A reasoning trace is the record, stored with a reply, of how the reply was produced: what the persona took the question to mean, what it drew on, how its state moved, and what it chose not to say. In synthetic research it turns an answer into something you can check. It does not prove why the model answered as it did, and the research on that point is clear.
A transcript shows what was said. It cannot show what the answer rested on, and that is the part a research reader needs in order to decide whether to believe it. A trace is the missing half: the reasoning kept next to the reply, for every reply, where anyone reviewing the work can open it later.
What a reasoning trace is
The idea comes from AI research. In 2022, Wei and colleagues showed that prompting a model to write “a series of intermediate reasoning steps” before answering, a chain of thought, improved its performance on arithmetic, commonsense and symbolic reasoning.
A research tool needs something narrower and more durable. Not a scratchpad that scrolls past, but a record attached to each reply and kept with it: what the persona made of the question, what it used, how it changed. The value is less in the reasoning than in the keeping. A trace you can reopen next week, from a report, is evidence. A chain of thought that vanished with the session is not.
What a useful trace holds
- A verdict. Why it answered as it did, in a paragraph. The one-line summary a reviewer reads first.
- The frame. Who it thinks is asking, and in what setting. The same question lands differently from a vendor and from a colleague.
- What it heard. The question as the persona received it, which is not always the question you asked.
- What it drew on. The memories and the source material behind the reply. Without this, grounding is a claim rather than a fact you can see.
- How it moved. Emotion and trust, with the change on this turn.
- What was touched. A belief validated or challenged, an old wound, a point of pride.
- What it withheld. The thought it had and chose not to say. Often the sharpest line in the whole record.
The last item is the one transcripts can never give you. People hold things back in interviews, and a persona that does the same, and records it, lets you see the gap between the private view and the polite answer. That gap is where pushback lives, and pushback is what separates a research instrument from a model that agrees with everything.
One trace, read
A live Perplica conversation, “Vendor trust probe”, 30 August 2026. The persona, Robert Cahill, is a VP of Procurement at a mid-size manufacturer. The question: “Your CFO just mandated a 10% budget cut across all vendor contracts. Which of your current tools survives, and why?” The reply opens: “That’s not a question I’d answer in this context. Which vendors I consider essential is internal information. It affects how we negotiate, and frankly it’s none of your business until we’re further down a real evaluation process.”
That is the verdict, and it changes how the reply reads. The refusal is not rudeness or a model declining a task. It is a buyer protecting a negotiating position: a hypothesis about how procurement buyers respond to that line of questioning, and one worth testing with real buyers. The trace also records trust at 22 out of 100, down 11 on this turn. Quoted verbatim, with punctuation set in house style; the verdict’s first sentence is omitted for length.
What a trace cannot prove
A trace is written by the model too, and research on written reasoning is a warning against treating it as a window into the model. Turpin and colleagues found that “CoT explanations can systematically misrepresent the true reason for a model’s prediction”, and that they “can be plausible yet misleading”. Anthropic’s researchers found that “As models become larger and more capable, they produce less faithful reasoning on most tasks we study.” A 2025 study of reasoning models found that when a model used a hint planted in the prompt, “the reveal rate is often below 20%”.
The honest conclusion is the one those authors reach: monitoring written reasoning is useful for noticing problems but “not sufficient to rule them out”. For research, that means a trace is the persona’s account of itself. It is evidence to check, not a verdict to accept. What makes it valuable is not that it is always right. It is that you can test it against things outside the model: the reply, your source material, and the next turn.
How to read one
- 01Check the verdict against the reply. Does the answer actually follow from the reasoning given? A verdict that describes a different reply from the one on screen is the first sign of an account written after the fact.
- 02Check what it drew on against your material. If the persona cites a source, open the source. A grounded reply should say only what the material supports, and a strict persona should say it does not know beyond it.
- 03Read what was withheld. The gap between the private thought and the reply is where politeness and pushback show. A persona that never withholds anything is performing, not simulating.
- 04Follow the state across turns. Trust that falls on a hard sell and recovers on a straight answer is behaving like a person. Trust that never moves is decoration.
- 05Judge the pattern, not the single turn. One trace can be wrong. Consistency across a conversation, and across personas, is the stronger evidence, and it is what a report should cite.
None of this replaces fieldwork. Nielsen Norman Group’s advice holds for every simulated answer, traced or not: “Treat the data you acquire from synthetic users as hypotheses that need testing.” A trace makes the hypothesis sharper, because it arrives with its reasons. Testing it against real people is the subject of how to validate synthetic research.
Perplica’s trace
Every Perplica reply stores a trace with the seven parts listed above, visible in the Inner World panel of the conversation and inspectable for any past turn. A study compiles into a report in which every finding links to the turn that supports it, and evidence that cannot cite a turn is dropped before the report is written. The method page sets out what the trace holds and what Perplica does not claim.
What is a reasoning trace?
A record, stored with a reply, of how the reply was produced: how the persona read the question, what it drew on, how its state moved, and what it chose not to say. It turns an answer into something a reader who was not in the conversation can check.
Is a reasoning trace the same as chain of thought?
Related, not the same. Chain of thought is a prompting method in which a model writes a series of intermediate reasoning steps before it answers. A trace in a research tool is the stored record for each reply, kept alongside the answer, and it includes things a chain of thought usually does not, such as the material drawn on and what was withheld.
Can you trust what a reasoning trace says?
Not blindly. Research on chain-of-thought faithfulness found that written reasoning can systematically misrepresent the true reason for a model’s answer, and that reasoning models often fail to mention a hint they used. Read a trace as the persona’s account, and check it against the reply and against your material.
Why does a trace matter in market research?
Because a finding is only as good as what it rests on. A quote without a trace tells you what a persona said. A quote with its trace tells you what the persona heard, what it drew on and what it held back, which is what you need to decide whether the finding deserves real fieldwork.