Brand mentions and share of voice in AI answers
An AI answer can recommend you without linking to you. It names three tools worth considering, describes what each is good at, and attaches its citations to a comparison article written by someone else. Your analytics records nothing, and yet the recommendation happened. This is why AI visibility has two units of measurement, mentions and links, and why measuring only the one your analytics can see understates the channel.
Mentions are not links
In classic search, visibility meant a ranked link: if you appeared, you were clickable, and clicks were countable. In generated answers the two come apart. A citation is a link the engine attributes part of its answer to. A mention is your brand appearing in the answer's text, linked or not. Unlinked mentions leave no referral trail, but they do the work a recommendation does, and the visitors who eventually arrive because of them are valuable ones: in the first quarter of 2026, AI-referred retail traffic converted 42 percent higher and produced 37 percent more revenue per visit than standard search. Tracking citations without tracking mentions misses the larger half of what the answers are saying about you.
Share of voice is a prompt-set metric
Share of voice in AI answers is the share of sampled answers that name your brand, across a fixed set of prompts your buyers would plausibly ask. Every part of that definition is load-bearing. The number only means anything relative to the prompt set it was measured on, run repeatedly, across phrasings and models.
The variance data shows why a single run is worthless. Three identical ChatGPT runs leave roughly 2.2 to 2.3 percent of cited sources stable across all three, and repeated phrasing alone accounts for 35 percent of the variance in LLM brand responses. One prompt, run once, is a dice roll. A share of voice figure is only honest as an average over many runs, reported with a confidence interval, compared month over month. The full requirements are in measuring AI search visibility honestly; a vendor quoting your share of voice as a precise daily number has skipped them.
You cannot manufacture mentions
Once mentions become a metric, someone will sell you a way to inflate them: placements in listicles, seeded forum threads, review-site coverage arranged by invoice. Google's guidance on AI features addresses this directly: seeking out inauthentic mentions is not effective. That matches how the systems work. Engines ground answers in retrieved content and weigh sources; a mention pasted into a low-trust page is exactly the kind of signal these systems are built to discount. The fuller argument, and what Google recommends instead, is walked through in what Google actually says about AI search optimization.
What earns authentic mentions
The mentions that survive are the ones other people make because you were useful to cite. The reliable route is to be the primary source for facts your industry repeats: publish original numbers, first-hand tests and data nobody else has, and your name travels with the figure each time someone else quotes it. The GEO research found that quotations, statistics and cited sources lift visibility in generated answers by roughly 41, 30 and 27 percent; a brand that is the source of the statistics collects the mention wherever those numbers appear. Vupie's content work is built around this route, creating citable substance rather than arranging placements. The practical playbook is in how to get cited by AI engines.
The honest summary
Track mentions and citations as separate metrics. Define share of voice on a fixed prompt set, sample it properly, and never read a single run as a trend. Spend nothing on manufactured mentions, and put the budget into being the source others have a reason to name.