Measuring your “share of answers” in AI

In SEO, you tracked your average position and your share of voice. In a world where AI answers instead of the ten links, a new metric emerges: your share of answers — how often you, rather than a competitor, get cited in AI engines' replies. Here's how to measure it without fooling yourself.

A metric for a world without ten blue links

Share of voice measured your presence in search results; share of answers measures your presence in generative replies. The principle is the same — what fraction of attention do you capture? — but the surface has changed: it's no longer a page of links, it's a prose answer citing two or three sources.

How it's measured, concretely

The method is hands-on but robust: you define a set of real buyer questions, ask them to the engines, several times, and count who gets cited. Share of answers is the frequency of your citations across the whole.

1. List the prompts: the real questions your customers ask ("best fragrance-free SPF 50 cream"…).
2. Query each engine: Claude, ChatGPT, Gemini, Perplexity, Google AI.
3. Repeat the runs: several executions per question, because the answer varies.
4. Detect citations: does your brand/product appear, and competitors'?
5. Compute the share: frequency of your citations over the total.

Why it's harder than a Google ranking

• The answer isn't deterministic: the same question can yield different citations from one run to the next.
• No clickstream: no fixed position to read off, you must interpret a prose answer.
• Engines differ: a product cited by one may be ignored by another.
• Phrasing matters: rewording the question sometimes changes the answer.

Variability

This is the main difficulty. Because generation is probabilistic, a single run proves nothing. You need multiple runs per question to get a stable frequency, not a random snapshot.

The per-engine breakdown

Aggregating all engines into a single number hides what matters. You can be heavily cited by Perplexity and absent from Gemini. A useful share of answers reads engine by engine — that's where you see where to act.

{
  "prompt": "meilleure crème solaire SPF 50 sans parfum",
  "runs": 5,
  "engines": {
    "claude":     { "cited": 4, "share": 0.80 },
    "chatgpt":    { "cited": 2, "share": 0.40 },
    "perplexity": { "cited": 5, "share": 1.00 },
    "gemini":     { "cited": 1, "share": 0.20 },
    "google_ai":  { "cited": 3, "share": 0.60 }
  }
}

On this question, you dominate Perplexity but are nearly invisible on Gemini. Without this breakdown, an average figure ("60%") would hide exactly the engine where you have work to do.

The trap: measuring on too little data

A share of answers computed on 3 questions and 1 run isn't a measurement, it's noise. For honest signaling, you need a floor: a minimum of prompts, runs and engines below which you don't show a percentage, but a plain "insufficient data". No number beats a wrong number.

What to do with the result

1. Spot your absences on the questions that truly matter.
2. Fix structure and attributes where you're missing.
3. Re-measure after a few weeks to see movement.
4. Track per engine over time, rather than a frozen global figure.

What isn't measured honestly can't be steered. And a bad number is worse than no number.