalates to a human by stage, and syncs the lead to Leadbox with attribution to TTO.
The panel is opt-in, and opt-in panels skew towards younger, female and lower-income users. Weighting 9,675 panelists to 2024 Census targets is the correction, and no category moving by more than 1.4 percentage points is the evidence that the correction held. It bounds the error; it does not remove it.
Most panelists were recorded on phones, so desktop and work-account use is likely under-represented, and API traffic is missing entirely. Anyone who runs campaigns for a B2B software team should assume those users are thinner in this panel than in their own account.
The definition is self-reported. A question that looks like purchase research is not a purchase, there is no control group, and no follow-up recorded a sale. Write the finding as "in that panel, 6.9% of prompts showed purchase research", never as a claim that assistant conversations lift or depress purchases by a given percentage.
Watch the denominator. Conversations, prompts and users are three different units, and the purchase-intent share and the 12% buyer-journey figure answer different questions. Mixing either with a click-based conversion rate produces a comparison that cannot be defended in a review.
The coding carries its own uncertainty. Two labelers agreed on 90% of category labels and on 97–99% of brand tags, and there is no separate agreement figure for purchase intent. A label that two readers would call differently moves the headline share more than the market does.
Finally, the study covers ChatGPT users in the US between Jan 1 and Oct 29, 2026. It says nothing about Gemini, Perplexity or ChatGPT users in Singapore, and it is not a measurement of any agency's results.

Before the next budget review, take your own two-quarters of enquiry and deal data and split it by whether the query named a brand, compared options or asked a direct buying question. That split tells you what share of revenue still depends on declared intent, and therefore how much of the budget the assistant layer can take without gambling. If the answer is most of it, keep the search account tight and audited. If branded and comparison queries are growing, the case for prompt monitoring gets stronger, and a sem agency should be able to show you the prompt counts and report fields before you sign.
For the platform side, read Google's AI Mode ad expansion across more markets and the ChatGPT search-share reading in the SEM brief. Together they show where the ad inventory is moving and how the assistant's traffic share has been read in this market.