Methodology & evidence

How the AI Image Score evolves over time: reading the curve

October 24, 2026 6 min read
Kamel Malek Director of SetRecord.ai

The AI Image Score — the composite measure of accuracy, presence, and sentiment across the model panel — is not a static number you check once and file. It is a curve that evolves over weeks and months, and the shape of that curve carries more diagnostic value than any single data point. This article teaches you to read the curve.

The score is a curve, not a number

When an organisation first measures its AI Image Score, the natural reaction is to focus on the number: "We are at 72 out of 100." But a single score is a snapshot — it tells you where you are, not where you are going. The score becomes useful when you have at least three measurement points and can see the trajectory. A score of 72 that has been stable for six months tells a very different story from a score of 72 that has dropped from 85 in the last quarter. The curve, not the point, is the unit of analysis.

The four curve shapes

In practice, we observe four canonical curve shapes. Stable: the score fluctuates within a narrow band (±3 points) over multiple cycles, indicating a narrative that is neither improving nor deteriorating. Rising: the score trends upward over three or more cycles, indicating that corrections are propagating and the narrative is improving. Falling: the score trends downward, indicating either new distortions entering the model layer or corrections that are not persisting. Volatile: the score swings by more than 10 points between cycles, indicating an unstable source layer where the model is drawing from contradictory or rapidly changing sources. Each shape requires a different response.

Don't know what AI says about you today?

A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.

Request an audit

What drives score changes

Score changes are driven by three mechanisms. The first is training-cycle refresh: when a model is retrained, its latent narrative updates, which can cause a step-change in the score — up if the new training data includes your corrections, down if it includes new distortions. The second is source-layer change: when web sources that the model retrieves from are updated, created, or removed, the web-search mode score shifts — this is often the faster-moving component. The third is query-environment change: when the patterns of questions users ask change (due to news, events, or competitive activity), the presence rate and sentiment components shift even if the underlying facts have not changed. Understanding which mechanism is driving a score change tells you whether to respond with a training-cycle intervention, a source-layer intervention, or a messaging intervention.

Reading a regression

A regression — a score that drops after a period of improvement — is the most diagnostically valuable event in a DNI programme. It tells you that a correction did not persist, which means either the source was not durable enough, or a new distortion has entered and overridden the correction. The first step is to identify which component dropped: accuracy (a factual error returned), presence (the entity appears less often), or sentiment (the tone shifted negative). The second step is to trace the regression to a specific source or training event. A regression is not a failure of the programme; it is the programme working — it has detected a drift that would otherwise have gone unnoticed, and it gives you the information to intervene precisely.

What the curve tells your programme

Over a six-month programme, the curve tells you three things that no single measurement can. It tells you whether your correction strategy is working (rising curve), whether it has plateaued (stable curve after initial rise), or whether it needs a fundamental rethink (falling or volatile curve after sustained effort). It also tells you the velocity of change — how fast corrections propagate, which sets expectations for how long future corrections will take. SetRecord treats the curve as the primary deliverable: a single score is a data point, but a curve is a story — and the story of your AI narrative is what your board, your executive committee, and your communications team need to understand.

See what AI says about you

A SetRecord audit checks what supported AI models say against your verifiable facts — within five business days.