The Method

Auditable customer intelligence.

Built for investment committees and boards.

Each read is traceable to the sentence that produced it.



Collect · Grade · Aggregate · Benchmark · Deliver
The process

5 stages.

1
Collect

Collect reviews from every major platform.

Unstructured review text, ratings, and metadata across review sites.

2
Grade

Grade each sentence 1–5 on each KPA.

Sentence-level grading captures context and nuance a single-review rating misses. +32% more accurate than legacy tripart (pos/neu/neg) sentiment.

3
Aggregate

Aggregate quantified reviews into group, regional and site reads.

Every sentence contributes to the aggregated view.

4
Benchmark

Benchmark vs competitors, substitutes or the system.

Compare performance to local and national competitors, substitutes or franchise systems.

5
Deliver

Delivered in a format that fits your workflow.

Dashboards, slide-ready exhibits, standalone reports, or structured data.

See it work

Collect. Grade. Aggregate.

This is what one customer review looks like as it moves through Sorrel, and how it feeds the aggregated read behind it.

01 · Grade
Each sentence is graded 1–5.
★ ★ ★ ★ ☆ Booking.com · Verified stay · 2 weeks ago
Stayed for three nights on a weekend break.Context Check-in was seamless, greeted by name with a signed note from the manager waiting in the room.Arrival experience · 5 The room was generously proportioned but the shower drainage was slow and the wardrobe door didn't close properly.Room size · 4 Staff throughout were attentive without being intrusive; the concierge went out of their way to secure a difficult reservation.Hotel service · 5 Given the rate, breakfast felt limited for what was charged, and the wine list at dinner was overpriced.F&B value · 3 Would return for the service, though I'd hesitate to recommend at rack rate.Intent · +
Arrival Room size Hotel service F&B value Intent
02 · Quantify
Each review is quantified.
Hotel service
4.7
Arrival experience
4.5
Room size
3.8
Hotel decor
3.2
F&B value
3.4
Value
3.6
Ambience
4.2
Showing 7 of 11 hotel KPAs. Neutral bars are contributed by other reviews.
This review's contribution to the aggregate. One of ~2M reviews across UK premium hotels over the trailing 24 months.
Every point on the dashboard clicks through to the sentence that produced it.
03 · Aggregate
Every review is aggregated into the group, regional and site read.
4.7
Hotel service · group read
Aggregated from 2,014,338 graded sentences across the estate, trailing 24 months.

The coral square is the review above. Every other square is another graded review contributing to the same number. No single voice moves the read, which is the point: the aggregate is what survives scrutiny, and every square in it remains individually traceable.

See what the framework uncovers.
Request a demo
The algorithm

Purpose-built.
Specialist-trained.

Legacy sentiment tools apply a generic positive / neutral / negative label to a whole review. Off-the-shelf LLMs guess without a framework. Neither is defensible in an investment committee or board where accuracy, consistency and transparency matter.

Sorrel's algorithm grades every sentence 1–5 against a proprietary, PhD-designed sector-aware language taxonomy, maintained by sector specialists and refreshed continuously as language and customer expectation evolve.

The result is a model that reads a review the way an experienced operator would, at scale, consistently, and with a full audit trail.

Tens of millions
Individual training data points annotated at sentence level, not review level.
PhD-designed
Sector-aware language taxonomy built by hospitality & consumer sector specialists, not a generic NLP schema.
Continuously refreshed
Taxonomy and model updates versioned engagement-to-engagement, so nothing drifts without an audit trail.
97%
Model agreement with ground truth. Sorrel’s algorithm is now more accurate than the human annotators it was trained by.
Why Sorrel

Built for capital.

How seven approaches to customer feedback compare
against what's required by investors and boards.

Approach Driver-level granularitySpecific quality, service and value drivers, not one score Like-for-like benchmarkingComparable vs sector peers, not internal trends only Sector-aware frameworkBuilt for the hospitality & consumer sectors ScaleMillions of signals, not hundreds DefensibilityStands up under IC scrutiny
SorrelBuilt for investors & boards
Positive / negative sentimentLegacy NLP
Generic LLMsChatGPT, Claude
Social listeningBrand mentions, built for marketing & PR
Review scraping servicesRaw data, built for engineering teams
Manual analyst reviewConsultancy teams
Customer surveysNPS panels

Scale = capability to process millions of customer signals*. Defensibility = capability to do it accurately, consistently and with rigour.

*Public reviews, ratings, and similar expressions of customer perception.

Output formats

The format follows the context.

Insights are delivered in the form that fits the situation.

Dashboards
Real-time portfolio oversight, asset management, diligence and research. Drivers, competitive position, momentum.
Slide-ready exhibits
Charts and tables formatted for IC packs, pitch books, information memoranda and board materials.
Standalone reports
Key findings, implications and commentary, ready to share with boards, ICs and external parties.
Structured data
Processed outputs delivered as API, CSV or warehouse feed for integration into your own analytics environment.

Tested and trusted by leading investment funds, advisory firms and operating groups.

Including
Your next question

Get in touch.

Schedule a call directly, or email us at sam@sorrel.ai.

Response within one working day. NDA-covered.