Skip to content

Glossary

Every abbreviation and bit of jargon used across these notes, grouped the way the site is. Where a concept has a note of its own, the term links to it.

Alpha (α)The false-positive rate you’ve agreed to accept, conventionally 0.05
Always-valid inferenceSequential methods whose p-values stay honest however often you look
Base rateThe share of tested ideas that contain a real effect. Sets how many of your “wins” are false
BaselineThe control’s current rate. One of the four sample-size inputs
BayesianUpdating a prior belief with the data to get a probability the variant is better
Benjamini-HochbergA multiple-testing correction controlling false discovery rate. Less brutal than Bonferroni
Beta (β)The false-negative rate. Power is 1 − β
BonferroniDivide alpha by the number of comparisons. Simple, and usually too strict
CIConfidence interval. The range of effects the data hasn’t ruled out
Cohort analysisGrouping customers by when they arrived and tracking each group forward separately
CovariateA pre-test variable used to strip noise out of a result
Credible intervalThe Bayesian interval, which does mean what people wrongly think a CI means
CUPEDControlled-experiment Using Pre-Experiment Data. Uses pre-test behaviour to strip noise out of the result
Delta methodThe variance maths that makes ratio metrics behave
Effect sizeHow big the difference is, as distinct from whether it’s significant
Expected lossThe average cost of shipping the wrong variant, weighted by how likely each wrongness is
Family-wise error rateThe chance of at least one false positive across a whole set of comparisons
FDRFalse discovery rate. The share of your “significant” findings that are false
FrequentistAsking how surprising the data would be if there were no effect. Produces p-values
H0 / H1Null and alternative hypothesis. H0 is the boring one
HARKingHypothesising After the Results are Known. Writing the prediction once you’ve seen the answer
Heavy tailA distribution where rare large values dominate the average
M-type / S-type errorRight direction but inflated / wrong direction entirely
MDEMinimum detectable effect. The smallest lift you commit to being able to see
Novelty effectEarly engagement inflated purely because the thing is new
p-hackingRunning comparisons until one crosses 0.05, then reporting that one
Point estimateThe single observed effect, before you look at the interval around it
Posterior / priorBelief after / before seeing the data
Regression to the meanExtreme early readings drifting back toward the truth, which looks like decay
ROPERegion of practical equivalence. The band around zero you’d call “no meaningful difference”
RPVRevenue per visitor. Conversion rate × AOV, and brutal to test on
ShrinkageExtreme estimates pulled back toward zero, usually correctly
SignificanceThe verdict that a result is unlikely enough under chance that you’ll act on it
Simpson’s paradoxA trend in every subgroup reversing once you combine them
SPRTSequential Probability Ratio Test. The original sequential method
Zero-inflatedA distribution where most observations are zero, like revenue per visitor
AdvertorialA long-form, editorial-styled landing page built for cold paid traffic
AscensionA customer moving up to a more expensive rung
Awareness stagesUnaware, problem-aware, solution-aware, product-aware, most aware
BlendedAveraged across all channels or cohorts, usually hiding the variation that matters
BOFUBottom of funnel. Decision stage
CACCustomer acquisition cost
CLV / CLTVCustomer lifetime value. Same thing as LTV, different house style
CROConversion rate optimisation
Demand captureGetting in front of buyers already shopping. Easy to measure, and finite
Demand generationCreating want that didn’t exist. Hard to measure, and where the ceiling is set
JTBDJobs to Be Done. What the customer hired the product to accomplish
LTVLifetime value. Total revenue from a customer over the relationship
MOFUMiddle of funnel. Consideration stage
OfferEverything the buyer is choosing between. Price, terms, bonuses, guarantee, urgency
Pre-sell pageA short warming page sitting between the ad and the product
TOFUTop of funnel. Awareness stage
Traffic temperatureHow much prior context a visitor arrives with. Cold, warm, hot
UTMThe tracking parameters bolted onto a URL to identify the campaign
Value ladderA deliberate sequence of offers at rising price points
AIDAAttention, Interest, Desire, Action. A copywriting skeleton
AnchoringThe first number seen setting the reference point for every number after it
BNPLBuy now, pay later. Klarna, Clearpay and similar
Charm pricingPrices ending .99, leaning on the left-digit effect
Cialdini’s six principlesReciprocity, commitment, social proof, authority, liking, scarcity
Cognitive loadThe total mental effort a page demands
Compromise effectThe pull toward whichever option sits in the middle, regardless of dominance
ConfirmshamingGuilt-worded opt-outs. “No thanks, I don’t want to save money”
Curiosity gapThe pull to close a gap between what you know and what you want to know
DecoyA third option added to make your target option look obviously better
Default biasWhatever is preselected gets picked disproportionately
Endowment effectValuing something more once it’s yours. Why free trials and easy returns work
Fogg behaviour modelB = MAP. Behaviour needs motivation, ability and a prompt at once
FramingSame fact, different presentation, different decision
Goal gradientPeople accelerating as they get closer to a visible finish
Halo effectOne good impression colouring judgement of unrelated things
Left-digit effectReading £19.99 as “nineteen-something” before processing the rest
Loss aversionLosses hurting roughly twice as much as equivalent gains please
Mental modelThe expectations a user arrives with about how your category works
Negative haloOne bad signal contaminating perception of everything else. Faster than the positive kind
Open loopAn unfinished task or question the mind keeps wanting to close
Paradox of choiceMore options producing fewer decisions, past a threshold
PASProblem, Agitate, Solution. A copywriting structure
Peak-end ruleExperiences remembered by their peak moment and their ending, not their average
Prestige pricingRound numbers, signalling confidence rather than value
Processing fluencyHow easily the brain decodes what it’s looking at. Fluent reads as true
ReactancePushing back because you’ve noticed you’re being pushed
Risk reversalMoving the perceived risk of buying off the customer and onto you
ScarcityLimited quantity. Urgency’s sibling, and the one that’s usually faked
Social proofLooking to others’ behaviour to decide your own
Sunk costContinuing because of what you’ve already spent, not what’s ahead
System 1 / System 2Fast automatic thinking / slow deliberate thinking
UGCUser-generated content. Customer photos, video, reviews with images
UrgencyLimited time. Scarcity’s sibling
Zeigarnik effectUnfinished tasks staying in mind in a way finished ones don’t
A/B testA randomised controlled experiment comparing two versions of something
AA testA control-vs-control experiment, run to check the testing infrastructure is honest
BanditAn algorithm that reallocates traffic toward the winner while the test runs
BucketingAssigning a user to a variant, usually by hashing their ID
Cluster randomisationAssigning whole groups instead of individuals, when spillover makes user-level assignment invalid
Concierge testThe feature appears to work and is fulfilled by hand behind the scenes
ControlThe existing version, held constant so the variant has something to be measured against
Epsilon-greedyThe simplest bandit. Mostly show the leader, occasionally explore
Experiment velocityHow many tests you get through. Compounds harder than per-test rigour does
Fake doorA painted door that admits the feature doesn’t exist yet
Friction analysisMapping where the user is working harder than they need to, before generating tests
GuardrailA metric you watch for damage, not for wins
Heuristic reviewWalking the site against a usability checklist, noting friction
HiPPOHighest Paid Person’s Opinion, overriding the result
HoldoutA slice kept on control after launch, to measure the long-run effect
HTEHeterogeneous treatment effect. The effect differing by segment
ICEImpact, Confidence, Ease. A prioritisation score
InterferenceOne user’s outcome depending on another user’s assignment
Listwise deletionDropping any record with a missing field. The silent default in most tools
MCAR / MAR / MNARMissing completely at random / at random / not at random. Only the last one biases you
Metric shoppingTrying metrics after the fact until one crosses the line
Mutual exclusionCarving traffic so two conflicting tests can’t overlap
MVTMultivariate test. Several variables at once, one variant per combination
Orthogonal assignmentIndependent bucketing per experiment, so concurrent tests don’t confound each other
Painted doorA button for something that doesn’t exist yet, to measure whether anyone wants it
PeekingChecking results before the planned sample, which inflates your error rate
PIEPotential, Importance, Ease. A prioritisation score
Pre-registrationWriting down metric, sample and stopping rule before the data arrives
Primary metricThe one number that decides the test, committed to beforehand
Proxy metricA stand-in for something you can’t observe inside the test window
Randomisation unitWhat you assign. It has to match what you analyse, or the variance is wrong
RICEReach, Impact, Confidence, Effort. A prioritisation score
SaltThe per-experiment seed that keeps bucketing independent between tests
Segment fishingSlicing the data until something wins, then reporting that slice
Sensitivity analysisRe-running the analysis under other reasonable choices to see if it holds
Smoke testThe lightest painted door. A button leading to “notify me”
SpilloverTreatment effects leaking into the control group
SRMSample ratio mismatch. The split isn’t what you asked for, so the test is invalid
SUTVAStable unit treatment value assumption. One user’s outcome doesn’t depend on another’s bucket
SwitchbackRandomising time periods instead of users, when the whole system shares state
Thompson samplingA bandit rule allocating by how uncertain each variant still is
TreatmentThe variant carrying the change being tested
UCBUpper confidence bound. A bandit allocation rule
VariantAny arm of the test, including control in loose usage
Win rateThe share of tests producing a real win. Above 30% is a diagnosis, not a boast
Winner’s curseShipped winners overstating their true effect, because you only ship what looked good
WinsoriseCap extreme values at a percentile instead of dropping them
Wizard of OzThe user thinks they’re using software. It’s a person
CAPIConversions API. Sending conversion events server-to-server rather than from the browser
CLSCumulative Layout Shift. How much the page moves about after load
CWVCore Web Vitals. LCP, INP and CLS
Data layerThe structured object on the page that every downstream tool reads from
DeduplicationMatching the browser and server copies of one event so it counts once
Event taxonomyThe naming and property conventions your analytics events follow
Feature flagA runtime switch deciding whether a user sees a feature, without a deploy
FIDFirst Input Delay. The Core Web Vital that INP replaced
FlickerThe control showing briefly before a client-side variant swaps in
FOIT / FOUTFlash of invisible / unstyled text. Web fonts causing late text shifts
FOOCFlash of original content. The same thing as flicker
GA4Google Analytics 4
GDPRThe EU and UK data protection regime. Why declined consent means no cookie
Growth loopA system where customers create the conditions for the next customers
GTMGoogle Tag Manager. “Server-side GTM” means the container runs on your infrastructure
Identity stitchingJoining a user’s activity across sessions and devices into one record
INPInteraction to Next Paint. Lag between input and visible response
Input metricAn operational metric that compounds into the north star
ITPIntelligent Tracking Prevention. Safari’s cookie restrictions
Kill switchTurning a feature off in seconds without a deploy
LCPLargest Contentful Paint. How long the main element takes to render
Match qualityHow well the identifiers you send let a platform tie an event to a person
Modelled conversionsPlatform estimates of conversions it couldn’t observe. Never analyse a test on these
North starThe single company-level metric everything is meant to compound into
PIIPersonally identifiable information. Hash it before it leaves your server
SPASingle-page app. React, Vue, Next and similar
TBTTotal Blocking Time. Not a Core Web Vital, still worth watching
TTFBTime to First Byte. Not a Core Web Vital, but upstream of all of them
Vertical sliceShipping one feature through every layer it touches, rather than each layer broadly
ActivationThe point a new user first gets real value. What trial conversion actually depends on
AOVAverage order value
ARRAnnual recurring revenue. MRR × 12, and a projection rather than money in the bank
ChurnCustomers leaving. Voluntary if they chose to, involuntary if a card failed
CM1 / CM2 / CM3Contribution margin after goods / after fulfilment / after acquisition
ContractionExisting customers paying less through downgrades
CPLCost per lead. Misleading on its own, because it isn’t cost per customer
CRMCustomer relationship management. Where the closed-won data lives
Dark patternA design that gets the outcome by obstructing the user rather than persuading them
DTCDirect to consumer
DunningRecovering failed subscription payments through retries, reminders and card updates
ExpansionExisting customers paying more through upgrades, seats or usage
First-order economicsThe maths of the first order, which in DTC usually loses money on purpose
FreemiumA permanently free tier with paid features above it. No clock
Geo holdoutTurning a channel off in matched regions to measure what it actually caused
Ghost adsShowing the control group an unrelated ad where yours would have run
GRRGross revenue retention. NRR with expansion taken out, so it can’t hide churn
IncrementalityWhether the spend caused the sale or just took credit for it
MERMarketing efficiency ratio. Total revenue ÷ total marketing spend
MQLMarketing-qualified lead. Someone who looked interested
MRRMonthly recurring revenue
NRRNet revenue retention. What a cohort’s revenue does over a year, before new customers
PQLProduct-qualified lead. Someone who has used the product and hit value
Progressive profilingBuilding a lead’s profile across visits instead of one long form
PSA testShowing a public-service ad to the control group to measure incrementality
ReactivationA churned customer returning
Repeat purchase rateThe share of customers who buy again. The early read on LTV
ROASReturn on ad spend. As the platforms report it, structurally overstated
SALSales-accepted lead. The rep has agreed to work it
Save flowWhat sits between the cancel button and the actual cancellation
Self-reported attributionAsking the buyer how they heard about you, because tracking can’t see the dark channels
SLAService-level agreement. Here, the committed maximum response time on a lead
Speed-to-leadHow fast a new lead gets contacted. The curve is brutal in the first hour
SQLSales-qualified lead. Worked by sales and confirmed worth pursuing
Time-to-valueHow long before a new user gets something useful. Sets the right trial length
Value metricThe unit you charge by. Upstream of every pricing-page decision
Win-backRe-converting a customer who has already churned
B2B / B2CBusiness-to-business / business-to-consumer
CPMCost per thousand impressions
CTACall to action
CTRClick-through rate
PDPProduct detail page
SKUStock keeping unit. One sellable product variant