Research: What Profound’s Public Disclosures Say About Its 1.9 Billion Conversation Dataset, Prompt Volumes and Index Methodology

This page brings together Profound’s public statements about its conversation dataset, Prompt Volumes, Answer Engine Insights and Profound Index into a single source record as of September 15, 2026.

It is based on first-party Profound material only, including product pages, Help Center documentation, engineering posts, research, public statements and press releases. Verbatim quotations are marked. Where different disclosures need to be read together to understand how the products relate, that relationship is stated. Where the public record does not establish a point clearly, it is left unresolved rather than inferred.

The aim is to document what Profound has publicly disclosed, distinguish directly observed data from modelled estimates and benchmark construction, and identify which methodological details cannot be determined from the available material.

This review follows a set of 31 methodology questions published in July 2026 for prompt-tracking platforms generally. It does not attempt to answer those questions by assumption. It records what can be established from Profound’s own disclosures and clearly marks where the public record ends.

Profound’s three products and what each measures

Profound’s public methodology spans three distinct products. Understanding which claims relate to which product is important because Profound’s public materials do not always present these distinctions explicitly.

Prompt Volumes is the discovery and research tool. It provides keyword-level volume estimates derived from observed AI conversations. The underlying data comes from real-user conversations collected via double-opt-in consumer panels. The displayed volume is a modelled population estimate, not a raw count — a characterisation confirmed by Profound’s own product page, engineering blog and CEO.

Answer Engine Insights (AEI) is the measurement and tracking tool. It sends prompts to AI systems via headless browsers and captures live front-end responses. The underlying responses are directly observed rather than forecast. Profound then calculates visibility metrics from those observed responses. In the Profound Index, the prompts supplied to AEI are derived through Profound’s published prompt-selection methodology.

The Profound Index is the public benchmark. It uses Prompt Volumes data to discover and cluster real-user prompts, filters the final set toward open-ended, brand-agnostic questions, selects representative prompts, and then runs that set weekly through AEI. Brand rankings are calculated from the resulting AI system responses.

The 1.9B+ real-conversation corpus is the source dataset used by Prompt Volumes. The brand-visibility measurement belongs to AEI. The Index bridges the two: the corpus informs prompt selection, and AEI produces the ranking. Claims that attach the headline corpus size to the Index or to brand-visibility scores are describing the relationship between these products, not a single measurement calculated from 1.9 billion observations.

Corpus size claims over time

Profound’s public descriptions of its dataset scale have progressed as follows:

Observations: Profound’s published figures show substantial growth in the scale of its real-user dataset over time. CEO James Cadwallader described “more than 1B real prompts, growing by about 150M each month.”^[12] However, current Profound material also describes 1.9B+ conversations within a six-month research window. The public materials reviewed do not clearly establish whether the headline 1.9B figure represents an all-time cumulative corpus, a rolling analytical window, or another scope. “Prompt” and “conversation” are also used interchangeably for the same billion-scale figures across multiple pages.

Monthly scale descriptions are inconsistent and may reflect different periods or scopes

Three different descriptions of monthly data scale currently coexist in Profound’s published materials:

• “tens of millions each month” — current Help Center FAQ ^[10]
• “approximately 150 million per month” — CEO, September 2025 ^[3]
• “hundreds of millions of conversations every month” — engineering blog, November 2025 ^[4]

The current Prompt Volumes product page contains both “hundreds of millions of prompts per month” and “tens of millions each month” on the same page.^[11]

The published corpus figures suggest that growth in the hundreds of millions per month is plausible during parts of this period. What remains unclear is whether “tens of millions,” “~150M per month” and “hundreds of millions per month” refer to different periods, different source scopes, different units or simply outdated copy. The public materials do not clearly reconcile these descriptions, including the current Prompt Volumes page using both “tens of millions” and “hundreds of millions” per month.

“Prompt” and “conversation” are used for the same headline corpus figures

Profound’s engineering material explicitly treats a conversation as a multi-turn chat thread containing context. Its semantic retrieval infrastructure breaks chat threads into overlapping context windows.^[4]

Despite this, corpus-level claims frequently use “prompt” and “conversation” for the same billion-scale figure:

• May 2026: heading says “1.3B+ real user prompts,” body says “1.3B+ conversations,” growth described as “~150M prompts monthly” ^[5]
• July 2026: same page uses “1.9B+ real user prompts” and “1.9B+ real conversations” ^[7]
• June 2026: Index relaunch uses “1.5+ billion real-user conversations” and “1.5+ billion real-user prompts” within the same page ^[6]

We found no published data schema establishing that these terms refer to the same atomic record.

Prompt Volumes reports modelled population estimates

Profound’s June 2025 GA announcement states that its probabilistic model “analyzes tens of millions of real conversations” and “extrapolates frequency, intent and sentiment.” It describes correcting for demographics, geography and panel churn.^[1]

The November 2025 engineering post is more direct. It describes keyword occurrence counts being converted into a “volume projection” by scaling the panel to the global user base.^[4]

The current Prompt Volumes product page, when asked whether data is direct or modelled, says: “Both.” It describes starting with "tens of millions" of real prompts from double-opt-in consumer panels, then applying statistical modelling to correct for demographic and geographic biases, with volumes “scaled to reflect the full population.”^[11]

CEO James Cadwallader has publicly characterised the volume data as “not perfect” and “directional,” while arguing that panel-based data can produce reliable estimates at scale.^[12]

Profound has also published an external validation exercise comparing distributions in its panel data against OpenAI’s published ChatGPT usage research, reporting “strong alignment.”^[4]

Profound’s Help Center separately states that “Every prompt in the dataset comes from real users interacting with AI answer engines” and that “Synthetic or simulated data is not used at any point.”^[10] Read alongside Profound’s other disclosures, this appears to refer to the underlying Prompt Volumes source corpus: the source records are observed, while the displayed volumes are statistically modelled and scaled. Profound documents synthetic prompt generation separately in downstream benchmark products such as its Black Friday Index.^[17]

Summary: The source prompts are observed. The displayed volume is a modelled population estimate. These are Profound’s own disclosures.

The Profound Index uses a two-stage architecture

Profound’s 2026 Index methodology describes the following process:^[6]

1. Semantically retrieve real-user prompts from the preceding six months across ChatGPT, Gemini, Perplexity and Claude
2. Filter and rank with relevance scoring, deduplication and translation
3. Cluster prompts using embeddings and query fan-outs
4. Select the top representative prompt for each cluster
5. Filter the final set toward open-ended, brand-agnostic prompts
6. Run the locked prompt set weekly through Profound’s Answer Engine Insights across major LLMs

Brand visibility is calculated from the resulting AI system responses, not from the source corpus directly.

The Summer 2026 landing page says the Index is “powered by 1.9+ billion real user conversations.”^[8] The August 26 press release headlines an “Analysis of 1.9 billion real user conversations.”^[9]

The detailed methodology shows that “powered by” means the corpus informs prompt retrieval, filtering, clustering and representative-prompt selection. The brand rankings themselves are calculated against fresh LLM responses to the resulting down-selected prompt set.

We found no public disclosure of the number of canonical prompts per industry, the number of LLM responses per ranking, or the effective sample size behind a specific leaderboard position.

The Index filters out brand-named prompts

The same methodology page contains an explicit description of how the final prompt set is filtered:^[6]

“Because the Index is designed to be a neutral industry view, we filter the final prompt set toward open-ended, brand-agnostic questions. Prompts that name a specific brand or read as too narrow to a single region are dropped.”

The following sentence states:

“The output is a prompt set that reflects how buyers in that category actually talk to AI, with no thumb on the scale for any one brand.”

These statements appear in the same section: the final prompt set is deliberately filtered to remove brand-named prompts, while Profound describes the resulting set as reflecting how buyers in the category talk to AI.

Excluding branded prompts is a defensible design choice if the intended measurement is spontaneous brand visibility within standardised, brand-agnostic category questions. It does, however, mean that the resulting benchmark does not reproduce branded interactions within the wider distribution of buyer conversations.

The resulting prompt set is therefore deliberately restricted to open-ended, brand-agnostic category discovery and excludes prompts expressing commercial intent through known brands, comparisons, alternatives, switching behaviour or branded products.

The Black Friday Index: from 400 million conversations to 50 synthetic prompts per category

The Black Friday Index provides a particularly explicit example of how corpus data is reduced to a benchmark prompt set.^[17]

The index landing page states:

“It starts with real user prompts from Prompt Volumes to identify the retail categories with the high Black Friday search volume. For each category, we synthetically generate 50 unbranded Black Friday related prompts. Half reflect broad shopping intent, such as ‘Where do I get the best Black Friday deals on laptops?’ and half include price or feature constraints, like ‘Where do I get Black Friday deals on laptops under 500 dollars?’”

The term “synthetically generate” is Profound’s own published language.

The hero section of the Black Friday Index landing page describes the index as "powered by 6+ million daily prompts and insights from hundreds of millions of real Al conversations".

A separate announcement blog post describes the Index as “Built on over 400 million real conversations”^[19] and sets out the same basic methodology: real-user data is used to identify high-volume retail categories, after which 50 unbranded prompts per category are used for the benchmark.

The disclosed methodology shows that the real-user corpus was used to identify high-volume retail categories, while the brand rankings were calculated from responses to 50 unbranded prompts per category.

Later in the same blog post, Profound attributes brand performance to cases where users asked branded questions: “When users asked ‘Is [brand] running a Black Friday deal?’ Answer Engines relied heavily on pre-existing, indexed page content. Brands with early preparation gained a measurable edge.”

That branded query could not belong to the disclosed unbranded prompt set. The article does not explain whether this conclusion came from a separate analysis of real-user conversations, an additional branded prompt set, or editorial observation.

How the Index is described outside the methodology section

The prompt clustering, representative-prompt selection and brand-agnostic filtering used by the main Index are described in the “Creating the Index prompt sets” section of the June 2026 Index blog post.^[6] The separate Black Friday Index explicitly documents synthetic generation of 50 unbranded prompts per category.^[17]

Elsewhere, the Index is described in broader terms.

Index landing page hero:^[18]

“Unlock the definitive industry leaderboard for AI Search, built on 1.5+ billion real user prompts.”

Index landing page FAQ — “What is the Profound Index?”:^[18]

“The Profound Index is a free, publicly available leaderboard that ranks brands by how often they appear in AI Search across their category. It’s powered by real user prompts and updated weekly.”

Index landing page FAQ — “How does Profound source the data behind its rankings?”:^[18]

“Profound analyzes real user conversations sourced from double opt-in consumer panels of active AI users.”

Index landing page — industries section:^[18]

“See who’s winning in AI Search across every major category, with rankings built on real user prompts and refreshed weekly across every major Answer Engine.”

Summer 2026 Index press release:^[9]

“Analysis of 1.9 billion real user conversations.”

None of these descriptions disclose that the final rankings are calculated from a filtered, standardised prompt set rather than by counting brand appearances across the underlying real-user corpus. The published methodology shows that the latter is not how the Index is calculated.

Profound’s comparison page for Adobe provides a further example.^[7] It describes Profound’s data as “demand data pulled from what your audience is asking, not a proxy for what search behavior suggests they might be asking” and states that “a modeled visibility score can diverge from what models are saying, and there’s no way to know how far off it runs without comparing it against something measured.”

Profound’s Answer Engine Insights does directly capture fresh front-end responses from AI systems. Prompt Volumes, however, is itself a modelled population estimate derived from panel observations — a characterisation confirmed by Profound’s own product page,^[11] engineering blog^[4] and CEO.^[12]

In the Adobe comparison, Profound contrasts a “modelled visibility score” with directly measured Answer Engine responses, while also describing Prompt Volumes as demand data “pulled from what your audience is asking.” Profound’s own Prompt Volumes methodology separately states that the underlying observations are statistically modelled and scaled to produce population-level volume estimates.

The 2026 commercial-conversation study

In August 2026, Profound published a study on commercial conversations in ChatGPT.^[13]

This study provides the clearest public example of how a Profound estimate is constructed:

• Used a 7.5M-conversation en-US ChatGPT sample over one year
• Measured commercial conversations per active user per week
• Extrapolated that rate against OpenAI’s reported/estimated weekly active user counts
• Ran a ±10% sensitivity check on the user-base assumption

This provides a concrete published example of Profound extrapolating an observed panel rate to a population-level estimate, consistent with its broader description of Prompt Volumes as modelled and scaled data.

The sensitivity analysis is on the assumed ChatGPT user-base size. It does not publish the sampling uncertainty of the underlying panel estimate itself.

The September 2026 persona study

On September 9, 2026, Profound published research examining how AI systems respond to persona-modified prompts.^[14]

Methodology: A single persona attribute was prepended to a base prompt: “I am [attribute value]. [prompt]”. Prompts were run across ChatGPT, Claude and Gemini over 14 days, producing 71,147 responses.

Key findings:

• Responses to different personas shared only about 1 in 4 brand mentions and 1 in 5 citations
• Changing the persona reduced brand overlap by 15 percentage points compared to baseline variation
• Income produced the clearest differences — low-income personas received resale platform recommendations; high-income personas received luxury brands
• High-income responses averaged 7.5 brand mentions vs 5.8 for low-income
• Models sometimes added assumptions absent from the prompt, such as associating older clothing shoppers with women

Profound’s own conclusions:

“Measure visibility by persona as well as platform.”

“An aggregate score can conceal whether a brand appears for the customers it most wants to reach.”

Relevance: Profound separately supports persona-conditioned analysis within Answer Engine Insights. The published Profound Index methodology, however, does not describe persona-conditioned execution or preservation of source-conversation state. Profound’s persona study demonstrates that adding a single demographic attribute can materially change which brands appear in responses.

Profound’s Answer Engine Insights measurement

Profound’s current Answer Engine Insights documentation confirms that its measurement layer is prompt-driven: Profound sends prompts to AI systems, captures the responses, and aggregates those responses into visibility metrics.^[15]

The current product page states that every tracked prompt is run daily and responses are captured from the “consumer browser experience” rather than an API.^[16]

What we did not find

Based on our review of Profound’s publicly available materials as of September 15, 2026, we did not find public documentation establishing the following at the corpus, Prompt Volumes or specific Index-ranking level, as applicable:

• The precise atomic record behind headline corpus counts (whether one “conversation” equals one turn, one thread, or another derived record)
• Whether growing threads are re-ingested or whether earlier turns are re-counted
• How conversations crossing month boundaries are assigned to monthly volume
• Whether identical or repeated prompts from the same user are counted as separate demand events
• Whether prolific users are downweighted relative to occasional users
• Panel provider identities and device-level collection mechanics
• Prompt Volumes weighting equations, inclusion probabilities, design effects and calibration targets
• Confidence intervals, margins of error or sampling-uncertainty estimates attached to Prompt Volume estimates
• Effective sample sizes behind narrow Index rankings
• Number of canonical prompts per Index industry
• Number of LLM responses supporting a specific leaderboard position
• How sensitive specific Index rankings are to alternative industry prompt selections or benchmark constructions
• Whether intent, topic and demand attribution for an individual turn uses that turn alone, a surrounding context window or the full conversation state
• How a multi-intent conversation is attributed when different turns belong to different topics or stages of a customer journey
• Whether canonical prompts preserve any of the conversational state from the real-user conversations from which they were derived
• Whether canonical Index prompts or clusters are weighted in the final ranking according to their observed Prompt Volume, cluster size or coverage of the underlying real-user corpus

These details were not located in the public materials reviewed. Their absence does not establish that Profound does not calculate or use them internally.

What the evidence does establish

Profound has a billion-scale, panel-derived dataset of real-user AI activity licensed from double-opt-in consumer panels.

It models and extrapolates that data using statistical correction for at least demographic and geographic bias.

Its CEO has characterised the resulting estimates as directional.

It has performed external distributional validation against OpenAI’s published usage data, reporting “strong alignment.”

It has published sensitivity analysis for some assumptions in specific research.

It has built both lexical and semantic retrieval infrastructure, including conversation-aware context windows.

It uses real-user conversation data to inform prompt selection for its Index rather than relying solely on hand-authored prompt lists.

Its Index rankings are calculated against fresh LLM responses to a standardised prompt set derived from that data.

The most defensible summary is: Profound has a very large, real-user, panel-derived conversation corpus. It uses that corpus as the observational foundation for statistical population-volume estimates and for semantic discovery of representative prompts. Prompt Volumes are modelled estimates grounded in those observations. The Profound Index uses the corpus to construct a category prompt set and then measures brand visibility in fresh AI system responses. The final Index set is deliberately filtered toward open-ended, brand-agnostic prompts, so its rankings describe performance within that constructed benchmark rather than the complete distribution of branded and unbranded buyer conversations. Profound’s public documentation describes several parts of that architecture in detail and includes validation and sensitivity work for some aspects of its data and measurement systems. The public materials we reviewed nevertheless do not establish the corpus counting unit, the demand weighting of the final Index panel, the effective sample size behind a specific ranking, or a claim-level uncertainty bound for a specific public Index result.