Share of voice in AI search is the percentage of category-relevant prompts in which an engine mentions your brand, weighted by prominence in the answer. It is the closest thing the AI era has to a rank tracker, and if you are not measuring it across ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews, you are guessing at whether your category is moving toward or away from you.
What share of voice means when the results page disappears
Classic share of voice is a paid media metric: your spend as a percentage of category spend, or your impressions as a percentage of category impressions. Neither concept survives the transition to AI answer engines. There is no impression count, no auction, and no results page. There is a generated paragraph and a set of citations that changes shape every time you ask.
In AI search, share of voice becomes a per-prompt binary that you aggregate. For each prompt in your battery, either the model mentioned your brand or it did not. Sum the mentions, divide by the number of prompts, and you have raw share of voice. Weight each mention by prominence (was it the first vendor named, one of many, a footnote) and you have weighted share of voice, which is the number you actually report.
The engines you need to cover
No single engine represents "AI search". Each of the majors draws from a different source graph, applies different guardrails, and reaches a different slice of your buyers. A demand generation program that measures only ChatGPT is measuring a large but distorted sample of its market.
- ChatGPT. The default assistant for most knowledge workers. Cites more sparingly than Perplexity but influences the biggest volume of buyer conversations.
- Perplexity. Built around explicit citations. The best window into which of your pages actually get pulled into answers.
- Claude. Popular inside enterprise and technical teams. Different citation behavior and stronger guardrails on brand claims.
- Gemini. The Google surface that most closely resembles a next-generation SERP.
- Google AI Overviews. The one your CMO cares about because it sits inside classic Google search and cannibalizes SEO clicks.
If your buyer skews technical, Claude matters more than average. If your buyer is a non-technical LOB owner, ChatGPT and Google AI Overviews dominate. A serious measurement stack covers all five, because you cannot tell in advance which one will move first when the category shifts.
Building the prompt battery
The prompt battery is the single most important design decision in any AEO measurement program. It decides what "share of voice" means for your brand. A weak battery of forty brand-name prompts will produce great-looking numbers that do not correlate with pipeline. A strong battery of one hundred category prompts will produce lower numbers that predict revenue.
Structure the battery around four prompt types.
- Category prompts. "What is the best X for Y." "Who are the leaders in X." These are the prompts a first-time buyer asks. They matter most.
- Comparison prompts. "X vs Y." "Alternatives to X." These matter when a shortlist is forming.
- Feature or capability prompts. "Which X does Y." "Best X for Z workflow." These matter when the buyer has narrowed the field to two or three vendors.
- Job-to-be-done prompts. "How do I do Y." "What is the right way to Z." These do not always surface vendors, but when they do, they surface them with high intent.
Aim for a battery of 60 to 120 prompts total, weighted 40/25/20/15 across those four types. Freeze the list. Do not add or remove prompts week to week or your trend line becomes meaningless.
Sampling for signal, not noise
Ask the same LLM the same question five times in a row and you will get five different answers. Sometimes the differences are cosmetic, sometimes they are structural. A single-shot measurement of share of voice is not a measurement. It is one sample from a distribution.
The practical fix is multi-sample runs. Run each prompt five to ten times per engine per week, then treat the aggregate as your data point. Attach a confidence interval so you know when a week-over-week move is real and when it is variance. If your tool does not show you confidence intervals, it is not doing enough sampling. Our Measure module ships this by default.
Prominence weighting
Not every mention is equal. A brand named first in a three-vendor answer carries more weight than a brand mentioned in the last paragraph. A pragmatic weighting scheme uses three tiers.
- Primary (weight 1.0). Named first, or given the most detailed treatment.
- Secondary (weight 0.5). Named in the answer body but not the lead recommendation.
- Passing (weight 0.2). Mentioned in a list or citation only, with no descriptive treatment.
Weighted share of voice becomes the sum of weights divided by the number of prompts run. It is a smaller number than raw share of voice, but it correlates more tightly with pipeline in every dataset we have looked at.
Reporting a share of voice number nobody argues with
The report that actually gets used has three views: the trend, the split, and the movers.
The trend is a single line chart of weighted share of voice per week for the last twelve weeks, with confidence bands. Executives read this in three seconds.
The split is share of voice per engine and per prompt type. This is where you see that your ChatGPT number is fine but your Perplexity number is collapsing, or that you win on category prompts but lose on comparisons.
The movers is the top five prompts where your position changed most week over week, in either direction. This is where the next content brief comes from. Our RevOps AEO playbook walks through how to turn movers into a weekly action loop.
Common mistakes
- Rewriting prompts every week. Kills your trend line. Freeze the battery and only rev it quarterly.
- Measuring only one engine. Each engine draws from a different source graph. Blind spots are guaranteed.
- Ignoring prominence. Raw mention counts overstate weak mentions and understate leadership.
- Chasing daily variance. LLM outputs are noisy. Weekly cadence is the right unit of decision.
- Not linking to pipeline. Share of voice without dollar attribution never survives a CFO conversation.
What good looks like at day 60
Sixty days into a serious measurement program you should have a frozen 100-prompt battery, five-sample coverage across four to five engines, a weekly weighted share of voice trend with confidence bands, a per-engine split, and a top movers list feeding a content backlog. If any of those pieces are missing, you are not yet measuring, you are collecting screenshots.
Where to start
The fastest way to see whether your brand is being cited today is to run a small live battery against your category. Our free AEO Grader fires six category prompts at three engines and returns a share of voice score in about twenty seconds. It is a fraction of a real prompt battery, but it is enough to prove whether the pattern exists in your market.