HYPV

A pre-launch prototype of a predictive-AI powered hyperpersonalisation platform for enterprise online retailers — request demo

Materially different from existing solutions (Segmentify, Algolia, Optimizely etc.) as it can run as a native online store frontend itself — ingesting maximum signal telemetry and able to algorithmically vary absolutely any visual or data aspect of content rendered by cohort (microsegmentation) — not just adding some product discovery to an existing platform UX and running a few A/B tests

This step-change in targeting capability is achieved by maximising the label density assigned to every data or visual object and tracking how endusers interact with those labels — adapting the content rendered based on the weight of the labels and velocity of the interactions over time — whilst enabling unlimited A/B test variants of any object to coexist in parallel

General or RAG-enhanced image classification can be used to detect and assign labels not already in the product attribute set (such as :sneaker_high_top or :skirt_long for clothing) to further assist the targeting effectiveness

In addition to optimising traditional UX journeys — early indications are that this native hypervariation mechanism will be particularly effective when used in combination with emerging bi-directional conversational AI voice services to drive ecommerce interactions, especially on mobile — refining display based on voice input and able to learn new labels from enduser response comments — "no, those are all too :new_neg_label"

Summary

• Combines deep product discovery and unlimited A/B testing features within bare metal platform
• Uses novel labeling system where all components of a page render (data, visual fragments and code/algos) have a uniform labeling format with weights and interaction directions/velocities
• Designed primarily for £200M+/yr stores with complex product ranges and high repeat-visit traffic — unlikely useful outside of that
• Pushes 'hypervariation ⟶ microsegmentation' content targeting concepts to extremes — without using any third-party tech other than supporting integrations with external AI/ML services
• Can run as an augmentation API to upgrade an existing platform, as headless frontend over a traditional platform (Big Commerce, Salesforce Commerce Cloud, Shopify etc.) or as a standalone platform itself — eventually
• Renders HTML natively as MPA or can seamlessly integrate with any popular frontend/SPA framework like React
• Proprietary multidimensional data structure that, theoretically, unlocks unparalleled predictive modeling capabilities

Content Hypervariation & Targeting

• Automatically optimises/AB tests content based on realtime behaviour cues unless attenuated by manual override rules (such as forcing :high_ltv visitors to see specific content UX and merchandising at every stage in journey vs :visit_first_time) or delivering a richer experience when viewing sponsored brands listed on the site (larger image creative, no competitor upsell etc.)
• No limit on what can be MVT'd — a whole new category taxonomy can be tested against current without breaking existing CMS config — an entirely new site layout/codebase can be tested on 5% of traffic
• Targeting packs/multivariate tests can comprise data, visual and logic components together — for instance: trialling different product imagery, descriptions, PDP layout and recommend strip logic vs current config as a unified test package
• Similarly, permanent content targeting can be created to show different product creative and descriptions to :female_under_30 vs :female_30_up or showing dynamic fashion model image creative based on fit size signals (generative AI is almost at point of being able to build these)
• On the automated side — content entropy can be added by creating manual or synthetic variants of any object (a product with 5 different main images or a home page campaign strip with 20 available promotion ads but can only display 4) — the platform can MVT which works most effectively with which enduser biases — this pattern repeats everywhere — trial 6 alternative versions of a landing page, A/B test 2 different 'most popular' algo calcs on category filter etc.
• All targeting logic runs at bare-metal level with supporting caches for maximum performance (<300ms page render) — these targeting model caches are recalculated every few hours — the caches can also be A/B tested enabling scenarios such as a Pepsi-challenge on using AWS vs Azure AI frameworks behind specific features to be performed
• Where there is no overall winner as individual object candidates in test perform better with different cohort groups — the test can become a permanent routing so each cohort group receives their optimal experience going forward
• 100Ks concurrent automated MVTs and learned cohort routing optimisations are able to exist in parallel delivering vast variation in UX journeys available once sufficient object entropy and signal telemetry is available
• Almost all product discovery visible (category pages, home page fragments etc.) go through a targeting layer and are being pushed because of earlier signals from enduser or cohorts they are being classified as belonging — on a specialist fishing tackle store, a :sea angler would see minimal :fly products after first couple of clicks unless indication of wider interest observed
• Continually learning from successful/unsuccessful recommendations as expected — but also able to run bespoke inference such as assigning synthetic bias label :vegetarian to enduser after a couple of sessions without buying or interacting with a :meat labeled product
• The targeting packs have unlimited configurability and refine in realtime — opening out potential to do things like shifting recommend biases based on whether it is sunny or raining in London for endusers in area or behave differently on a Monday vs Friday (macro scenarios that can be tested against control even though they have site-wide impact)
• Able to bias range targeting to lean towards a financial goal-seek objective — such as prioritising clearing aged-stock over revenue in weeks before end of the half

Multidimensionable Data Model

• 4-dimensional data model that allows variant objects to seamlessly co-exist with control object and provide maximum signal resolution enabling targeting of content down to 1-v-1 granularity — in additional to normal measurement of quantity of interactions, the velocity of those interactions is tracked in realtime — and whether any object involved has been modified over time or is a test variant of the control object
• That high-resolution measurement granularity is able to detect anomalies such as conv rate drop after a product title change or identify that a variant test object is underperforming with :urban_male but overperforming for :rural_female vs control object
• Big focus on changes in velocity of signals, especially negative — as soon as :size_f_12 signals drop/stop (but :size_f_14 signals accelerate) — the realtime loop immediately starts filtering products available in 12, but not 14 from cache sets — even though it was a primary targeting label for the enduser in previous months
• The importance of velocity (trending) repeats everywhere as can be more interesting in product discovery scenarios than simply using static 'top' based calcs — repeat visitors already knowing the best sellers but maybe not aware of a product emerging from 80th to 30th place in category in last 7 days
• In addition to velocity, the weight of bias labels is also a key factor in maximising data available to predictive models — an attribute such as :colour being significantly less heavy a selection signal as :size

HYPV can be run initially in augmentation mode (like Algolia etc.) where it is only ingesting telemetry and feeding back some targeted content over API for insertion on existing site — fails silently if any issue, low risk — at a minimum you likely get some views on trading position and enduser behaviour not available in other analytics tools

Currently starting limited trials on live sites

contact@uncommerce.com

[Updated: 2025-02-25]