Semantic fingerprint
A semantic fingerprint is how an in-product agent stores an interface element without depending on its markup. Instead of a CSS class or XPath, it records the element's role, accessible name, landmark position, and neighbours, then matches candidates against that fingerprint at runtime, so a restyle or re-render doesn't break the reference.
New to the category? Read what AI customer support is and how it works.
Why not use a CSS selector?
Because selectors encode the DOM, and the DOM changes on nearly every deploy. A button that moves from one wrapper to another keeps its role, name and neighbours but loses its selector. A fingerprint survives the change; a selector silently points at nothing, which is how selector-based tools break every release.
How is a match scored?
Each signal contributes to a combined confidence between 0 and 1. Above 0.85 the agent acts; between 0.60 and 0.85 it re-reads the live interface first; below 0.60 it stops and hands off rather than guessing. The threshold is the safety mechanism, not a heuristic to tune away.
What does a semantic fingerprint look like in practice?
A designer moves the Export button from a toolbar into a dropdown and changes its CSS class. A selector-based tour now points at nothing. A semantic fingerprint still matches, because the button kept its role, its accessible name and its neighbours, so guidance keeps working after the redesign without anyone re-recording a step.
How does Peeve handle semantic fingerprints?
Peeve stores each element by role, accessible name, landmark position and neighbours, never by CSS class or XPath. At runtime it reads the live accessibility tree and scores candidates against that fingerprint: above 0.85 it acts, between 0.60 and 0.85 it re-reads the live interface first, and below 0.60 it stops and says so.