Measurement Methodology - External

Confidential ：prepared for client use only. Not for redistribution.

How does our GEO tool Work

GEO is built to answer four questions that traditional Search Engine Optimisation (SEO) simply can't:

Does my brand appear in AI answers, and how often?
When it does appear, what's its rank, and what does the AI actually say about it?
Where are competitors beating me — in which scenarios, on which sources?
Which content and channels should I prioritise to close the gap?

1 Data Collection

Each measurement cycle constructs two types of inputs to simulate real consumer search behaviour across AI platforms.

System Prompts: these define the AI's research persona for that query. We deliberately use multiple distinct roles (advisory, purchasing, and research perspectives) so we capture how the AI behaves across different consumer mindsets and intents, rather than relying on a single framing that might bias results.

Brand Prompts: generated on a per-brand basis. We analyse your brand's digital presence to extract its category positioning, core claims, and competitive context. From this, we produce a structured set of "no-name" queries — questions that don't mention the brand directly, mirroring how a real consumer would search the category. This tests whether the AI surfaces your brand unprompted, which is a far stronger signal than asking the AI about the brand by name.

2 Natural Language Processing (NLP)

Every AI response is parsed through a structured extraction pipeline with four components:

Brand recognition — aliases and product-line names are mapped back to a single canonical brand identifier, so variations in how the AI refers to your brand aren't missed or double-counted.
Position extraction — where the AI returns a list, we parse the literal rank; where the response isn't list-formatted, we use first-mention order in the text as a proxy for prominence.
Sentiment labelling — brand description text is segmented at the sentence level and each sentence is labelled positive, neutral, or negative.
Source attribution — any Uniform Resource Locators (URLs) cited by the AI are categorised by platform type: official website, social media (Weibo, Xiaohongshu, Douyin), review/blog, video platform, third-party database, or other.

3 Scoring

The outputs above feed into five dimensions, each weighted 20%, combining into an overall AI Search Score out of 100:

Metric	The question it answers	How it is derived
Presence	Is the brand recalled at all?	Measures the share of query executions in which the AI includes the brand in its response — the baseline test of whether the brand exists in AI awareness of the category.
Ranking	Where does it appear in the list?	Measures the brand's relative position within AI-generated recommendation lists. Brands consistently named first or second score higher than those appearing further down or only occasionally.
Perception	Does AI place it in the right consumer contexts?	Measures how often the AI surfaces the brand in the consumer contexts that matter strategically to that brand — not just any mention, but mention in the right intent context.
Persuasion	Does AI endorse it, or merely list it?	Analyses the language AI uses when describing the brand — whether it actively recommends and advocates, uses neutral listing language, or includes negative qualifiers. Derived from sentence-level sentiment analysis.
Consistency	Is performance stable across AI models?	Measures variance in the brand's recall rate across the AI models included in the measurement set. A brand with strong, even presence across all models scores higher than one that dominates one model but is largely absent from others.