Share of AI voice, measured against your named rivals

You want to know whether assistants name you or name them. The answer is usually both, in unequal proportion: in the category benchmarks I have run, the leading brand takes 35 to 55 percent of all brand mentions on its own category questions, and the client paying for the benchmark is rarely the leader. This engagement produces that split as a measured number, for you and for three to five rivals you name, on the questions your buyers actually type. The question this answers: Not whether you are visible, but who gets listed when a buyer asks for options, in what order, and on the strength of which sources. Those are three separate findings, and teams act on each of them differently. A brand that is named late in every answer has a different problem from one that is never named at all, and a rival whose lead rests entirely on one marketplace listing is a different competitor from one quoted out of its own documentation. Method: You name the rivals and define the category boundary. I build a question frame across five intents: open discovery, shortlist building, direct comparison, objection handling and switching. Each question is asked thirty times per assistant in separate sessions, and every answer is parsed for brand entities using a disambiguation rule set that I hand-audit against a sample of two hundred answers before any number is reported. Entity precision is stated in the deliverable. If it lands under 95 percent for a brand with an awkward name, you see that figure rather than a clean-looking chart resting on bad matching. What arrives: A share-of-mention table by rival, by intent and by assistant. A co-mention matrix showing who gets named in the same breath as you and who never appears beside you, which tends to reveal how these systems have grouped the category, and it often disagrees with how your own deck groups it. A citation ledger listing the domains and specific URLs feeding each rival's presence, counted, so you can see whether a competitor's lead rests on its own documentation, a review site, a marketplace listing or one well-placed thread. An attribute map of the claims attached to each brand in the answer text. The raw JSONL. A written read of around twelve pages, and one working session. Where the comparison breaks down: Assistants sometimes decline to rank commercial products. Refusals are recorded as their own class with a rate per assistant and never dropped from the denominator, because a category where two of five assistants recommend nobody changes what any share figure means. Language matters more than buyers expect: the same question in Mandarin returns a different competitive set rather than a translated one, so an English-only benchmark will mislead you if Greater China carries revenue. And this is one measurement in time. Anything measured once can be a rollout artefact, which is why the read states what would need to repeat before you act on it. Not this: Not ongoing monitoring, which is a separate monthly engagement and priced as one. Not a teardown of competitor marketing, pricing or positioning. Not a messaging workshop. And no recommendations dressed as findings: where the evidence supports an action, it goes on one clearly separated page, and where it does not, that page is short. Before you commission: Bring a real competitor list. I have refused to add two names to a set because no buyer in that market mentions them, and padding a benchmark with brands nobody names inflates your own share by arithmetic alone. Bring an honest category boundary too. Analytics and product analytics return different rivals and different winners, and picking the wider boundary because it flatters you is a choice you will have to defend in front of your board later. Who runs it: Wei-Lin Tan, working alone, in English and Mandarin. I read every parsed answer set myself before the numbers leave my machine. Fifteen business days from the day the competitor list and category boundary are signed off. I will not start against a provisional list, because a set that changes mid-run produces a benchmark that compares to nothing, including its own later re-run.