Globally unique identifier automatically assigned to all objects with embedded timestamp
ID
VID
CID
GID
OBJECT
:_id8 vid1 c0 draft
:vid1 c0
:cid c0
:gid
:hermes_oran_sandal
current
:_id7 vid1 c0 committed
:_id6 vid1 c0 previous
:_id5 vid1 c0 off
:_id4 vid0 c0 off
:vid0 c0
:_id3 vid1 control draft
:vid1 control
:cid control
current
:_id2 vid1 control committed
:_id1 vid0 control off
:vid0 control
:_id0 vid0 control off
This shows all the :_id that may exist for a single object
The :hermes_oran_sandal example is a single product with
a control variant and one test candidate
c0 against it — both have had some edits so versions
v0, v1 etc. exist
Traditional ecommerce data models have a single row per object (with one id per object) such as
a product or category in a table — which gives no mechanism for multiple variants of the
object to exist in parallel for A/B or MVT testing
Advanced MongoDB data model designed specifically for HYPV that stores, as equals, any
variation clone of an object and every change to those objects (similar to Git branches
+ version history)
Multiple copies of an object can be live concurrently — including test variants (e.g.
improved product description vs control) or version variants where a title has been modified
A method of comparing two objects (variant A vs. variant B) to determine which one performs
better in achieving a specific goal
Actionable Intelligence
Metrics and insights that are not just informative but can be directly used to make decisions
and implement changes, such as the percentage of top selling products in stock today
That is a measurement that can be obviously impacted with activity — vs metrics like
AOV which usually have no mechanism to influence directly
Average order value — only tracked in places as a byproduct — it is not usually a
particularly useful metric for performance targeting as trying to force it higher can often
reduce overall gains
Targeting specifics such as 'attachment rate' or bundling metrics can be more useful in that
area
Attachment Rate
Whether an optional item is attached to an order or upsold, cross-sold
The process of identifying and assigning credit to the various marketing channels such as
Meta Ads — it is critical to know as much as is possible where ad traffic is landing and
how it is converting to get a view on ROAS and identify areas to increase/decrease spend at a
granular level
Augmentation API
A subset of the HYPV engine functionality to deliver recsys and other UX improvements over
API endpoints to existing platforms
In the context of HYPV, bare metal is when concepts that usually require add-ons such as
A/B testing, search and recommend systems all run as part of the platform itself without
needing external tech
This increases the granularity of the signal telemetry available for targeting and means all
objects involved are peers — so the performance is within platform control, not the
lowest common denominator of the add-ons in place
Behavior
The set of actions, interactions, and patterns exhibited by an enduser while engaging with
the platform — this includes clicks, page views, purchases, search
queries, time spent, and UX journeys
We are interested in the positive actions, the negative actions (things they are not doing)
and also what we can infer from those at a higher level — such as infering vegetarian
label from someone who is not purchasing or even viewing meat over several sessions
A deviation or inference of normal from enduser behavior — for instance
[:size_12, +6], [:size_14, +2] from a session filtering clothing results
Blit
The process of taking the denormalised backend data which includes all the variate instances,
and transposing it to 'origin' objects (so there is a reference point back to where the data
came from) — then flattening everything out to the 'edge' database servers
This happens every time a deploy of the frontend is initiated
The creation of a wholly new structure for each deploy allows absolutely any combination of
MVT configurations to run in production — such as running two completely different
category taxonomies
A specific object variation (e.g. content element) that is being considered or tested for its
potential effectiveness or suitability for a particular segment
Candidates can also persist permanently — for instance when a set of creative assets are
performing better than control for a specific cohort
:cid
The unique identifier for a candidate of an object, it can have many versions (vids)
In the context of the 4D Model, clone is a copy of an object created to
serve as a basis for a new variation or test — clones allow for parallel development and
testing of different object states without altering the original
CLV
Customer lifetime value
Control
In A/B testing or experimentation, the 'control' is the existing or baseline version of an
object against which new candidate variations are compared to measure their impact
Code & Algorithms
In the context of hypervariation, this refers to the software logic, machine learning models,
statistical methods, and rule-based systems that power the HYPV platform
These behave exactly like the data and visual objects when testing — so two code
libraries or algos can be tested against each other
AI tech that enable computers to understand, process, and respond to human language in a
natural, conversational way
HYPV native platform is experimenting with conversational AI to drive shopping experiences
for enhanced user interaction
Cross-sell
Encourage enduser to buy additional items or attach an optional accessory
The overall experience a customer has with a retailer — of which the UX is only part
It includes email communications, shipments, CS queries, returns and marketing
In the context of hypervariation, this encompasses all information processed and utilized
by the HYPV system which is delivered to the enduser
This primarily includes 'objects' (e.g., products, content, UI elements) and their
variations
A defined period of time from which data is collected and analyzed
Shorter windows will show velocity of targetable user behavior such as increase in interest
in fishing tackle for a user that has never interacted with category before
Longer windows are useful to see seasonal trends ahead of time which may not yet be visible
in the short windows (so start pushing garden furniture before Easter not after)
Demographic Segment
Characteristics like age, gender, income and occupation
In the context of A/B/MVT testing or optimization within HYPV, the process of selecting
a specific variation (candidate) to be targeted at a cohort or enduser
Which may be for an already calculated bias reason or because the user happens to be in an
experiment group
An individual interacting with the platform — they can be anonymous initially before
becoming known once logged in
Experiment Group
All endusers are assigned membership of some preset experiment groups, which enables a fast
mechanism for selecting 5% or 10% of whole audience for any site wide tests
An attribute or characteristic of an object (usually product) that can be used for
categorization, filtering, or searching — for instance size or color
Fragment
A distinct, often reusable, component or section of the user interface or page layout
Examples include header, footer or product recommendation strip
These can be varied for testing as with object data and other targetable components
Funnel
A defined series of pages that a user progresses through to achieve a specific goal
Recsys sets that are built from multiple sources to form a single set
For example purposes, a 20 object set could be formed from:
15 short-window trending object interactions in same category for cohorts enduser is member
15 wider-window trending object interactions in same category for cohorts enduser is member
Backfill appended with trending objects related to all endusers who interacted with object
Removal from set of anything we do not want to promote like out of stocks — or say the
product is from a sponsored brand so we remove any competitors
Interlace into the set 3 sponsored products that are in proximity but did not appear natuarally
in the earlier calcs
That is just an illustration — the real calcs would do more to filter earlier etc.
Hypervariate
The concept that multiple candidates of any object may exist and be targeted concurrently
Enabling completely unique experiences to be delivered on a per-enduser basis
Hyperpersonalize
This refers to the practice of using detailed enduser data and predictive AI to tailor
experiences, content, and offers to individual users or microsegments at an extreme level
of granularity
HYPV has the ability to vary any data, code or visual content rendered by cohort or individual
enduser
HVE, HVE50 and HVE1000
High-value endusers — pre-built cohorts that can be used for targeting
and track in some of the intelligence visualizations
Giving opportunities to vary:
run an ad-free, upsell-free experience for HVE1000, excluding all brands likely to be of
low interest &mdash changing layouts for HVE1000 to use larger image collateral
offer specific products only to HVE1000 which are not available to general public but
appear within the normal UX and product discovery — or push some sponsored luxury
brands harder
Inferred Label
The calculation that an enduser is vegetarian as they never interact with meat labeled
products or categories
Input Metric
Metrics that we have some power to improve for instance '% Top 250 SKUs In Stock'
Improving input metrics with action should result in improvements to output metrics like
'GMV' or 'NPS'
Interlace refers to the technique of alternating or combining different rec algorithms
or item sets to create a compound set for delivery to the enduser
This may include pulling in some sponsored or promoted objects to the calculated set
IPO
Items per order
Label
A unique text identifier for an object using a symbol format
Labels are fundamental in HYPV for training supervised machine learning models and for
triggering targeting rules
A numerical value assigned to a label indicating its importance
For instance size labels are more important than more subjective labels like color
Label Velocity
The rate at which labels are applied, changed, or accumulated over time for a user, object,
or segment
This can indicate trends, evolving interests, or the freshness of certain
characteristics, which HYPV can use for dynamic adaptation
For instance it is useful for targeting to know the interactions with size_12 are increasing
and at the same time size_14 is decreasing
The practice of measuring, monitoring, and analyzing delays (latency) within a system or
across various components (e.g. HYPV API response times, page load times with variations,
data query speeds)
Pagespeed is core to UX performance so every aspect in HYPV that can be measured for speeed,
is
Layout
In the context of hypervariation, this refers to the arrangement and structure of visual
elements on a page or screen
Layouts can be dynamically altered by HYPV, including the
inclusion, exclusion, or rearrangement of 'fragments', entire 'pages', or sequences of
pages ('funnels')
These layouts themselves can be treated as 'objects' subject to variation
An older technology system, application or e-commerce platform that is still in use but may be
outdated, difficult to modify, or lack modern capabilities
HYPV can integrate with legacy platforms via an Augmentation API to add predictive-AI features
not available from other sources
As HYPV can be modified per client there are no restrictions on what can be done to meet
constraints in client technology
Some aspects of HYPV can be fully automated (synthetic), partially automated or manual
Manual refers to situations like the ability of merchandisers to push a specific object
regardless of what the recsys feels should be pushed
A highly specific and granular subdivision of a larger user segment, often defined by a narrow
set of characteristics, behaviors, or preferences
Microsegments enable more precise targeting and hyperpersonalization by HYPV
For instance [:female_30_up, :city]
Mobile/Web
Refers to user experiences delivered through web browsers on mobile devices (smartphones,
tablets) — this is targeted in same was as desktop
This is distinct from native mobile apps
Refers to user experiences delivered through native applications installed on mobile devices
(smartphones, tablets)
For instance apps downloaded from app stores for iPhone or Android
HYPV can target these environments with API variance but it is usually more limited than
with web interactions
Multivariate Testing — a method of testing in which multiple variations of multiple
elements on a page or within an experience are tested simultaneously to determine the
combination that performs best
A core capability supported by HYPV which can have unlimited MVTs running in parallel
Native Platform
The use of HYPV as a full platform (not just an augmentation API) which unlocks the full
hypervariation feature capabilities of the 4D data model
Negative Labels
In the inference context these are labels that explicitly indicate a negative for instance
vegetarian from lack of meat interactions
In conversational AI user interface context this is where voice feedback is received that
can assign negative labels to objects which were not previously know — "no those are
too boho, show me more mainstream"
Object
In the HYPV system, a fundamental unit of data or content that can be managed, versioned,
and varied using the 4D Model — this can include products, categories, content articles,
UI elements (fragments, layouts), media assets, or any other distinct entity (_id) that is
subject to personalization or testing
Origin
Intermediate objects created during a blit to allow link to edge objects
1. to reference where they came from
2. to enable an efficient mechanism for soft updates (price, stocklevel etc.) that do not
require a full deploy
Metrics which cannot be changed directly but do tell us whether our input metric covered improvements are having an impact on the wider goals
Examples are 'repeat visit frequency', 'NPS', 'blended ROAS' and revenue / margin
Output metrics are primarily for monthly reports and the exec team
A single document or screen within a website or application, typically identified by a unique
URL or view state
Pages are often composed of multiple 'fragments' and can be part of a larger
user 'funnel'
HYPV can vary content at the page level or sub-page (fragment) level.
Page Type
The classification of the page (the template)
This enables measurement of all instances of page template and how they interact — for
example the frequency of transition (CTR) between a category page and a product page
PAGE
ABBREV
DESC
:home
HME
Home page
:category
CAT
Category listing page
:product
PRD
Product page (PDP)
:brand
BRD
Brand or vendor page
:list
LST
Listing page — a collection of products usually from multiple categories
:search
SRC
Search page — the search results page or advanced search page
:lander
LAN
Landing page — a bespoke page for a specific campaign or similar
:cart
CRT
Cart page — the first page of the funnel or 'view cart', which can run upsell
:account_home
AHM
Account entry page, which may have recently viewed or similar on there
:account_view_order
AVO
Account page for looking at a previous order
Polyfill
In context of HYPV this refers to additional features and APIs created when working with
legacy platforms to add logic beyond that normally needed for hyperpersonalization but is
missing and impacting UX
For instance some platforms do not have tiered categories or breadcrumbs, HYPV can also
supply these if needed
Predictive AI
The use of historical and real-time data to make predictions about future outcomes, behaviors,
or trends of a cohort or individual enduser
It employs machine learning algorithms (like regression, classification, clustering) and
statistical modeling to analyze patterns and build a range of targeting caches, decision
trees and lookup tables
Used to disambiguate with Generative AI when talking about the core HYPV technology
Product Discovery
The process by which users find and learn about products on an ecommerce platform
This includes search, nav, recommendation strips and other features that help users
identify objects relevant to their needs or interests
Psychographic Segment
Lifestyle, values, attitude and personality signals
A full targeting output generated by HYPV and refreshed every few hours which contains
pre-calculated decision trees, large scale caches and similar to ensure very fast UX even
though all rendered objects are targeted specifically
Web server logic that adapts quickly to changes in signal biases when it becomes clear the
quant pack caches are too stale to use for the current enduser session
For instance the signal change cliff that happens at the end of Christmas shopping period
and users start immediately behaving very differently
The ability of site to keep endusers returning over time
A core KPI for HYPV and monitored at a granular level (by category, by device, by UI medium
etc.)
ROAS
Return On Ad Spend — even though HYPV is mostly focused on targeting, this is part of
the core offering as it is critical to wider performance
Where is ad traffic landing and how is it behaving
Routing
All pages, fragments and algos go through a router which elects whether the enduser get
control or candidate of any object
A frontend end source pack which contains all visual code for rendering
Multiple can be in existance at any time to enable different visual content for testing or
targeting
Can also be used to create different layouts for whitelabel sites or international variants
A group of endusers sharing one or more common characteristics, behaviors,
or attributes (e.g., demographic, geographic, behavioral, psychographic)
Segmentation Strategy
Partitioning of a range into macro segmentation groupings (sea vs coarse in fishing, male vs
female on a fashion site) that is used in obvious targeting situations and reporting
Session
Despite restrictions on cookie-based tracking and many analytics-type tools moving away from
the concept — it is still a very important targeting concept
HYPV can take telemetry direct from platform when running as an augmentation API, avoiding
issues — when running headless or natively it is first party anyway
Sessions go through substantial post-processing to extract behavior patterns and deduplicate
attribution — to enable additional bias computations on the enduser and get a clearer
view on ad spend
Signal & Signal Granularity
An individual pageview, click, purchase or search query
Any automated generation of an object — for instance a pack of creative assets used
to feed a merchandising strip
The packs can be created synthetically for targeting to cohorts (vs manually created packs
by the merchandising team)
A series of pages or a funnel when analyzing and visualizing enduser behaviour or impact
of tests on transitions from a page — for instance what happens next downstream from a
landing page and which tracks are converting higher than others
TRX
Transaction — an order — HYPV is mostly only interested up to the point of
receiving one, everything after that is out of scope
Upsell
Encouraging an enduser to choose a higher-priced or upgraded version of the same product they
were initially considering
When any object is modified in CMS a new version is created (like a git commit) this enables
the performance of changes to be tracked over time — for instance if a title is modified
the CR for product can be seen before and after by version id
Shopping using voice commands on web or mobile — in context of HYPV it benefits from
the labeling system in use and also allows collection of user contributed labels from
interactions &mdash "not like that, those are all too boho, show me more mainstream"