A
BLIND / NATURAL
Can the brand enter naturally?
The target brand is not supplied. The shopper asks from the need itself.
OBSERVE
Whether the brand enters the answer or recommendation set, and which alternatives appear.
METHOD
A method for testingwhat AI actually recommends.
PersonaSignal starts with structured shopper contexts, then changes selected test conditions—brand context, product information, model, repeated run, or Search / Grounding—to compare observable recommendation behavior.
OBSERVABLE BOUNDARYWe study answers and recommendation behavior we can observe. We do not claim access to hidden reasoning, model weights, or proprietary ranking systems.
01
RESEARCH OBJECT
Knowing a brand exists
is not the same as recommending it.
A model may describe a named brand when asked directly. That does not tell us whether the same brand will appear when a shopper asks for a solution without naming it.
OBSERVED STATE ABRAND RECOGNITION
≠
AI can identify the supplied brand, product, or website information.
OBSERVED STATE BRECOMMENDATION
AI selects a brand or product for a defined shopper need.
SHARED STATUS LANGUAGE
These labels describe observable outcomes, not permanent brand scores.
- Recommended
- The brand or product is explicitly presented as a recommended option.
- Mentioned
- The answer names the brand but does not select it as a clear recommendation.
- Not Selected
- Another option is selected, or the target is not chosen within the defined result set.
- Unknown / Insufficient Evidence
- The answer does not support a reliable classification under the recorded rules.
A status records what happened in an answer. It does not automatically explain why it happened.
02
PERSONA FIRST
Define who's asking.
Then observe what AI recommends.
We do not begin with a loose audience keyword. We build a structured shopper context that connects the person, task, full need, screening criteria, and relevant product information.
- 01
PERSON
Who is deciding?
Life stage, experience, household, or other details relevant to the purchase. - 02
CONTEXT / TASK
What are they trying to do?
Market, climate, use environment, timing, and the practical task. - 03
FULL NEED
What must the choice deliver?
Goal, use, budget, preferences, must-haves, and exclusions. - 04
SCREENING CRITERIA
What shapes the shortlist?
Explicit conditions used to compare or rule out options. - 05
PRODUCT MATCHING
What fits those criteria?
How public product information aligns with the stated need. - 06
AI RECOMMENDATION
What does the model select?
The recommendation, mention, non-selection, or alternative observed.
TEST-DESIGN FRAMEWORK
This is an external test-design framework—not a claim about the model's hidden reasoning process. Structured scenarios help us compare answers; they do not represent every real shopper.
03
THREE TEST CONDITIONS
Three conditions.
Three different research questions.
The conditions are related, but they do not measure the same thing. Brand Understanding is not automatically “better performance” than a Blind / Natural result.
B
BRAND UNDERSTANDING
How is the brand understood?
Brand information is supplied so the model can respond to the stated positioning and offer.
OBSERVE
Positioning, fit, constraints, uncertainty, and information gaps in the description.
C
PRODUCT FIT
Where does the product fit?
Relevant product context is supplied, then compared across defined shopper needs.
OBSERVE
Which situations fit, which do not, and which alternatives may be selected.
04
CONTROLLED COMPARISON
Hold the shopper stable.
Change selected information.
Controlled comparison helps isolate observable differences between test conditions. It does not prove why a model changed its answer.
HELD CONSTANT
SAME PERSONA+SAME PURCHASE NEED
NO BRAND PROMPT
Blind / Natural
Observe natural brand entry, recommendations, and alternatives.BRAND INFORMATION
Brand Understanding
Observe positioning, fit, gaps, constraints, and uncertainty.PRODUCT CONTEXT
Product Fit
Observe fit judgments, recommendation states, and alternatives.COMPARE OBSERVED OUTPUTS
Compare brand entry, description, recommendation status, and alternatives to identify differences worth testing further.
05
REPEATED RUNS
One answer
can be noisy.
We repeat recorded test conditions to see which patterns persist and which results move around.
SAME TEST CONDITION
PERSONA P-07 · BLIND / NATURAL · SAME MODEL- RUN 01Recommended
Target Brand is explicitly recommended.
ALTERNATIVE · Brand A - RUN 02Mentioned
Target Brand appears but is not clearly selected.
ALTERNATIVE · Brand B - RUN 03Recommended
Target Brand enters the recommendation again.
ALTERNATIVE · Brand A - RUN 04Not Selected
Another brand is selected instead.
ALTERNATIVE · Brand C
OBSERVE
- States that recur
- Variation between runs
- Alternative brands
- Answers that remain unclear
INTERPRETATION BOUNDARY
Repeated runs support robustness and pattern comparison. They do not, on their own, establish statistical significance or represent the wider market.
06
MODEL COMPARISON
The same question
can surface different answers.
With the shopper, need, and test mode recorded, we can compare observable outputs across GPT, Gemini, or other selected models.
HELD CONSTANT
SAME PERSONA+SAME NEED+SAME TEST MODE
GPT
ILLUSTRATIVE OUTPUT
- RESULT
- Recommended
- MAIN ALTERNATIVE
- Brand A
- DESCRIPTION NOTE
- The answer emphasizes ingredient fit and screening criteria.
Gemini
ILLUSTRATIVE OUTPUT
- RESULT
- Mentioned
- MAIN ALTERNATIVE
- Brand B
- DESCRIPTION NOTE
- The answer emphasizes use context and public product information.
COMPAREObserve model-specific differences in recommendation state, alternatives, and description. This is not a provider ranking.
The research question and agreed Scope determine whether one model or a model comparison is useful. Differences are not reduced to an “intelligence level.”
07
SEARCH / GROUNDING
When search is available,
sources become observable too.
Some projects examine what changes when a model can retrieve public information. We can record whether a source or brand appears and compare grounded with non-grounded output.
SAME PERSONA + SAME NEED
ILLUSTRATIVE COMPARISONNON-GROUNDED
Without public retrieval
- RECOMMENDATION
- Record recommendation state
- MENTION
- Record whether the brand appears
- ALTERNATIVE
- Record the selected alternative
- SOURCE OBSERVATION
- Not available in this condition
SEARCH / GROUNDED
With public retrieval
- RECOMMENDATION
- Record recommendation state
- MENTION
- Record whether the brand appears
- ALTERNATIVE
- Record the selected alternative
- SOURCE OBSERVATION
- Record public sources that appear
SOURCE TYPES OBSERVED
- 01Official
- 02Retail
- 03Editorial
- 04Community
COMPARE OBSERVED OUTPUTS
Compare recommendation, mention, alternatives, and source observations across recorded conditions.
SOURCE BOUNDARY
A source appearing in an answer does not prove that it caused the recommendation. The research does not claim access to hidden source weights.
08
HUMAN REVIEW & TRACEABILITY
Automation can scale runs.
Human review makes the record defensible.
Each finding is mapped back to the shopper context, prompt condition, model, run record, recommendation status, and available source observation.
- 01
RAW RUN
Model answer
- 02
STRUCTURED RECORD
Recorded fields
- 03
CLASSIFICATION
Status label
- 04
SOURCE / CONTEXT CHECK
Evidence check
- 05
HUMAN REVIEW
Reviewed record
- 06
REPORT FINDING
Written finding
TRACEABLE STRUCTURED RECORD
Every finding has a route back to the test.
These illustrative fields use the same logic as the Sample Report's Raw Data structure.
RUN IDRUN-024
PERSONA IDP-07
TEST MODEBLIND / NATURAL
MODELGPT
GROUNDING STATENON-GROUNDED
RECOMMENDATION RESULTMENTIONED
MAIN ALTERNATIVEBRAND A
SOURCE RECORDNOT APPLICABLE
REVIEW STATUSREVIEWED
HUMAN-REVIEW BOUNDARY
Human review checks classification, context, and evidence records. It does not turn judgment into model certainty or decide what AI should answer.
09
EVIDENCE BOUNDARIES
What the research supports.
What it does not prove.
A useful research record is explicit about its boundaries. Findings are tied to the tested model, version, time, shopper context, prompt, and condition.
SUPPORTED BY THIS RESEARCH
Reviewable observations
- Recommendation behavior
- Whether the target was Recommended, Mentioned, Not Selected, or could not be classified.
- Recurring patterns
- Which outcomes persist or move across repeated recorded runs.
- Competitor substitution
- Which alternatives appear when the target is not selected.
- Model differences
- How observable outputs differ across selected models and conditions.
- Source observation
- Which public sources appear when Search / Grounding is available.
- Questions worth reviewing
- Public information that appears missing, ambiguous, inconsistent, or worth testing.
NOT AUTOMATICALLY PROVEN
Claims outside the evidence
- Market representativeness
- Structured shopper scenarios are not a statistically representative consumer sample by default.
- Hidden reasoning
- Observable outputs do not reveal internal chain of thought, weights, or ranking logic.
- Causal attribution
- Controlled differences do not prove why an output changed.
- Source effect
- A source appearing does not prove that it caused a recommendation.
- Future behavior
- No test guarantees how a model will recommend later.
- Business outcomes
- No result guarantees ranking, traffic, recommendation, sales, or an optimization effect.
METHOD SUMMARY
Not one AI answer.
A record you can compare and review.
PersonaSignal turns recommendation questions into structured evidence through shopper contexts, controlled conditions, repeated runs, model comparison, source observation, and human review.
The record remains bounded by the Persona, model, test condition, time, and public information available.
This page will open in a later stage.