Share of AI voice, measured across repeated runs

One prompt asked once is not a measurement. Put the same buying question to ChatGPT five times across three days and the brand set changes often enough that any benchmark built from single screenshots is noise presented as a trend. This engagement builds a prompt set from questions your buyers actually ask, runs each one repeatedly across ChatGPT, Perplexity, Google AI Mode and AI Overviews, Gemini and Copilot, and reports citation rates with the spread attached, so a real gap can be told apart from a bad afternoon. Where the prompts come from: Your Search Console query export, your sales call recordings and your support tickets. Not a keyword tool. The questions that decide deals are long, comparative, and often contain a competitor's name, and they do not appear in volume data. We write them in English and in Brazilian Portuguese where you sell in both, because the source set a model reaches for in pt-BR is materially different, and a practice reporting only English numbers to a Brazilian company is reporting somebody else's market. Three events, counted separately: Being mentioned, being cited with a link to your URL, and being the recommendation. Most reporting in this category collapses the three, which is how a brand with a healthy mention rate finds out late that it is being mentioned as the alternative inside a paragraph recommending someone else. We keep the answer text, the cited URLs and the run conditions for every pass. The finding you probably do not want: On most engagements the pages the models cite for your category are neither yours nor your competitors'. They are a Reddit thread, a review-site category page, the documentation of an adjacent tool, and one listicle from 2019. When that is the shape of the source graph, publishing more of your own content is the wrong response, and the intervention list says so: which third-party page to get corrected, which comparison a competitor owns because they wrote the honest version of it, which of your claims is currently being attributed to somebody else. What we refuse to produce: A single visibility score. It compresses away the only parts anyone can act on, and it exists to be put on a slide. We also label sampling method rather than blur it, because API results and results from a logged-in consumer session with search and memory enabled are different populations, and quoting one while implying the other is the standard dishonesty in this field. And we do not sell a monthly dashboard. Citation patterns for a considered purchase move more slowly than a monthly cycle can honestly show. Quarterly is the interval at which a change means something. Not included: No content production, no rewriting, no outreach to the third-party sources we identify. This engagement ends at the intervention list and the evidence behind each item on it. Who this is not for: Anyone who needs a number for a board meeting next Thursday. The runs alone take two weeks, and the output that matters is a list of arguments rather than a metric. What you get: The prompt set as a file you own and can re-run without us. A per-run log with timestamp, model, mode, answer text and cited URLs. Citation and recommendation rates for you and up to five named competitors, with variance. A source graph of the domains the models draw on for your category, ranked by how often they appear and by how easily each can be influenced. Then the intervention list, ordered by what we would do first with your money, with the items we think are not worth doing marked as such and the reason given.