Where Prompt Narrative Intelligence Fits Within PARSE
The PARSE framework organizes the AI Product Recommendation Enablement tech stack into five distinct layers. Each layer answers a different question about how brands and products appear in AI recommendations – and what organizations can do to improve that position.
The complete model is explained in The PARSE Framework: The Tech Stack for AI Product Recommendation Enablement. This article focuses specifically on P: Prompt Narrative Intelligence.
| Letter | Layer | Primary question |
| P | Prompt Narrative Intelligence | What narratives are people asking – or should the brand systematically test? |
| A | AI-Selected SoA Citation Patterns | Which visible sources, pages, and publishers do AI systems select around the narrative? |
| R | Response Patterns (Target Narrative Position) | What do AI systems say, and how consistently does the brand receive the desired association? |
| S | SoA Corpus Semantic Association Mapping | What does a deliberately selected authoritative corpus support, contradict, or associate with the brand? |
| E | External Web Narrative Intelligence | What narratives are forming and spreading across the wider web and public conversation? |
The following diagram shows where each PARSE layer sits within a simplified AI response process.
The PARSE Tech Stack for AI Product Recommendation Enablement.
P – Prompt Narrative Intelligence
Prompt Narrative Intelligence identifies the questions, needs, comparisons, attributes, and Decision Narratives that should be monitored. It can begin with observed prompt behavior, a proxy for likely demand, or a deliberately constructed testing portfolio.
P is the entry layer of PARSE. Before a brand can measure which sources AI systems cite or how frequently its products are recommended, it needs a defensible set of narratives to test. Poor prompt selection can distort every downstream measurement by overrepresenting certain questions, overlooking meaningful variants, or treating strategically convenient prompts as though they represent market demand.
Primary question: What Decision Narrative patterns can we observe in prompts submitted by people to AI systems – and which additional narratives should the brand systematically test?
P1, P2, and P3: Three Types of Prompt Narrative Intelligence
Prompt Narrative Intelligence can be separated into three evidence types:
| Code | Prompt evidence | Meaning |
| P1 | Observed prompt narratives | Aggregated prompt activity observed from consenting users, panels, or another disclosed behavioral dataset. |
| P2 | Modeled prompt narratives | Demand inferred from search behavior, query fan-out, semantic reconstruction, referral data, or another proxy. |
| P3 | Constructed prompt narratives | Prompts selected by the customer, curated by the vendor, or synthetically generated for systematic testing. |
Clickstream and panel data should not be treated as synonyms. Clickstream describes the data – an ordered record of digital interactions. A panel describes one way participants are recruited and sampled. Panel participants do not ordinarily sit together in a controlled room; they may consent to monitoring software and use their devices normally. Clickstream can come from a panel, but it can also come from applications, browser extensions, websites, or data partners.
P1 also does not automatically mean that OpenAI, Google, or another AI provider licensed private prompt records directly to a vendor. Narr Theory has not found public documentation of such a feed for the vendors examined here. When a vendor discloses a panel, clickstream source, or another collection method, that provenance is identified. When it merely says “real prompts,” uncertainty remains.
The Prompt Narrative Intelligence Vendor Landscape
The vendors examined in this article occupy different portions of P:
| Vendor | P classification | Publicly demonstrated role |
| Profound | P1 | Panel observations and population modeling |
| Semrush | P1 + P2 + P3 | Clickstream, search-derived modeling, and custom tracking |
| Evertune | P1 + P3 | EverPanel observations and custom prompts |
| Goodie AI | Claimed P1 + P3 | Claimed real-query monitoring with constructed prompts |
| Ahrefs | P2 + P3 | Search demand transformed into AI-style questions |
| Rankscale | P2 + P3 | Semantic reconstruction and tracked prompt groups |
| Peec AI | P3 | Attribute-based experimental prompt families |
| Centium | P3 | Customer- and vendor-defined portfolios |
| Gumshoe | P3 | Persona-, use-case-, and attribute-based prompts |
| Handraise | P3 | Balanced narrative-testing methodology |
| Talkwalker | P3 | Prompt templates and automatic generation |
| Brandwatch + Trajaan | P2 + P3 | Search intelligence and monitored prompt sets |
Prompt Narrative Intelligence Vendors
Profound
Profound’s Prompt Volumes documentation says the product draws from real conversations across ChatGPT, Gemini, Claude, and Perplexity and is grounded in double-opt-in consumer panels. It reports keyword volume, topic trends, intent, demographics, related prompts, and the prompts associated with particular citations.
Mechanism: P1 panel observations plus population modeling
Profound licenses anonymized and aggregated conversations from multiple consumer panels. It says the underlying records are scrubbed of personally identifiable information and refreshed weekly. Related prompts are normalized and clustered semantically, while probabilistic modeling adjusts demographic and geographic bias and extrapolates frequency to a broader population.
Profound is therefore a clear P1 platform, but its displayed market volume is not a census of every prompt entered into an AI product. The observations come from participating panel members; the broader-market estimates are modeled from that sample. That is still materially stronger provenance disclosure than an unexplained claim of “real prompt volume.”
Semrush
Semrush says its AI Visibility database contains more than 317 million prompts and responses and sources billions of real prompts from AI-search clickstream data and Google keyword data. It groups semantically equivalent prompts into topics and estimates demand at the topic level.
Mechanism: P1 clickstream plus P2 search-derived modeling
The clickstream component belongs in P1 to the extent that it records sampled interactions with AI-search products. Semrush separately describes consented clickstream networks and its relationship with the data provider Datos. However, its public documentation does not conclusively identify the acquisition path for every AI-prompt record. The safe description is therefore P1 with partially disclosed provenance.
The Google keyword component is P2. Search demand observed in Google is transformed into an AI-search demand estimate, then combined with third-party interaction data and machine-learning models. Semrush also provides P3 functionality through custom Prompt Tracking. The overall system is a hybrid of observed interactions, search-derived proxies, semantic clustering, extrapolation, and customer-defined tests.
Evertune
Evertune says its suggested prompts are grounded in EverPanel, a proprietary consumer-intelligence panel containing more than 150 million real user conversations. Its prompt system covers unaided brand awareness, attribute-specific consumer preference, and aided word association.
Mechanism: P1 EverPanel data plus P3 custom prompts
Evertune uses its panel to identify consumer language, intent patterns, category questions, and relative topic frequency. It then translates those observations into a prompt library, while allowing clients to add exact prompts of their own.
That supports a P1 classification, with a diligence qualification: Evertune’s public material does not provide the same detail as Profound about recruitment, collection instrumentation, geographic coverage, and the conversion from raw observations into modeled population volume. The panel claim is public; the full chain of custody is not.
Goodie AI
Goodie’s Prompt Research page says it monitors millions of real customer queries across major AI platforms and analyzes conversation volume, seasonality, intent, query fan-out, and competitive opportunity. It expressly distinguishes AI queries from ordinary Google keyword behavior.
Mechanism: claimed P1 plus P3, with undisclosed provenance
Goodie publicly describes the output but does not identify a panel, clickstream supplier, browser extension, direct platform license, or other collection channel. It is also not completely clear whether “monitors” means passively observing independent user prompts, executing a large prompt library, or combining both.
The appropriate classification is therefore claimed P1, supplemented by P3 prompt construction. A prospective buyer should ask who supplies the records, whether users consented, whether displayed volumes are counted or extrapolated, how follow-up prompts are handled, and how semantically equivalent phrasings are normalized.
Ahrefs Brand Radar
Ahrefs explains that Brand Radar begins with real queries from its keyword database, expands them into natural-language questions using People Also Ask and semantic fan-out, and then runs those questions through supported AI systems. Its separate methodology article says there is no significant, reliable source that completely reveals what users prompt across proprietary AI platforms.
Mechanism: P2 search demand transformed into AI-style questions
Ahrefs estimates AI-adjusted volume from a parent Google keyword’s demand and applies platform adjustments derived from aggregated AI-referral traffic relative to Google organic traffic. It also permits exact customer-defined prompts, which adds P3 testing.
This is a strong prompt-prioritization mechanism, but it is not a direct panel of ChatGPT or Gemini prompts. The underlying demand is primarily observed in search and transformed into a plausible AI question. People Also Ask is a public Google feature; its use does not indicate that Ahrefs is owned by Google or has privileged access to private Gemini prompts.
Rankscale
Rankscale describes its Prompt Research mechanism as “Prompt Decoding,” using semantic reconstruction and model simulations to estimate the density of an intent. The company explicitly says it does not track individual users.
Mechanism: P2 semantic reconstruction plus P3 tracking
Rankscale attempts to infer the shape of prompt demand from model distributions rather than from live user telemetry. It can then turn the reconstructed themes into tracked prompt groups.
This is intellectually interesting because it may reveal semantic intent clusters missed by literal keyword tools. Nevertheless, a model-estimated prompt density is not the same as a population-level monthly count. The product belongs in P2 unless Rankscale separately documents how a probability is calibrated into observed market volume.
Peec AI
Peec recommends building groups of questions around brand attributes such as reliability, price, safety, and support. These prompt families are then used to study visibility and sentiment across models and markets.
Mechanism: P3 experimental prompt design
Peec’s value at the P layer is coverage design rather than demonstrated demand measurement. A user can establish unaided recommendation prompts, branded perception prompts, competitive comparisons, and attribute-specific questions. That can produce a rigorous testing portfolio, but it does not reveal how frequently the general population asks each question.
Centium
Centium’s Brand Perception product tests category and buying questions and groups the resulting brand descriptions into common themes. The platform can organize prompts around attributes, competitors, buyer needs, and different stages of decision-making.
Mechanism: P3 customer and vendor-defined prompt portfolios
Centium helps determine which questions will expose a desired narrative association, but its public material does not demonstrate population-level prompt telemetry. Its P capability should therefore be interpreted as prompt construction and organization, not prompt-volume observation.
Gumshoe
Gumshoe’s Product Perceptions workflow analyzes the explanations appearing in monitored AI answers and organizes them into buying criteria. That workflow begins with customer-configured or persona-oriented prompt portfolios.
Mechanism: P3 designed prompts and personas
Gumshoe can broaden a test set by generating prompts around use cases, personas, and product attributes. The resulting portfolio is useful for determining what to ask repeatedly. Public documentation does not establish that those prompts reflect measured population demand.
Handraise
Handraise’s AI brand-perception methodology recommends balanced prompt families covering factual, comparative, evaluative, narrative-specific, adverse, and open-ended questions. It treats prompt coverage as part of experimental validity.
Mechanism: P3 designed narrative tests
Handraise’s prompt layer is methodological. It creates enough variation to determine whether a brand association survives rewording, adverse framing, competitive comparison, model changes, and retrieval conditions. That is valuable P3 intelligence, but Handraise does not publicly claim that the constructed portfolio measures total prompt demand.
Talkwalker
Talkwalker’s LLM Insights allows users to create custom prompts or automatically generate lists based on goals such as brand monitoring, competitor intelligence, product feedback, trends, and risk detection.
Mechanism: P3 prompt templates and automatic generation
The product constructs an observation set and lets the user select the models to monitor. Nothing in the public product page establishes that these suggestions come from an observed panel of AI users. Talkwalker therefore belongs in P3 rather than P1.
Brandwatch and Trajaan
Brandwatch’s Trojan-powered Search Intelligence monitors search behavior across traditional, social, shopping, and generative-AI environments and tracks brands across thousands of prompts. It combines customer questions, search signals, LLM monitoring, and prompt-scale analysis.
Mechanism: P2 search intelligence plus P3 monitored prompts
Brandwatch clearly demonstrates search-derived demand and a large monitored prompt universe. Its public page is less specific about whether the generative-AI demand layer contains independently observed user prompts, modeled demand, generated prompts, or a mixture. Until that provenance is disclosed, the conservative classification is P2 plus P3—not confirmed P1.
What P Can and Cannot Establish
P identifies the Decision Narratives worth testing and, in some products, supplies evidence about their relative market importance. It can reveal observed questions, model likely demand, and create systematic coverage where observation is incomplete.
P does not establish how AI systems answer those narratives, which Sources of Authority they select, or whether the external evidence actually supports their conclusions. Those questions belong to A, R, S, and E.
The practical objective is therefore not to identify one supposedly perfect prompt. It is to build a transparent portfolio of Decision Narratives, preserve whether each narrative came from P1, P2, or P3, and avoid presenting modeled or constructed demand as though it were directly observed population behavior.
Leave a Reply