Why Per-Engine Measurement Beats a Blended Average

Measure AI visibility per engine, not as one blended average: Profound found only about 11% citation overlap across engines, so a single "AI visibility" number averages together separate universes and hides where you actually win or lose. Each engine is its own game; a blended score is a figure that corresponds to no real surface.

ChatGPT

Perplexity

AI Overviews

Gemini

“best X for Y”

Often

Rarely

Sometimes

Rarely

“X alternatives”

Sometimes

Often

Rarely

Sometimes

“is X worth it”

Rarely

Often

Sometimes

Rarely

strong needs work critical

Why does a blended average mislead?

A blended average misleads because it combines engines that barely agree on what to cite, producing a number that matches no real surface. Profound found only about 11% of cited sources are shared across the major engines — they're effectively separate universes. So if you average a 60% share of voice in Perplexity with a 5% share in Google AI Overviews, the "32%" blended figure describes nothing you can act on: you're not at 32% anywhere. Worse, it can stay flat while one engine quietly collapses, because a gain elsewhere offsets it.

A worked example: how the blend hides a collapse

Watch what a single averaged number conceals. Suppose you track three engines and report a naive average:

The blended 40% is identical both months, so a dashboard built on it reports "steady — nothing to do." But Google AI Overviews just halved, masked by a gain in Perplexity. If AI Overviews is where your buyers are, you're losing badly and the average is actively hiding it. The per-engine view makes that collapse impossible to miss — and tells you exactly which engine to investigate.

How little do engines actually overlap?

Engines overlap remarkably little: about 11% of cited sources are shared across them, per Profound. The cause is structural — each engine runs a different index, a different retrieval-and-rerank pipeline, and a different weighting of freshness, authority, and community sources. On top of that, citations are volatile within each engine: Semrush found 40–60% of LLM-cited sources change month to month. Ahrefs has watched a single source's presence swing from 76% to 38% across a tracking window, and Semrush found Google keeps a given URL in an AI Overview only about 3.87 days on average. So each engine is not only different from the others but a moving target in its own right. Low overlap across engines and high churn within them is exactly the situation where a single averaged number is least trustworthy.

What does per-engine measurement look like?

Per-engine measurement reports a separate share of voice for each engine you track, on the same fixed prompt set, so you can see exactly where you stand on each. The mechanics are the same as any share-of-voice program — you just never collapse the results into one figure.

1. 1

   Run one prompt set across each engine

   Use the same fixed questions on every engine so the engines are comparable to each other, not to a moving target.

2. 2

   Score each engine separately

   Compute brand and competitive share of voice per engine — ChatGPT, Perplexity, Google AI Overviews, and any others your audience uses.

3. 3

   Prioritize by audience

   Weight your attention toward the engines your buyers actually use, not toward whichever inflates the average.

4. 4

   Track each trend over time

   Watch the slope per engine; a healthy total can still hide a sharp decline on the one that matters most.

Why do engines differ so much in the first place? Because they retrieve and rank differently — the mechanics are in ChatGPT vs Perplexity vs Google AI Overviews: how citation differs and how AI engines choose citations.

How should you report per-engine results?

Report per-engine results as small multiples, not one merged line — a compact panel per engine showing its share-of-voice trend, side by side. That layout keeps each engine's reality visible while letting you scan the whole picture at a glance. Three practices keep the report honest:

• One row per engine, with its trend. Show current SOV and the direction over the last several readings, per engine — never just a single combined figure.
• Order by audience, not by score. Put the engines your buyers actually use first, so attention follows revenue rather than whichever number looks best.
• Annotate changes. Note when an engine shipped a known update or when you changed tactics, so a swing has context — the kind of evidence the Adaptability pillar runs on.

Is a blended number ever OK?

A blended number is acceptable only as a rough executive headline, and only when it's reported alongside the per-engine breakdown so it can't hide a collapse. If leadership wants one figure to glance at, give them one — but make every decision about where to invest from the per-engine view, because that's where the signal is. The moment a blended score starts driving strategy, you're optimizing for an average of universes that don't overlap.

If you must roll up to one figure, a usage-weighted index beats a naive average: weight each engine's share of voice by the share of your audience that actually uses it, so the number at least tracks the surfaces that matter to you rather than treating a niche engine and your buyers' main engine as equals. It still hides per-engine detail, so keep it as a headline sitting on top of the breakdown — never as the figure you optimize against.

Where this fits in the Canon

Per-engine measurement is the Adaptability pillar made operational: the engines change monthly and overlap little, so you measure each one and adapt on its own evidence. It's the discipline behind share of voice and the reason a single "AI visibility" number is a trap.

Related: ChatGPT vs Perplexity vs Google AI Overviews, how to measure your AI visibility, and the best AI visibility tools (track each engine separately).