How to track brand presence in AI search results and LLM outputs

Start from purchase intent, not from your brand name. Most buyers do not ask "tell me about Brand X"; they ask "what is the best bank for a small business in South Africa?" or "which car insurance is cheapest for a young driver?". Your brand either shows up in those answers or it does not.

Write ten to thirty questions covering the whole journey: open category questions, comparison questions, problem-led questions, and a few direct brand questions as a control. Keep the wording natural. If a question sounds like a marketer wrote it, a buyer never asks it.

This is the step most tracking skips, and it is the one that changes the results most. AI models tailor answers to the person asking: the same insurance question from a student and from a retiree produces different recommendations, different reasoning, and different sources.

Define the audience segments that matter to your business (life stage, income band, region, role) and put each question in the context of each segment. Also run every question with no audience context at all. That baseline is your reference point: the difference between the baseline answer and each segment answer is where the insight lives.

Your buyers are spread across ChatGPT, Gemini, Claude, Grok, Perplexity, DeepSeek, and whatever ships next quarter. Each has different training data, different web search behaviour, and different opinions. A brand can dominate on one model and be invisible on another.

Monitor at least the three or four models with the largest share of your market’s usage, and query them live rather than relying on cached or synthetic answers. Where models cite web sources, capture those too: the sources explain why a model says what it says.

Now run the full grid: every question, for every segment plus the baseline, on every model. One complete run of that grid is a wave, and it produces a matrix of real AI responses. For a modest setup (twenty questions, five segments, five models) that is already six hundred responses per wave, which is why doing this by hand in a chat window does not scale past the first afternoon.

Keep the raw responses. Summaries are useful, but when a stakeholder asks "what did it actually say?", you want the full text with the mentions highlighted.

From each response, extract every entity you care about: your brand, competitors, and any category-specific attributes (fees, reliability, service quality, whatever your research needs). Group variant spellings and abbreviations into a single clean entity so "FNB" and "First National Bank" count as one brand.

The core metrics to compute from those mentions:

• –Share of voice: the percentage of responses that mention each brand, by model and by segment.
• –Top-of-mind recall: which brand the model names first when the question is open-ended.
• –Characterisation: which attributes recur next to each brand (affordable, premium, slow, trusted).
• –Sources: which websites the models cite when they answer, and which of them mention you.

A single measurement is a snapshot; the value compounds when you repeat it. Run the same grid on a schedule (monthly is common) and compare waves: which brands gained share of voice, which segments shifted, whether a campaign or a PR event moved the answers.

Keep the questions stable between waves. If you change the questions every wave, you are measuring your questionnaire, not the market.

Everything above can be done manually, and for a one-off spot check that is a fine way to start. llmeknow exists for the version that scales: it generates audience segments and personas from a market description, runs every question through every major model in each persona’s context, extracts and deduplicates entity mentions automatically, and tracks share of voice, characterisation, and cited sources across waves.

The output is the same measurement described in this guide, produced in hours instead of weeks, with the raw responses kept for inspection.