Product Case Study · April 2026

phia Smart Match

Making novel fashion images searchable, matchable, and trustworthy inside phia Lens.

Karishma Nageshwaran

02 · Context

Phia is exploding — and visual search is the pressure point.

Phia's stats

1M+

users under 1 year

6,200+

retail partners

Visual search is the new default

~20B/mo

Google Lens visual searches
4B are shopping-related ^[1]

80%

of Pinners start shopping with visual search ^[2]

"People use Google Lens for nearly 20 billion visual searches every month." — Google, 2024

Target user

Maya, 26 — urban, fashion-conscious, price-sensitive. Screenshots celebrity fits on Instagram.

Job to be done

"When I see something on social, find it — or close — in my budget, and show me why it matched."

★ North Star Metric

Daily active users completing a tap-through to a retailer product page.

Why this metric: it's the single signal that captures the full value loop — user intent, AI trust, and retailer commercial outcome. One click proves all three at once.

References: [1] Google, "Google Search updates" — blog.google, Oct 2024 · [2] Orr Consulting, citing Pinterest user research

02 / 13

03 · Problem

Phia's visual matcher fails on most novel images uploads

Phia · Green Sequin Skirt

Olive green sequin mini skirt uploaded. Returns black and brown sequin skirts as top results, then green suede and denim skirts — color & key attribute match failed.

Phia · Kendall's Pap Shot

Classic Pinterest inspo image.Results are Blue floral wallpaper and wide-leg pants — the matcher ranked on a single attribute with no consistency.

Phia · White halter

Returns tops with a close neckline match but wrong product style and attributes. No reasoning, no visibility into what the AI detected.

Repeatable across several uploads of different images. It's a consistent failure pattern and every fallback leaks a session to Google Lens.

03 / 13

04 · Solution

Smart Match — a VLM-powered fine-grained attribute layer that makes Phia Lens work on any image.

Image uploaded

Smart Match detects all products and accessories in the frame of any image uploaded.

Attributes extracted

VLM extracts fine-grained attributes per detected item into structured JSON — garment, color, silhouette, fabric, design details, and accessories.

Attribute-aware query

Retrieval scoped per product type queried separately for cleaner, categorized results.

Scored & interleaved

Weighted scoring, 6:4 New:Used render. Best matches first, per product type.

Demand signals surfaced

↗ Trending ⊘ Low stock ⏱ Selling fast — lifts purchase rate.

"Matched on:" trust signal

Per-card reasoning pill with AI context. Exact match surfaced at best price across when confirmed product is found.

04 / 13

05 · Diagnosis

Not a rebuild. An intelligence layer upstream.

The current system works when the image is already in Phia's catalog. It breaks on screenshots, paparazzi shots, and TikTok grabs — the fastest-growing image upload types.

Before · Phia

User uploads novel image

↓

Embedding lookup fails

↓

Broad category fallback

↓

Attribute-mismatched results

↓

User leaves for Google Lens

After · Smart Match

User uploads novel image

↓

VLM extracts structured attributes
garment · color · silhouette · fabric

↓

Attribute-aware query + fallback

↓

Weighted scoring · 6:4 New:Used render

↓

Trusted results with "Matched on:" reasoning ✓

05 / 13

06 · Proof

Same upload, same index. Different answer.

Google Lens

$Google Lens result$

Detects the product accurately with strong attribute context. No resale inventory, no demand signals, no price comparison.

Phia

Returns results on a 1:1 match. No trust signal, no demand context, no match visibility — reduces intent to purchase.

Smart Match

Demand signals, save counts, and urgency pills surfaced. Fine-grained attributes detected from all products per type, with exact match at best price.

06 / 13

07 · Competitive & Market Benchmarking

Nobody combines fashion attribute extraction with resale aggregation.

Google Lens

Reverse image search on the open web.

S Nails catalog-indexed ASOS / Zara items.

W No fashion attributes, no resale, no sustainability lens.

Shop with Google

Retailer catalog depth, Google-powered.

S Very deep new-retail catalog with real-time price & inventory.

W Shallow fashion attribute extraction. No resale.

Pinterest Lens

Visual discovery for style & vibe.

S Style-native discovery; strong Gen Z audience.

W Shoppable pins exist, but match quality is poor.

ShopSavvy / Honey

Price comparison and coupons.

S Strong price-compare mechanics + coupon stacking.

W Requires a SKU. Not visual-discovery-native at all.

The Wedge

Phia · Smart Match

Attributes + resale aggregation, on ANY image.

S Fashion attribute extraction upstream of retrieval.

S 150+ resale platforms already aggregated.

S Shows its reasoning per card — trust built in.

The category gap is attribute-precision × resale-aggregation. Phia is the only player positioned to own it.

07 / 13

08 · Prioritization

The smallest test that validates the thesis.

Reach / Month

600,000

~60% of 1M monthly users engage visual search. Prioritized as the entry point with highest leverage on the thesis.

Impact

3 · High

Fallback on novel images is a trust-breaker on Phia's primary entry point. Chose this because it targets fundamental accuracy — not a UX refinement.

Confidence

80%

Rapid-prototyped with Replit Agent 4. Validated across Claude Sonnet 4.6, GPT-4o, and Gemini on 20+ uploads using SerpAPI Google Shopping for real product retrieval. Scored results showed real tier variance — the mechanism holds up before scale.

Effort

2 · Medium

Scoped to contained Python scoring layered upstream of retrieval. No embedding-index rebuild — why I chose to ship additive, not replacive.

720K

Rice Score

(600,000 × 3 × 0.80) / 2

MVP · Ships to 10% Traffic

VLM based fine-grained attribute extraction on upload path; model selected via benchmark on fashion-specific accuracy, latency, and unit economics.
Attribute-aware query composition over existing retrieval — reuses what works, adds what's missing.
Weighted scoring — Strong ≥ 75 / Similar ≥ 50, no floor. Thresholds chosen to reward precision over recall on novel images.
6:4 Used:New render with ratio-warning logging. Interleave chosen to balance resale depth with freshness.
"Matched on:" line per card. Included because trust requires visible reasoning, not just a match.
Structured diagnostic logging (scores, pool, fallback). Chosen to make the first 30 days observable, not intuited.

08 / 13

09 · Demand Signals

Attribute match finds the piece. Demand signals convert the session.

Exact Match — AI-described + price-verified

When attributes resolve to a confirmed product, Smart Match surfaces the best price across 6,200+ partners — visual search becomes price comparison.

Strong match — confidence pill

Tells the user why this card matched — exact color, garment, silhouette. Trust before price.

↗ Trending ⊘ Low stock ⏱ Selling fast

— inventory pressure

Three demand signals lift purchase rate on highest-intent sessions.

🔖 Saves — social proof signal

High save count signals desirability before price registers — the strongest pre-purchase intent indicator on the card.

Why this matters for Phia

Phia is AI commerce for high-intent shoppers. Retrieval finds the item. Demand signals sell it without rebuilding the catalog.

09 / 13

10 · Success Looks Like

User Outcome

< 2s

User uploads an image photo/screenshot. Sees a resale match in under 2 seconds.

INTENT SIGNALS

↑ Tap-through on result cards
↑ Return visit within 48h
↓ Session abandonment

Business Outcome · north star

−40%

Fallback-to-Google-Lens on novel-image sessions. The one number I'd monitor weekly.

CONVERSION SIGNALS

↑ Retailer tap-through
↑ Purchase intent (add-to-cart proxy)
↑ GMV per visual-search session

Model Outcome

1 → N

VLM extraction becomes reusable infrastructure across 3 downstream features.

INHERITS THIS LAYER

→ Celebrity detection (v1.1)
→ Multi-item outfit tabs (v2)
→ Partner API (Q4)

Risks & mitigations

TRUST

Wrong matches erode trust faster than right matches build it.

Guardrail — luxury price-plausibility rule, false-positive counter held below 3%.

LATENCY

VLM latency breaks the sub-2s UX budget.

Mitigation — cache by image hash, parallelize with catalog pre-fetch.

10 / 10