Each of 15,500 synthetic reviews was scored across 25 domains in 6 categories using the SSR framework.
| Category | Avg Score | Mentions |
|---|---|---|
| Concerns | 3.51 | 45,531 |
| Awareness | 3.60 | 15,801 |
| Procedures | 3.68 | 19,414 |
| Satisfaction | 3.93 | 9,580 |
| Loyalty | 3.68 | 5,890 |
99.7% of synthetic reviews trigger at least one domain, averaging 6.8 domains per review.
Sanity check: do synthetic scores show realistic spread and ranges across domain categories?
Box plots showing score variance within each category. Tighter distributions indicate consistent LLM output; wider distributions reflect diverse persona/concept interactions.
Which treatment concepts trigger which sentiments? The heatmap reveals concept-level score patterns across key domains.
Rows = treatment concepts, columns = sentiment domains. Green = positive sentiment, red = strong concern. Concepts with more red in concern columns create stronger patient barriers.
How do different patient types react to treatments? This heatmap shows persona-level score patterns.
Rows = patient personas, columns = sentiment domains. Variation across personas confirms that demographic conditioning produces differentiated reactions.
The key validation: do persona attributes produce expected score patterns?
| Group | Avg Score | n |
|---|---|---|
| High | 3.562 | 2319 |
| Medium | 3.668 | 3100 |
| Low | 3.968 | 1433 |
| Group | Avg Score | n |
|---|---|---|
| First-time | 2.875 | 595 |
| Occasional | 3.085 | 4 |
| Regular | 3.320 | 23 |
| Group | Avg Score | n |
|---|---|---|
| First-time | 3.194 | 3084 |
| Occasional | 3.306 | 189 |
| Regular | 3.749 | 714 |
Correct directional signals across all checks confirm that persona conditioning is working as designed.
Which synthetic sentiments travel together?
| Domain A | Domain B | r |
|---|---|---|
| Biostimulator | Laser/Energy | 0.93 |
| Expectations | Biostimulator | 0.87 |
| Dermal Filler | Laser/Energy | 0.87 |
| Domain A | Domain B | r |
|---|---|---|
| Needle Fear | Provider Referral | -0.24 |
| Needle Fear | Web Search | -0.07 |
| Needle Fear | Provider Loyalty | -0.01 |
Which domains are triggered together in the same synthetic review?
| Domain Pair | Co-mentions |
|---|---|
| Cost Concerns & Expectations | 4,944 |
| Expectations & Web Search | 4,851 |
| Expectations & Immediate Results | 4,797 |
| Dermal Filler & Expectations | 4,457 |
| Downtime Concerns & Expectations | 4,277 |
High co-occurrence rates reflect realistic review content where patients discuss multiple aspects of their experience simultaneously.
Correlation with real National data (124,705 reviews): r = 0.785
| Domain | Difference |
|---|---|
| Pain/Discomfort | -0.726 |
| Permanence Concerns | -0.636 |
| Expectations | -0.593 |
| Health Concerns | -0.510 |
| Provider Loyalty | -0.484 |
Rank-order correlation is the key metric — domains should rank in the same order, not match exactly. Domains where synthetic scores are higher than real represent silent concerns — things the real patient population feels but the review-writing population under-reports.
Violin plots comparing synthetic (blue) vs real (green) score distributions. Distribution shape similarity (measured by Overlap Coefficient) matters more than matching means.
Lower synthetic means are expected — this is the selection bias (real reviews come from satisfied self-selected customers). What matters is that relative domain ordering and distribution shapes are preserved. Operational domains (bedside_manner, operational_friction, provider_loyalty) are excluded from comparison metrics.
Average word count: 74 words per review.
Left: Word count distribution by treatment concept. Right: Average word count by persona. Variation shows LLM output diversity across different inputs.