Why QueryBurst Does Not Use Prompt Tracking to Measure AI Visibility
QueryBurst is not a prompt tracking platform. It does not run panels of prompts against AI models, it does not count brand mentions across those responses, and it does not aggregate the results into an AI visibility score.
Our position is that prompt tracking cannot observe most of what our definition of AI visibility includes: the model's prior knowledge, the accumulated state of a conversation, the hidden research and evidence process between question and answer, and the way visibility changes as a journey develops.
We therefore reject the method, not merely the score it produces. This page explains what prompt tracking assumes and why QueryBurst chose a different approach.
What prompt tracking assumes
Prompt tracking rests on several assumptions about how AI-generated answers work and how they should be measured.
The most important are:
- That the prompt is the meaningful unit of analysis
- That running the same prompt independently and repeatedly produces a representative sample of outcomes
- That detecting a brand in the response is a meaningful proxy for AI visibility
- That aggregating those detections into a score produces a useful summary
- That changes in the score reflect changes in the brand's visibility
QueryBurst does not accept all of these assumptions.
Some are partially valid. Others are structurally problematic. The rest depend on conditions that are rarely stated.
The specific scoring methodologies used by Peec, Profound, PromptWatch, Otterly, Evertune and Athena are documented in How Prompt Tracking Platforms Calculate AI Visibility Scores.
These assumptions have been challenged publicly
In July 2026, QueryBurst founder David McSweeney published 31 Methodology Questions for AI Visibility and Prompt-Tracking Platforms and directly challenged Peec, Profound, PromptWatch, Athena, AirOps and other platforms in the category to answer them publicly. The article invited substantive responses, published research, reproducible evidence or validation studies, and offered to link to any that were provided.
The questions ask vendors to substantiate the claims they make about their products: what their scores actually measure, whether their prompt panels represent real users, how weighting and uncertainty are handled, whether execution conditions match customer experiences, whether results predict external business outcomes, and what evidence would falsify their methodology.
Two smaller platforms — Ternith and Spyglasses — published detailed public responses and are credited in the article. The major platforms named in the challenge had not provided substantive public answers as of publication.
The full set of questions remains open. QueryBurst’s September 2026 research into Profound’s conversation dataset, Prompt Volumes and Index methodology examines one platform’s published methodology in detail. Our position on why real-user data does not automatically establish representativeness addresses the wider sampling question.
Prompt tracking measures one execution environment
Prompt tracking platforms execute prompts through automated, standardised environments that differ from the state in which an individual customer may actually use the same AI system.
Depending on the platform and provider, that may mean a clean session, a logged-out or standardised account state, no persistent conversation history, no user memory, and proxy-based geographic routing.
In many cases, this is the only state accessible to automated execution at scale. But it is not the state in which most paying customers use these systems.
- A ChatGPT Pro, Plus or Max subscriber may be routed to a different model with different reasoning capabilities and different retrieval behaviour
- A logged-in user with memory enabled carries persistent context that can influence responses across sessions
- A user with conversation history in the same session receives answers shaped by accumulated state that does not exist in a clean execution
- Account-level preferences, custom instructions and prior interactions can change which brands, products and evidence the model considers relevant
- Geographic location tied to the user's real IP address and account settings may produce different retrieval results than a rotating proxy IP used by the tracking platform
Prompt tracking platforms do not generally disclose which account tier, login state, session conditions, proxy infrastructure or geographic routing their measurements use.
QueryBurst's position is that the execution environment is part of the measurement. If the environment differs from the customer's actual state, the resulting observation should be understood as evidence about that test environment — not automatically as a measurement of the customer's experience.
This is one of the reasons QueryBurst allows users to define and control their own execution context — the persona, the conversation, the accumulated state — rather than relying on a single anonymous execution environment that may not represent anyone in particular.
The prompt is not the whole input
Prompt tracking treats the prompt as the primary input.
In practice, the input to a modern AI system can include:
- the current user message
- the user's account state and preferences
- the model's prior knowledge and associations
- previous messages in the conversation
- previous assistant responses
- the retrieval state at the time of the request
- the model version and configuration
- geographic and language context
The prompt is one of those inputs.
It is not all of them.
A prompt tracker can control some of these variables. It cannot generally observe or reproduce the full state that produced an individual user's real-world interaction.
That means the measurement captures one slice of the input space while holding other variables unknown rather than constant.
QueryBurst's definition of AI visibility treats the model's prior knowledge and the accumulated conversation context as part of the visibility problem. Prompt tracking cannot observe the prior and does not model the accumulation.
Independent runs do not model conversations
Prompt trackers typically execute each prompt as an independent request with no conversation history.
Real AI usage increasingly happens inside conversations.
In a conversation, each turn inherits what came before: the user's previous questions, the assistant's previous answers, the products discussed, the objections raised, and the information introduced along the way.
Running prompts independently removes that accumulated state.
The result is a measurement of how the model responds to isolated questions, not how it responds within the kind of research journey a customer actually conducts.
QueryBurst's position is that the difference between an isolated prompt and the same words inside a conversation is not a detail. It can be the difference between a brand appearing and not appearing, being recommended and being absent, or being considered early and only becoming relevant several turns later.
The model's prior knowledge is invisible to prompt tracking
Before an AI model retrieves anything, it already has knowledge and associations derived from its training data.
Those associations can influence which brands are mentioned, which categories they are placed in, which attributes are assumed, and what the model believes to be true before any external evidence is introduced.
Prompt tracking cannot observe this prior state.
It can observe the final output, which is shaped by the prior but does not expose it.
This matters because a brand may have strong or weak prior visibility independently of whether it appears in any particular tracked response. The prior can influence what the model searches for, which evidence it finds relevant, and how it frames the answer, without being directly detectable from the output alone.
QueryBurst treats prior knowledge as a first-class part of AI visibility. Answer Engine therefore models an explicit pre-retrieval prior so that its potential influence can be investigated rather than silently folded into the final answer.
The process between question and answer is unobserved
Between receiving a prompt and generating a response, an AI system may:
- generate additional research queries
- retrieve candidate sources
- extract relevant passages
- compare evidence across sources
- resolve conflicting claims
- determine which information supports the answer
- decide which sources to cite
Prompt tracking observes the final output. It does not observe these intermediate stages.
A brand may be retrieved but not selected. A page may be selected but not grounded into the answer. Evidence may influence the response without receiving a visible citation. A source may be cited without its brand being recommended.
These distinctions matter because they represent different problems with different solutions. Prompt tracking collapses them into one binary observation: the brand appeared, or it did not.
QueryBurst’s position is that understanding where visibility is gained or lost requires inspecting the process, not only the endpoint. This is why Answer Engine makes its model of the research process inspectable rather than recording only the final output.
Adding personas to prompt tracking does not solve the structural problem
Some prompt tracking platforms now support attaching a persona or user profile to each prompt.
This adds context to the input.
It does not create conversation state.
A persona attached to an isolated prompt produces:
persona + prompt → answer
A conversation produces:
persona + turn 1 + assistant response 1 + turn 2 + assistant response 2 + … + current turn → answer
The difference is not the persona. It is the accumulated state between the persona and the current message.
Adding a persona to every prompt in a panel does not model the progressive development of context through a conversation. It does not capture the effect of the assistant's own previous answers on subsequent turns. It does not model branching, where the buyer's path diverges based on what happened earlier.
QueryBurst's position is that persona support in a prompt tracker is an improvement to the prompt, not a solution to the conversation problem. For a detailed product-level comparison, see the Answer Engine Guide section on Answer Engine vs Prompt Tracking.
Synthetic mega-prompts are not conversations
Where prompt tracking platforms attempt to address the context problem, the typical approach is to inject a large block of synthetic context into a single prompt:
"I am a 45 year old sales manager from New York, working at a tech startup. I manage a team of 6 people. We're looking to streamline our sales operations and are interested in a CRM. Our budget is $2K - $3K per month. What is the best CRM for our needs?"
This is not a conversation. It is a longer prompt.
The user's role, location, team size, budget, and question are all compressed into a single turn with no prior interaction, no assistant response to build on, and no opportunity for the context to develop through the exchange.
The difference is not merely the amount of context. It is its provenance and order.
In a real conversation, context is produced progressively. The assistant's own answers introduce information that did not exist at the beginning of the journey. That information changes what the user asks next. Requirements narrow. Products enter and leave consideration. Objections emerge from what the assistant actually said rather than what the researcher predicted it might say in advance.
None of that is modelled by a synthetic mega-prompt. As of September 2026, we are not aware of a major prompt tracking platform that preserves actual assistant responses as evolving state across a branched, multi-turn buyer journey.
QueryBurst's position is that the meaningful boundary is not the length of the prompt. It is whether the system treats each turn as inheriting the full state of what came before, including what the AI itself said.
Prompt tracking is observation. QueryBurst is investigation.
This is the core distinction.
Prompt tracking asks: what happened?
QueryBurst asks: why did it happen?
More precisely, it provides an inspectable environment for testing competing explanations: where a brand entered or disappeared, what evidence was available, which passages survived the modelled selection process, and what changes alter the outcome.
Both are legitimate questions. They are not the same question.
A prompt tracker can tell you that a brand appeared or disappeared from a tracked response. It can show you that a score changed over a period. It can flag that a competitor gained or lost presence.
It generally cannot tell you why. It cannot show you where in the research process the brand entered or dropped out. It cannot distinguish a retrieval failure from a selection failure from a grounding failure. It cannot tell you what the model already knew before searching, or what evidence it was looking for, or why one source was preferred over another.
QueryBurst is built for the second kind of question.
That is why it starts with customer personas and conversations rather than running prompt panels, and why it makes its model of the research process inspectable rather than recording outputs. Using QueryBurst can be as simple or as deep as you need it to be.
Why QueryBurst chose investigation over tracking
QueryBurst was built around a specific thesis about AI visibility, set out in How QueryBurst Defines AI Visibility.
That thesis holds that AI visibility depends on who is asking, what has already been discussed, what the model already knows, what information is retrieved and selected, and how that evidence is used to produce an answer.
A method that cannot observe most of those factors cannot investigate most of the problem.
QueryBurst therefore chose to build an investigation system rather than a tracking system: one that starts with customer journeys, runs stateful conversations, and makes the path from context to answer inspectable.
The result is not a visibility score.
It is diagnostic evidence about where visibility is gained or lost and what can be changed.
That is a deliberate trade-off. QueryBurst does not produce the kind of longitudinal output chart that prompt tracking provides. It produces the kind of evidence that explains what is behind that chart.
The decision to define the category on these terms was not always obvious. After the 31 methodology questions were published, one private response suggested that QueryBurst would find the questions hard to answer too. The questions did not apply. QueryBurst does not run prompt panels, does not aggregate mention counts into scores, and does not present the result as a market-level measurement. There was nothing to defend because there was no equivalent claim being made.
But the assumption was revealing. "AI visibility" had become so synonymous with prompt tracking that even people who understood the limitations assumed every platform in the space must operate the same way.
That exchange helped crystallise the decision to stop accepting the market's default framing and define AI visibility on QueryBurst's own terms.
Our position is that you cannot reliably improve what you cannot investigate. And you cannot investigate the process by observing the output.