Attribute Sentiment Scoring: Fit, Fabric, Scent, Value (August 2026)

Aug 19, 2026 by Merciv Team

Most review analysis tells you how a customer felt about a product. That's rarely the question your team actually needs answered. The question is which attribute is dragging repeat purchase and which is carrying it, and a single sentiment score can't tell you that. Scoring fit, fabric, value, and scent separately, the way aspect-level sentiment analysis works, is what turns a flat 78% positive SKU read into four separate signals you can route to the right team.

TLDR:

• A single star rating hides conflicting attribute verdicts; ABSA scores fit, fabric, scent, and value separately so each team gets a routable signal.
• The real finding in aspect-level analysis is the gap between two attribute scores on the same SKU, not any individual score in isolation.
• In Merciv's analysis, attribute complaints in reviews cluster three to six weeks before aggregate star ratings slip, giving you a brief window to act before the retailer conversation turns defensive.
• Cross-retailer attribute divergence on the same SKU points to fulfillment, listing, or shopper mismatch, not a product design problem.
• Merciv runs aspect-level scoring inside SKU-level continuous trackers with vertical-specific taxonomies and dual-source alerting before any signal fires.

Why a Single Sentiment Score Buries the Finding

Consider a four-star review of a mid-tier moisturizer: "Cleared my dry patches in a week, texture is silky, packaging feels premium, but the scent is aggressively floral and I can't wear it around my partner." Averaged into one score, that review reads as mildly positive. The star rating hides four verdicts, three enthusiastic and one a purchase-killer.

A brand manager sees a healthy 78% positive SKU. The formulator sees nothing actionable. The fragrance team never gets the memo. When repeat purchase softens two quarters later, the diagnosis lives buried in the same paragraphs that praised efficacy.

The finding a consumer intelligence for brand teams exercise needs is rarely "how do consumers feel about this product." It's "which attribute is dragging repeat purchase, and which is carrying it." Answering that requires scoring attributes separately, at the sentence or clause where the opinion actually lives.

What Aspect-Level Sentiment Analysis Actually Does

Aspect-level sentiment analysis (ABSA) pulls every distinct product attribute out of a piece of consumer language and scores each one on its own. One review, four verdicts. One star rating becomes four separate signals a brand team can route. Getthematic's ABSA guide frames this well: the moment you stop scoring the whole document and start scoring each attribute, the analysis becomes a routing system instead of a summary.

Three outputs sit inside every ABSA pass:

• Aspect extraction: identifying which attributes the consumer mentioned (fit, fabric, value, scent, texture, packaging, delivery, efficacy). This tells a merchant whether shoppers are talking about the seam or the sizing.
• Opinion term linking: connecting descriptive words to the aspect they modify. "Silky" attaches to texture, "aggressively floral" attaches to scent.
• Polarity classification: scoring each aspect as positive, negative, or neutral, independently of the others.

The output is a per-attribute read on a single verbatim, which rolls up into per-attribute reads on a SKU, a launch window, or a retailer.

How It Differs from Document-Level and Sentence-Level Sentiment

Sentiment analysis operates at three grains, and the choice determines what a brand team can act on.

Document and sentence approaches collapse conflicting signals inside the same text. Only ABSA separates them, which is why aggregate star ratings can look flat while a specific attribute quietly drags repeat purchase. Research on ABSA applied to e-commerce reviews confirms the pattern: composite scores systematically obscure attribute-level divergence that matters for product decisions.

The Four Attribute Clusters That Matter by Vertical

The taxonomy is the analysis. Wrong aspects, wrong answers downstream. Below are the clusters that hold up across four priority verticals, tuned to how consumers actually talk about each category in reviews.

Apparel and footwear

For brands using consumer intelligence for fashion & apparel, the key attributes are:

• Fit (true to size, length, arch)
• Fabric quality (hand feel, weight, stretch recovery)
• Construction (seams, stitching, hardware)
• Wash durability (pilling, shrinkage, color fade after 3-5 washes)

Beauty and personal care

Merciv's consumer intelligence for beauty & wellness tracks these core attributes:

• Scent
• Texture (slip, absorption, residue)
• Efficacy against the claim
• Packaging (pump, cap, travel integrity)
• Skin compatibility (irritation, breakouts)

Food and beverage

Merciv's consumer intelligence for food & beverage tracks:

• Flavor
• Texture or mouthfeel
• Packaging (reseal, leak, freshness)
• Portion value
• Ingredient clarity (label, sourcing, "no seed oils")

Wellness

• Efficacy ("did it work for me")
• Taste or format (pill size, powder mixability)
• Ingredient transparency
• Price-to-result ratio

Define the aspect set with a merchant or formulator in the room before the first job runs.

When Attribute Scores Diverge: That Is the Finding

The strategic value of ABSA is not the individual attribute score. It is the gap between two attribute scores on the same SKU. That gap names the problem.

Three divergence patterns show up repeatedly in CPG and apparel review data:

• Texture drops, flavor holds. A snack reformulation lands, flavor sentiment stays at +40, and texture sentiment slides from +25 to -10 over six weeks. The fix sits with the process engineer, and the diagnosis lands before syndicated velocity confirms the softening.
• Efficacy positive, value negative. A serum scores strongly on "worked for my skin" verbatims while price-perception sentiment trends down across Sephora and Ulta. The response is a size ladder or a bundle, not a reformulation.
• Fit high at one retailer, low at another. As shown in the athletic apparel competitive review analysis, a denim SKU shows fit sentiment at +35 on the brand DTC site and -15 on Amazon. Sizing execution or listing accuracy at one channel is off, and the buyer conversation moves to which channel is misrepresenting the fit.

One attribute moved, the others held, and the intervention has a name.

Aspect Scores as an Early Warning System

Review volume climbs year over year, and attribute-level signal piles up faster than any manual coding pass can absorb. The early-warning value of ABSA lives here.

The reformulation pattern repeats across categories, covered in depth in hero SKU threat monitoring: texture, scent, or flavor complaints cluster in reviews typically three to six weeks before the aggregate star rating slips. Efficacy, packaging, and price sentiment hold steady, so the four-star average masks the break. One attribute is failing; the composite hides it.

For an insights leader making the internal case, the argument is temporal. Waiting for the star average to move means waiting for satisfied buyers on other attributes to be outnumbered by the ones the reformulation broke. By then, syndicated velocity has turned, and the retailer conversation is defensive. Attribute-level tracking catches the shift while it still fits on a one-page brief to the formulator.

Cross-Retailer Attribute Divergence Reveals Distribution Problems

Same SKU on Amazon and the brand DTC site should return the same attribute scores. When they don't, the divergence itself is the diagnostic.

Three patterns show up in beauty and apparel review data:

• Fabric sentiment negative on Amazon, neutral on DTC. Same construction, same fiber. The gap points to third-party fulfillment: warehouse handling, transit humidity, or unauthorized resellers moving older stock. Product design is the wrong owner.
• Packaging sentiment negative on Amazon, positive on Sephora. A pump that arrives broken through one channel and intact through another isolates a shipping or box-spec issue, not a supplier defect.
• Fit sentiment high at Nordstrom, low at an off-price channel. Same denim, different shopper. The score gap reflects segment mismatch, so the response is listing copy or channel strategy, not a pattern revision.

You have controlled the product variable. What changed is the channel, the handling, or the shopper, which points the investigation at the right function before a week gets burned redesigning something that ships fine everywhere else.

Configuring Aspect Categories: The Taxonomy Decision

Two teams running ABSA on the same 40,000 reviews will surface different findings if their taxonomies drift. The drift shows up in three places.

Granularity. Aspects need to be broad enough to catch related language but narrow enough to separate genuinely different dimensions. In skincare, "silky" attaches to texture, "runny" attaches to consistency, and a reformulation can move one while holding the other. Collapse them into "feel" and the finding disappears.

Scope. Fit on a denim cut belongs at SKU level. Return experience belongs at brand level. Running everything at SKU scope buries brand-wide packaging failures; running everything at brand scope hides the SKU where the pump broke. Most brands need both, tagged separately so a brand manager and a category lead get the right slice.

Get a merchant, an insights lead, and someone who has read a thousand reviews in the same room before the taxonomy locks, a principle covered in depth in the CPG consumer insights practitioner's guide. Consumer vocabulary rarely matches internal vocabulary, and the taxonomy has to speak the consumer's language to be extractable at scale.

Confidence and Traceability: What Makes an Attribute Score Defensible

Three properties separate a defensible ABSA output from a colorful chart:

• Per-label traceability. Every attribute score clicks through to the underlying verbatims, the source (Sephora, Amazon, Reddit), and the retrieval date. "Fabric sentiment is trending negative" opens into the sentences that produced it, not a black-box aggregate.
• Confidence tiering at the label level. High confidence requires multiple sources in agreement and recent data. Directional means the pattern is real but thin. Exploratory means one feed, one week, treat as a hypothesis.
• Volume gating. An attribute mentioned in 12 of 800 reviews is not a defensible read. Flag it as low-coverage until mentions clear a pre-set threshold.

The CMO test: show me where you got this, a pressure point covered in citing AI sources in a readout. A clickable path from score to verbatim to source with a retrieval date survives the room. "The model returned it" does not.

Where ABSA Breaks Down: Real Limitations for Practitioners

Three failure modes show up consistently in production ABSA runs. Naming them upfront prevents wrong decisions downstream.

• Sarcasm flips polarity. "Oh, the scent is amazing, if you enjoy smelling like a hotel lobby" reads positive to most classifiers. The score looks clean while the verbatim contradicts it. Sample positive-classified verbatims manually every run, especially on scent, flavor, and value where irony clusters.
• Implicit aspects are harder than explicit ones. "Broke out after one use" is a skin compatibility verdict with no aspect noun. Expect lower recall until the taxonomy has been tuned against thousands of category-specific verbatims.
• Low volume produces unstable scores. A SKU with 30 reviews and 6 fit mentions cannot support a defensible read. Gate SKU-level output at a mention count agreed on before the first job runs, and label anything below it exploratory.

How Merciv Applies Aspect-Level Scoring at the SKU Level

Merciv applies aspect-level scoring inside SKU-level continuous trackers, with attribute taxonomies defined separately for beauty, apparel, and food and beverage. No generic aspect set gets pushed across verticals, because the vocabulary that surfaces "seed oils" in F&B has nothing useful to say about denim wash durability.

Three properties govern how the output holds up in a category review:

• Per-label confidence tiers (High, Directional, Exploratory) are applied as defined in the Confidence and Traceability section above.
• Per-label audit trail. Each score clicks through to the verbatims, sources, and retrieval dates behind it, supporting the goal of turning insight into defensible decisions. When a brand manager tells a buyer "fit sentiment is deteriorating at Nordstrom," the claim opens into the sentences that produced it.
• Dual-source alerting: trackers require two independent sources before an alert fires. Confidence tier definitions are covered in the Confidence and Traceability section above.

Final Thoughts on Turning Review Data Into Actionable Attribute Signals

Attribute-level scoring does one thing well: it names the problem before the aggregate metric does. Your formulator, packaging lead, and channel team each need a different slice of the same review data, and a single document-level score gives all three of them nothing to act on. Build the taxonomy with the people who know what the product does and how consumers talk about it, gate the output at a defensible mention count, and the signal pays for itself. Merciv's enterprise work covers how per-label confidence tiers and dual-source alerting keep those reads from sending the formulator after phantom spikes.

FAQ

What is aspect-level sentiment analysis, and how does it differ from document-level sentiment analysis?

Aspect-level sentiment analysis (ABSA) scores each product attribute separately within a single review (fit, fabric, scent, efficacy) instead of collapsing the whole text into one score. Document-level analysis returns one mixed verdict on a four-sentence review that praises texture, scent, and packaging but flags irritation; aspect-level analysis returns four independent scores, so the one attribute dragging repeat purchase is visible before it moves the star average.

Should I use aspect-based sentiment analysis or standard sentiment analysis for SKU-level product reviews?

For SKU-level decisions (reformulation calls, retailer conversations, channel diagnostics), aspect-based sentiment analysis is the right choice because standard sentiment collapses conflicting signals inside the same text. A review that scores "fit: positive, fabric: negative" reads as mixed in document-level analysis and generates no actionable finding; ABSA separates the two verdicts so the formulator and the merchant each get the signal that belongs to them.

How do I build an aspect-based sentiment analysis taxonomy for CPG or apparel product reviews?

Start by defining your aspect set with a merchant or formulator in the room before the first job runs, because consumer vocabulary rarely matches internal vocabulary, and a taxonomy tuned to how consumers actually write ("broke out after one use," "fell apart in the wash") will catch the highest-signal verbatims. The taxonomy decisions that matter most are granularity ("texture" and "consistency" are different dimensions in skincare; collapse them and the finding disappears), implicit reference coverage (complaints that carry no explicit aspect noun, like "fell apart in the wash," are often the highest-signal verbatims), and scope (fit on a specific SKU belongs at SKU level; return experience belongs at brand level). Volume-gate SKU-level outputs at a minimum mention count agreed on before the first job runs, and label anything below that threshold as exploratory.

What are the common failure modes in aspect-based sentiment analysis on product reviews?

Three failure modes show up consistently in production ABSA runs. Sarcasm flips polarity: "the scent is amazing, if you enjoy smelling like a hotel lobby" reads positive to most classifiers, so manual sampling of positive-classified verbatims on scent, flavor, and value is a required quality check. Implicit aspects produce lower recall until the taxonomy has been tuned on thousands of category-specific verbatims, because complaints like "broke out after one use" carry no explicit aspect noun. Low-volume SKUs produce unstable scores, and a SKU with 30 reviews and 6 fit mentions cannot support a defensible read; treating that output as anything above exploratory invites wrong decisions downstream.

How can I use aspect-level sentiment scores as an early warning system before syndicated velocity data confirms a problem?

In Merciv's analysis, attribute complaints cluster three to six weeks before star ratings slip, while efficacy and packaging sentiment hold steady. For the full pattern and alerting setup, see the Aspect Scores as an Early Warning System section above.