Monthly Brand Mention Tracking Across the Major LLMs

Ask the same Japanese prompt five times, same model, same hour, and the set of brands named changes in roughly half of the runs. Any tracker that reports one monthly figure without telling you how many runs sit behind it is reporting noise with a decimal point. We run each prompt five times per model per cycle and report a proportion with the interval attached, so you can tell a real move from the model having a different afternoon. How the measurement is built: The prompt set is frozen at the start. We change it only at quarter boundaries, and when we do we keep the old set running in parallel for a month so the series stays joinable. Prompts come from your sales calls and support tickets, not from keyword volume, because nobody types a keyword into an assistant. Every run records the model identifier, the date, whether retrieval fired, and the full response text. Japanese and English are tracked as separate series and never averaged, because they diverge, and the direction of that divergence is usually the most useful thing in the report. What we report: Share of answer with its interval. Position in the list when the answer is a list, because third place in a five-brand answer is not the same as being named alone. The competitor set the model actually reaches for, which in energy is rarely the set your team benchmarks against. Whether retrieval fired at all, per prompt, since a category that answers ungrounded is a category you influence through published sources rather than through your own site. And every factual error stated about you. The part that earns the retainer: Errors have sources, and we go and find them. A wrong founding year usually traces to a Wikidata statement nobody has looked at since it was imported. A discontinued product persists because a retailer page is still live and still indexed. A cosmetic efficacy claim the model puts in your mouth traces to an affiliate review that quoted a 2019 release loosely. Each error in the report comes with the correction path: the page to fix, the record to file, the release to publish, the retailer to email, and who on your side can do it. We then track whether the answer changes after the correction lands, which is the only way anyone learns how long these things take. What we refuse to build: A dashboard with daily movement. Daily movement in this data is sampling variance, and a chart of it will be screenshotted into a channel and argued about for a week. Alerts on single runs, for the same reason. An AI share of voice leaderboard for your category, because the denominator is invented and the number will be quoted back to you as fact for years. Who this is not for: Companies under two years old. There is not enough said about you in Japanese for the series to be anything but flat, and the money is better spent on being written about. Brands with no Japanese-language site or coverage who want Japanese tracking. Teams who want the report to be good news; roughly half of what we send is a correction list addressed to someone in your organisation who did not ask for it. What the monthly cycle includes: The runs, the written report, the correction paths, and one working session. Content production, PR placement and technical implementation are not included. We name the fix precisely and hand it to whoever owns it, including your other agencies, and we work alongside them without asking to run them. Minimum term is three cycles. One month of this data tells you nothing you can act on, and we would rather say that before you sign than after.