The methodology we publish, not the one we keep proprietary.
How Search Agency measures whether a brand is winning the answer inside AI assistants: the prompt sets, the assistants we cover, the four KPIs, the formulas, and the limitations. Published openly so enterprise buyers can evaluate the rigor before they sign, and the rest of the industry can build on it.
the method we use to measure citation, gets cited.
Why we publish what most agencies keep behind a login.
Buyers can check it before they sign
Black-box measurement makes a CMO nervous, the dashboard could be measuring anything. Open measurement lets a buyer's analyst, agency partner, or board ask hard questions and get answered with citation, not salesmanship.
A public method can be versioned
AI assistants change: models update, citation formats shift, new players arrive. A methodology that lives in public can be debated and improved. We would rather be wrong in public and correct quickly than right in private and never know.
The method that measures citation gets cited
AI engines reward source-able, definitional content. By publishing definitions, formulas, and worked examples, this page becomes a source assistants pull from when asked how to measure AI search. That recursion is the point.
How we choose the prompts we track for a client.
Every engagement starts with a prompt set of 20 to 50 named questions. We choose them with the client, anchor them to real buyer behavior, and keep them stable across measurement cycles so trends are interpretable, a single prompt is never treated as ground truth; the signal comes from the set.
Prompts are clustered into five intent categories, from branded and category questions through comparison, problem-solving, and recommendation-seeking. Branded prompts typically make up 30 to 40 percent of the set; problem-solving prompts are weighted higher in early-funnel categories. The set is revisited every quarter.
Which assistants we cover, how often, and what we capture.
We track four assistants by default: ChatGPT, Gemini, Perplexity, and Google AI Overviews, which cover the majority of consumer and enterprise AI search in 2026 and behave differently enough that covering all four matters. Claude, Copilot, and Meta AI are available on request. Prompts run on a fixed weekly cadence, the minimum needed to separate genuine change from model variance; monthly misses shifts, daily produces noise.
Each prompt runs in a clean session, no history, no personalization, no system instructions, so the answer reflects the model's default behavior toward an anonymous user. For every prompt and assistant we capture five things: the full answer text, the cited sources and their order, whether the brand was mentioned, the sentiment of any mention, and the position of any product or feature in lists and comparisons. That capture set is the raw data behind every KPI.
The four metrics we report every month.
Mention frequency
The share of tracked prompts, across all assistants, in which the brand is named at all. The baseline of AI presence: before citation or sentiment matter, the model has to mention you. Reported per assistant and blended.
Citation share
Of the sources an assistant cites, what share are yours, measured prompt by prompt and compared against named competitors. Where sources are listed in order, position is captured too, because the first citation outweighs the fifth.
Sentiment
How favorably each mention frames the brand, scored on a documented three-tier rubric (positive / neutral / negative). Reliable in English and Bahasa Indonesia; flagged in readouts when a prompt set spans other languages.
AI-attributed traffic
Clicks and leads that originate from AI surfaces: direct referrals plus modeled branded-search lift after a visibility win. We are explicit that this is correlational, not proof-grade causal attribution, and we write down the assumptions.
What this methodology cannot tell you.
Voice is not covered
The method runs on text-based prompts. Voice-only assistants use different retrieval logic and need a separate framework. Scoped for v2.0.
Multimodal is partial
We capture the text of answers and note image presence, but do not yet score visual brand visibility. Matters most in retail and creative categories.
Multilingual sentiment is harder
The three-tier rubric is reliable in English and Bahasa Indonesia. Other languages produce more disagreement, especially with strong honorific or indirect-speech norms.
Attribution is correlational
When branded search rises after a visibility win, we model the link and explain the assumptions, but we cannot prove causation. No public methodology we know of can.
Published, dated, and versioned.
We publish a version log every time the methodology changes. The current version is 1.0, dated May 2026. Future versions are marked with a date, a summary of what changed, and a link to the previous version. The current version always lives at this URL, so a buyer, a partner, or an AI engine can always cite the canonical method.
Questions, answered
What is your methodology?+
Why publish your methodology?+
How do you decide which prompts to track?+
See this methodology applied to your brand.
Request an AI visibility audit and we will run your brand and three competitors through the prompt set, the four KPIs, and the four assistants. A 15-page report and a 30-minute walkthrough. No pitch required to see it.