The terms behind autonomous conversion optimisation, each defined in a sentence or two, with a note where IntellQ does something specific.
Agentic CRO
- Conversion rate optimisation carried out by an AI agent rather than by a person running individual tests. The agent forms hypotheses, runs experiments and removes losing changes continuously.
Contextual bandit
- A bandit that chooses the best option for the current situation rather than one winner for everybody. The context can be the visitor's intent, source or history.
- In IntellQ: IntellQ uses seven intent contexts, such as pricing_friction and checkout_recovery, and ranks strategies within each.
Upper confidence bound (UCB)
- A bandit rule that scores each option by its observed success rate plus an exploration bonus that shrinks as evidence accumulates, then picks the highest score. It is deterministic, which makes decisions easy to audit.
- In IntellQ: IntellQ's score combines prior fit, a Beta posterior, lift over control, an exploration bonus and a negative outcome penalty.
Thompson sampling
- A bandit rule that draws a plausible success rate for each option from its posterior distribution and picks the option with the highest draw. Uncertain options win draws more often, which is how it explores.
Holdout control group
- A share of visitors who never see any change, kept aside so the effect of every change can be measured against doing nothing. Without one, an optimiser can report seasonality as lift.
- In IntellQ: IntellQ holds back 20 percent of visitors by default, configurable from 5 to 50 percent.
Sticky bucketing
- Assigning each visitor to a variant with a deterministic hash of their id, so they see the same variant on every page and every visit. Flipping visitors between variants makes results meaningless.
- In IntellQ: IntellQ hashes the experiment salt and visitor id with SHA-256.
Lift against control
- The difference between the conversion rate of visitors who saw a change and visitors in the control group, usually as a percentage of the control rate. It is the causal effect, unlike a before and after comparison.
Last exposure attribution
- Crediting a conversion to the most recent change a visitor saw inside a set outcome window, and counting it once. It stops one purchase being credited to several experiments.
Signal decay
- Weighting recent behaviour more than old behaviour, usually with an exponential half life, so a visitor's current intent is not drowned out by what they did weeks ago.
- In IntellQ: IntellQ uses a 48 hour half life.
Block level attention
- Time a visitor spent with each section of a page actually on screen, measured per block rather than per page. It separates a section that was read from one that was scrolled past.
- In IntellQ: IntellQ counts seconds while at least half a block is visible and the tab is in front, up to twelve blocks per page, with no text or field values recorded.
DOM slot
- An element on a web page that the site owner has marked as changeable by an optimisation tool, so the tool can only ever reach those elements.
- In IntellQ: In IntellQ a slot is any element with a data-intellq-slot attribute. The agent can set its text, show it, hide it, or point it at a path on the same site.
ChangeSet
- IntellQ's name for one proposed change to a page: the slot operations, the hypothesis, the evidence, the expected outcome and a sandbox preview, reviewed before it goes live.