Measuring Brand Mentions in AI Responses

Classic rank tracking tells you where a page sits on a results page. It says nothing about whether ChatGPT recommends you, whether Perplexity cites your documentation, or whether a Google AI Overview names a competitor instead of you. Measuring brand mentions in AI responses is the discipline of turning those fuzzy impressions into numbers you can trend, defend in a report, and act on. This guide lays out a repeatable method that avoids the two failure modes of AI visibility work: eyeballing a single ChatGPT session, and drowning in unstructured screenshots.

Define the prompt set before you measure anything

Your measurement is only as good as your prompts. Build a set of 30 to 50 questions that reflect how buyers actually ask in your category: unbranded demand ("best invoicing app for freelancers"), competitor comparisons ("alternatives to a named rival"), and problem framed queries ("how to reduce churn in a SaaS trial"). Freeze that set and version it. The instant you change prompts, you lose the ability to compare this week to last week, which is the entire point of measurement.

Sample across engines, and repeat each prompt

Two rules separate signal from noise. First, treat each engine as a separate channel: ChatGPT, Perplexity, Gemini, and Google AI Overviews return different answers with different citation behavior, so a blended number hides more than it reveals. Second, never trust a single run. Generative answers are probabilistic, so the same prompt can name you once and skip you the next time. Run each prompt three to five times and record a rate rather than a binary. The scorecard you report is a set of rates, not a set of yes and no flags.

prompts = [
  "best project management tool for agencies",
  "top alternatives to ",
  "which invoicing app integrates with Xero",
]
engines = ["chatgpt", "perplexity", "gemini", "aio"]
runs_per_prompt = 5

for engine in engines:
  for prompt in prompts:
    for i in range(runs_per_prompt):
      answer = ask(engine, prompt)
      row = {
        "engine": engine,
        "prompt": prompt,
        "mentioned": brand_in(answer),
        "cited": domain_linked(answer),
        "rank": first_position(answer),
        "sentiment": tone(answer),
      }
      log(row)

The loop above is deliberately simple. Whether you drive it with a paid platform, a headless browser, or an API, the structure is identical: iterate engines and prompts, capture the raw answer, then classify it into a row you can aggregate.

Classify every answer three ways

For each captured answer, extract three facts. Was your brand mentioned anywhere in the text? Was your domain cited or linked as a source? And what was the sentiment of the framing when you were named? Keep mention and citation strictly separate, because a model can praise you by name while linking a competitor as its source, and that gap is exactly the thing you want to close. For sentiment, a coarse positive, neutral, or negative label is enough at first, applied consistently.

The five metrics worth reporting

Roll the classified rows up into a small, stable scorecard. Five metrics cover almost every decision a stakeholder needs to make.

Share of voice is the one executives grasp fastest, because it frames the result competitively: of every brand named in answers to your category questions, what fraction were you. First position rate is the sleeper metric, since being named first correlates strongly with what a reader remembers.

Tooling: build, buy, or both

You have three practical paths. Purpose built AI visibility platforms such as Profound and the AI tracking modules now shipping inside Semrush and Ahrefs automate collection and give you dashboards out of the box, which is the fastest route for a team. A custom script gives you full control of the prompt set and classification logic, and is cheap if you already have API access. Many teams run both: a platform for the executive trend line and a lightweight script to interrogate specific prompts when a number moves. Whatever you choose, store the raw answers, not just the rollups, so you can re classify later without rerunning history.

Turn measurement into action

Numbers only matter if they change what you ship. A low citation rate despite a healthy mention rate means the models know your name but are not reading your pages, so the fix is quotable, well structured, machine readable content. A weak share of voice on comparison prompts points at missing comparison and alternatives pages. Negative sentiment on a specific prompt is a content and reputation task, not a technical one. Tie each metric to an owner and a next step, or the scorecard becomes a wall decoration.

Related reading

This metric sits inside the wider practice of generative engine optimization. To go deeper, see how the disciplines relate in SEO versus AEO versus GEO, the tactics behind becoming a cited source in AI answers, and the full set of generative engine optimization guides.

Frequently asked questions