Share of voice in AI search: measuring what Gemini and ChatGPT cite
Brand-radar metrics aren't new — the inputs are. Here's how we adapted share-of-voice tracking when the answer engine started writing the snippet itself.
Share of voice was a paid-search idea before it was an organic one. The translation to organic worked because the variables were the same: impressions, positions, click-through. With AI-generated answers, the variables changed. The new question isn't 'what fraction of the top ten are you?' — it's 'when an AI summary writes the answer for your query, how often does it cite you?'
The metric we actually track
For each tracked cluster, we measure two things across Google AI Overviews, ChatGPT, and Gemini:
- Citation rate: of the queries in the cluster where an AI summary is generated, what percentage cite at least one of our brand's pages.
- Position-in-citation: when we are cited, are we the first reference, the second, or buried in a 'sources' footer that the user doesn't see.
Together these answer 'are we in the room when the answer is being written.' That's the modern SOV question. Click-through metrics are downstream; if you're not cited, no amount of CTR work matters.
What moves the citation rate
Six months of A/B work across client portfolios has produced a short list of things that reliably increase citation rate, and a longer list of things that didn't.
What moves the needle:
- Direct, answer-shaped paragraphs in the first 200 words of the page. The AI summary writes in declarative sentences; pages that contain matching declarative sentences get cited.
- Internal data and original numbers — even modest ones. AI summaries cite pages with specific figures (percentages, counts, dates) far more often than pages that paraphrase. Originality is detectable.
- Author entities. Pages whose author has a public profile that the model has previously seen — bylines in established publications, conference talks, GitHub history — get cited more often than equivalent pages without an identifiable author.
What didn't move the needle, despite confident takes online: schema markup at the page level, hyper-detailed FAQ schema, and aggressive internal linking to the target page. These changed nothing in our tests.
The instrumentation problem
AI summaries are non-deterministic. The same query, asked twice in the same hour, can return different citations. The honest measurement approach is to sample — fifteen to thirty runs per query per week — and report distributions, not point estimates. Any tool that quotes a single 'AI citation rate' without sampling is reporting noise.