OMX innovation · 08 / 25 · Web Optimisation
Three phases.
One we closed. One we're building. One next.
Phase 1 Search (Lucene rebuild) closed May 2026 — AI visibility 12 to 80%, CSAT 8.0 to 9.0. Phase 2 Hierarchy + Filters (300,000 products classified at 93.1%) in build. Phase 3 PDP (5 archetypes, $1.9M/yr 1pp lift sized) is next.
DECK 08 · WEB OPTIMISATION · ANCHORED ON THE LIVE PROGRAMME
01 / 10
The programme

Three phases.
One Search mandate.

Phase 01
Lucene Search
Rebuild. AI visibility. CSAT lift.
Closed · May 2026
→
Phase 02
Hierarchy + Filters
442 L5 categories. 93.1% confidence.
Build + test
→
Mar 2026 said win Search. We didn't wait — Phase 1 shipped, Phase 2 is shipping, Phase 3 is queued. The numbers below are the real ones, not projections.
02 / 10
Phase 01 · Lucene Search · Closed 2026-05-24
What we did + the results

Lucene rebuilt. AI visibility 12 → 80%.

What we did
  • Lucene rebuild — search index reconstructed from canonical product data (32,675 products in scope)
  • V11 simulator — controlled environment to test queries + relevance changes before production
  • llms.txt + 11-bot allowlist — site discoverable by AI crawlers (Anthropic, OpenAI, Google AI, Perplexity, etc.)
  • Schema.org + structured markup — every PDP machine-readable
  • GA4 measurement — every change measured against baseline
What we got
12%→80%
AI visibility (citation in LLM answers)
8.0→9.0
CSAT (customer satisfaction)
Defects we fixed in flight
  • Circular query loops — redirect rules that bounced between two synonyms forever; broken chains untangled and capped
  • Zero-result queries — the long tail of searches returning nothing; root-caused (mostly missing attributes / mis-mapped synonyms) and resolved
  • Brand list integrity — full brand catalogue rebuilt + reconciled; brand facet now reliable, brand-name searches actually land
  • Boost / trim / redirect rules audited end-to-end; bad rules retired, good rules promoted, every change versioned
  • Query translation table — cleansed and certified; legacy edge-cases removed
The dataset trigger

Closing Lucene also triggered the ingestion of every search event to Snowflake — queries, results returned, click-through, abandonment, conversion. We now have a stewarded source-of-truth for search behaviour, modelled in dbt next to sales.

That unlocks the next-order analytics: search → sale attribution (which queries actually convert, by category, by customer tier) and search → churn signals (lost-search rate as a leading indicator of an account about to walk). The platform shift, not just the relevance lift, is what makes Phase 2 + 3 measurable.

Closure deliverable: A-Project-1-Search-CLOSED.pdf · sourced from web-optimisation/Current/50-search-and-seo/. The lift is banked. The data pipe is live.
03 / 10
Phase 02 · Hierarchy + Filters · Build + test
The data underneath

300,000 products classified. 93.1% confidence.
Every new product — onboarded automatically.

This is the plan we're building now — the classifier doesn't stop at the back-catalog; every new SKU lands enriched, classified, and PDP-ready on day one.
104
L1 categories
Top-level navigation
690
L2 departments
Sub-categories
442
L5 (Tech)
Fine-grained classification
300,000
Products classified
@ 93.1% confidence
Filter UX (Phase 2 design)
☑ Furniture (2,450)
☑ Chairs (892)
☐ Office Chairs (345)
☐ Stacking Chairs (187)
☐ Task Chairs (215)
☐ Desks (641)
☐ Storage (517)
Brand HP ×128 Brother ×64 Canon ×42
Price <$200 ×412 $200-$500 ×285 $500+ ×167
Stock In stock ×823 Same-day ×512
Sustainability FSC ×118 Recycled ×245 EnergyStar ×87
Filter requirements at 30-web-hierarchy-and-filters/01-FILTER-REQUIREMENTS.md. Brand Owner field design + Filter Automation feeds this layer.
04 / 10
The spine

Attribute Library DB.
1.07 GB. 9 sources. Built.

OMX backbone
16 L1 · 17 L2 · 68 attrs · 1,050 values
GS1 GPC (May 2026)
45 segments · 5,318 bricks · 2,085 attrs · 13,733 values
UNSPSC
36 segments · 18,347 codes
Amazon BTG
22 cats · 31,851 nodes · 4,449 refinements
ETIM 10.0
159 groups · 5,640 classes · 17,377 features · 201k links
Google Shopping
5,595 paths · 72 attrs · 61 values
Schema.org Product
72 properties
GS1 GDM v2.17
5 cats · 3,019 attrs
Icecat Open Catalogue
6,804 categories · 35,002 features · 8,879,106 values
Cross-walks generated
368 OMX → external candidates · 45 OMX-L1 → external segments
SQLite single-file (~1.07GB). Postgres + Snowflake DDLs ready for migration. Path: 30-web-hierarchy-and-filters/attribute-library-db/
05 / 10
The AI connector
Filter database × AI classification

Every new product. Onboarded automatically.

01
New SKU arrives
Supplier feed · Plex-CI scrape · quote-driven addition · PDX intake
→
02
AI classifier
Embeddings cache + prompt library. Reads against the Attribute Library DB. Confidence-scored output.
→
03
Filters auto-applied
L1 + L2 + L5 + brand + price band + sustainability tags + Icecat features — all populated from one inference.
→
04
Live on the site
PDP populated · filter rails respected · search index updated · ready for the customer.
93.1%
Classification confidence (Tech, 300,000 products)
442
L5 categories AI-trained against
9 sources
Attribute library feeds the classifier
~$0.05
Per SKU AI cost (Claude Sonnet)
Today: manual taxonomy + manual filter assignment + manual PDP. Tomorrow: AI classifier reads the Attribute Library DB and ships the SKU. Path: product-hierarchy/Current/20-tech-taxonomy/30-ai-classification/
06 / 10
Phase 03 · PDP · Investigation
The decision moment

5 archetypes. $1.9M/yr · 1pp lift.

A1
Tech · Computers + Peripherals
Spec-heavy; configurator pattern
Apple MacBook PDP reference
apple.com · MacBook
A2
Office Supplies · Consumables
Quick-add; reorder-pattern emphasis
Officeworks AU paper PDP reference
officeworks.com.au · paper
A3
Furniture · Bulky goods
Dimensions + delivery window pattern
Herman Miller Aeron PDP reference
hermanmiller.com · Aeron
A4
Safety + Compliance
Certification + compliance pattern
RS Components safety PDP reference
rs-online · safety
A5
Workwear + Apparel
Sizing + uniform-list pattern (BTS-27 link)
Carhartt overalls PDP reference
carhartt.com · overalls
$1.9M/yr
Sized opportunity
1pp conversion lift on current site revenue base. Conservative.
9 pre-reqs catalogued · sourced from C-Project-3-PDP-Investigation-and-Plan.pdf
Phase 3 needs Phase 2's data + classifier ready. Once shipped, every PDP gets the cues, the trust signals, the alternatives engine (Deck 20), and the Ask Max handoff (Deck 01).
07 / 10
The compound

Each phase stacks. The lift compounds.

Phase 01 · Done
Lucene
+1.0
CSAT 8.0 → 9.0 + AI visibility 12 → 80%
+
Phase 02 · Doing
Filters + Data
+10%
Filter automation projection
+
Phase 03 · Next
PDP
+1pp
$1.9M/yr · conversion lift
=
Compound
Programme
$1.9M+
Per-year revenue lift, conservative · plus Phase 1 AI/CSAT gains banked
Filter automation impacts (from 2026-05-22 design): +10% conversion · -15% bounce · +5% AOV · -25% support tickets. PDP adds on top. Real numbers, programme-document sourced.
08 / 10
Architecture

Same data. Three layers. One spine.

Phase 01 · Lucene
Search index · 32,675 SKUs
V11 simulator
llms.txt + 11-bot allow
Schema.org markup
GA4 measurement
→
Phase 02 · Filters + Data
Attribute Library DB
9 sources · 1.07 GB
Stibo Brand Owner field
AI classifier · 93.1%
Filter automation
→
Phase 03 · PDP
5 archetypes
Decision cues
Inline stock + delivery
Trust signals
Alt engine (Deck 20)
Ask Max handoff (Deck 01)
No new platform. Same Snowflake (data) + Lucene (search) + Stibo (MDM) + Lens design language + the existing site shell. AI classifier is the only new piece — and it runs on Claude.
09 / 10
The ask

Close Phase 2. Open Phase 3.
Bank what Lucene proved.

  • Phase 2 finish — Stibo Brand Owner field deployed · filter automation in production · AI classifier ingesting new SKUs daily
  • Phase 3 commit — 5 PDP archetypes designed; sized at $1.9M/yr 1pp lift; sequenced for H2 2026
  • 9 pre-reqs — catalogued in C-Project-3-PDP-Investigation-and-Plan.pdf; resourced + scheduled
  • Compound measurement — Lens dashboard tracking AI visibility · CSAT · conversion · AOV end-to-end
  • Connection to adjacent decks — Deck 16 PDX (data quality) · Deck 20 Product Matching (alternatives) · Deck 01 Ask Max (PDP handoff)
  • Stage 3 Switching — Deck 09 picks up an optimised funnel; this programme unlocks it
Audience: Chief Digital Officer · CMO · Commercial Director · CFO · Data Council
10 / 10