What Answer Engine Inspects
QueryBurst Answer Engine exposes the research process behind each AI-generated answer. Rather than starting with the final response, it makes each stage inspectable: the model's pre-retrieval position, the fan-out queries it generates, what evidence those queries are looking for, which sources are retrieved, which survive selection, and how the final answer is grounded.
What does Answer Engine show that a prompt tracker usually cannot?
A prompt tracker primarily starts with the visible answer.
Answer Engine can expose stages before it, including:
- the model's pre-retrieval position
- the fan-outs generated from that position
- what each fan-out is trying to find
- candidate sources
- candidate passages
- passage scores
- evidence that survives filtering
- evidence that gets discarded
- grounding
- citations
- the final response
So instead of only asking:
you can ask:
What is the pre-retrieval answer?
For many search-oriented questions, Answer Engine models an initial answer-shaped position before live retrieval.
This is the model's starting hypothesis.
It may contain:
- brands
- claims
- comparisons
- expected answer structure
- facts or attributes it believes need verification
It is not necessarily identical to the final answer.
That is the point.
Retrieval can verify it, refine it, or challenge it.
Why does the pre-retrieval answer matter?
Because it lets you distinguish very different problems.
For example:
Your brand was absent before retrieval began.
That suggests one problem.
Your brand entered during retrieval but lost during selection.
That suggests another.
Your evidence survived selection but never influenced the recommendation.
That is another problem again.
A final citation count collapses these into one outcome.
Answer Engine keeps them separate.
What are fan-outs?
Fan-outs are research queries generated to investigate the information needs behind the current answer.
A CRM recommendation might create separate research tasks for:
- pricing
- onboarding
- comparisons
- reviews
- cancellation policies
- security
- official product documentation
The interesting question is not simply:
It is:
"Why did that search exist?"
Answer Engine keeps the relationship between the model's initial position and the research it generated.
What is an evidence target?
For semantic retrieval, Answer Engine can create an answer-shaped representation of the evidence a fan-out is trying to find.
For example, a pricing research task should ideally retrieve passages containing plans, figures, billing terms, limitations, and similar pricing information.
The candidate passages can then be scored against that information need.
This is one plausible retrieval approach, not a claim that ChatGPT uses one universal scoring method for every query.
What models does Answer Engine use?
Answer Engine currently uses GPT-family models to model ChatGPT's behaviour.
That is deliberate.
The aim is to model current ChatGPT behaviour, so using an unrelated local or open-source model would introduce another major variable.
Other parts of the wider QueryBurst workspace may use different model providers where that distinction does not matter.
What does Answer Engine use for web search?
Answer Engine currently uses Brave Search for live retrieval.
The exact search provider is not the central claim of the system.
AI products can use different retrieval sources, change providers, combine indexes, or use multiple retrieval methods.
The more important questions are:
- why a research need existed
- what kind of evidence was sought
- what was retrieved
- what survived selection
- what influenced the answer
Will Answer Engine return exactly the same citations as ChatGPT?
Not necessarily.
Different retrieval providers, indexes, timing, caches, and product infrastructure can produce different candidate sources.
Answer Engine is not claiming to be a byte-for-byte replica of ChatGPT.
The useful part is making the research process observable.
The citations are the endpoint, not the entire system.
Does Answer Engine reproduce ChatGPT exactly?
No.
ChatGPT is a proprietary stateful product containing models, system instructions, memory, retrieval, tools, orchestration, ranking systems, product logic, and other components that can change.
Answer Engine models the mechanics that matter for investigating AI search.
It is a model.
A scraped ChatGPT output is an observation.
Neither should be confused with a universal ground truth. How we research and build these models is documented separately.
Why simulate ChatGPT instead of simply scraping the real UI?
Because the final UI response does not expose most of what produced it.
A scraped answer can show that a particular answer happened.
It generally cannot explain:
- why those particular fan-outs existed
- what the model thought before retrieval
- what evidence it expected to find
- which sources competed
- which passages lost
- where your brand entered or exited the process
If you want to make those stages observable, you need to model the process forwards.
Isn't a scraped ChatGPT result more "real"?
It is real in the sense that it happened.
That does not make it representative.
The answer belongs to a particular combination of:
- user
- account
- memories
- conversation history
- location
- product state
- model routing
- retrieval state
- time
- stochastic generation
There is no single canonical "real ChatGPT answer" for a prompt.