Measuring AI visibility: what actually matters.

Measurement has always been half the SEO discipline. AI search is no different, except the layer is newer and vendors have not converged on definitions. This post is opinionated about which metrics are load-bearing, which mislead, and what an ecommerce team should put on the wall. The short version: measure per SKU, per engine, per query intent — then roll up. Everything else is a derivative.

The metric primitive

Start with the atomic question: for this product, on this engine, when buyers ask with this intent, how often is the product named, cited, and characterized accurately? That is the primitive. Every useful rollup — category visibility, catalog score, trend lines — derives from it. Tools that do not store at this granularity can only produce aggregates, and aggregates are the metrics most likely to mislead. The AI visibility score definition shows how the primitive composes into one number per SKU; in eCommerce Insights it splits further into a citation score (is the engine recommending the product) and an agent-readability score (can a shopping agent parse the PDP) because the two fail independently and get fixed by different teams.

Why share of voice misleads

Share of voice (AI) sums citations for a brand across a prompt set and expresses it relative to competitors. Useful for comms. For ecommerce it can move in ways that do not predict revenue: a long-tail SKU grabbing mentions offsets a franchise SKU losing them, and the rollup hides the swap. This is the same blind spot covered in brand-level tracking is missing your revenue — any metric aggregated above the SKU loses the information an ecommerce team acts on.

Weekly vs quarterly instruments

Weekly data is noisy — establish each metric's noise floor once before reading trends. Quarterly metrics smooth the noise and match how leadership reads the story. Do not make quarterly metrics do weekly work; they are different instruments.

What not to track

Skip metrics that feel rigorous but add no decision value: total prompt-set impression counts (ambiguous denominator), single-engine share of voice as a solo metric, AI-sentiment scores uncalibrated to product specifics, and click-through from AI answers while the attribution layer stays immature — most of what is reported there is inferred. The test for any metric: can you define what success looks like before running the measurement? If not, it is ornamental.

The dashboard that moves work

The view that drives action is deceptively simple. One row per SKU. Columns: revenue weight, score per engine, recent citation count, trend arrow. Filters for category, collection, variant. A single sort: revenue-weighted visibility gap, descending. The top ten rows are the PDP triage list for the next two weeks; everything else is read-only context.

Most AI visibility tools the team has reviewed in 2026 bury this view under brand-level charts. The charts are fine for executive readouts; they are the wrong instrument for the person fixing PDPs. Build or buy for the action view first — it is the design center of SKU-level tracking and of the product generally.

Tying metrics to revenue, honestly

Attribution is not mature. Engines pass referrer data inconsistently when users click cited links — Google's GA4 channel-group documentation shows how recently AI sources became distinguishable at all — and many answers are read without any click. What works as of mid-2026 is directional correlation: track visibility and revenue by SKU, look for relationships where visibility moved meaningfully, report the pattern with caveats. Treat visibility as a leading indicator, not a revenue metric. AI traffic analytics is where the attribution story will eventually land; the product AI visibility guide covers interim approaches.

Key takeaways

• Measure per SKU, per engine, per query intent. Roll up afterward.
• Two scores per SKU — citation and agent-readability — because they fail independently.
• Share of voice is a comms metric; it misleads as a revenue predictor.
• Weekly: citations, source mix, characterization flags. Quarterly: threshold percentages and trends.
• The load-bearing dashboard is one row per SKU sorted by revenue-weighted gap.
• Attribution is forming; treat visibility as a leading indicator.