Methodology | How AI Search Lab Tests and Iterates

Four principles

Do not treat a single screenshot as a fixed ranking

01

Fixed questions

Each round uses the same query wording and taxonomy, covering brand, non-brand, comparison, local and decision questions.

02

Recorded conditions

Save platform, model, date, region, login state and visible sources; do not merge results collected under different conditions.

03

Repeat sampling

Treat important findings as observations across multiple independent runs; a one-off appearance or disappearance is a sample, not a trend.

04

Stay inside limits

Answers shift with index, personalisation, model updates and available sources; reports do not rewrite correlation as causation.

Query design

Question sets should mirror real buyer journeys

A query set is not a keyword list. It should cover how buyers understand a category, build a shortlist, compare vendors, check risk and decide the next step.

Brand: What the brand offers, which market it serves, and whether the description is accurate.
Category: Which GEO / AI search services exist in Hong Kong and which businesses they suit.
Comparison: Methods, platforms, cost factors, fit and non-fit scenarios.
Local: Traditional Chinese and English terms, Hong Kong service areas, compliance and local sources.
Decision: Whether to start with Audit, content, technical work or ongoing monitoring.

Metric definitions

Keep mentions, citations, accuracy and traffic separate

Metric	Definition	What it is not
Brand mention rate	Share of samples where the brand name appears on relevant queries	Not a recommendation, positive tone or click
Source citation rate	Share of samples where the official site or a named source is listed	Not proof that the page alone caused the answer
Description accuracy	Whether services, region, people and limits match official facts	Not brand preference or commercial outcomes
AI referral	Sessions/conversions analytics can attribute to AI platforms	Does not cover zero-click exposure or all in-app traffic
Competitive visibility	Mention and citation mix of other brands on the same questions	Not comparable across different query sets

GEO Guide rubric v1.0

How list star ratings are produced from public scores

Each brand is scored 0–20 on five dimensions: entity clarity, citable answers, machine-readable structure, external evidence layer and AI sample performance. 88–100 is three stars, 72–87 two stars, 55–71 one star; below 55 is not listed.

Score band	Meaning	Evidence required
0–5	Missing, contradictory or unverifiable	Insufficient public evidence found
6–10	Partially present, weak or outdated coverage	Single source or few intents covered
11–15	Basics in place, clear gaps remain	Official sources verifiable; some external echo
16–18	Clear evidence on most important intents	Multi-page / multi-source consistency; sample descriptions mostly accurate
19–20	Category benchmark: deep, stable evidence	Official, external and repeat samples corroborate each other

Editors may pause an upgrade for documented reasons—for example stale high-risk data, incomplete sampling or missing commercial disclosure. Any exception must appear on the brand detail page; paid relationships do not add points.

View the list and each brand’s full scorecard →

Publication standards

What can ship as a public research conclusion

Repeatable

Publish query types, platforms, observation windows, re-run methods and scoring definitions so others can understand how results were produced.

Traceable

Third-party stats carry original links and dates; our own observations are labelled as internal samples—we do not dress others’ numbers as our research.

Bounded

State region, language, industry, sample and model limits; do not generalise sample results into universal causation for every brand.

Client project data is confidential by default. Without public authorisation we use anonymous aggregates or clearly labelled scenarios—never fabricated client names, testimonials or outcomes.

Commercial attribution

How do you tell whether GEO contributed to revenue?

The “revenue path” on brand detail pages is a hypothesis inferred from public business models—not proven revenue. Sound measurement records visibility, on-site behaviour, qualified conversion and closed deals layer by layer. Without referral or conversion evidence, AI mentions stay supporting signals only.

Visibility: Mentions, source citations and description accuracy on fixed queries.
On-site: Identifiable AI referral, brand search and direct traffic changes.
Micro-conversions: Product pages, routes, bookings, WhatsApp, comparisons or application starts.
Business results: Qualified opportunities, completed applications, bookings, deals or lower service cost.
Controls: Flag concurrent campaigns, seasonality and other media so correlation is not sold as causation.

Want to build an AI visibility baseline with the same method?

View AI Visibility Audit

FAQ about Methodology | How AI Search Lab Tests and Iterates

Do you publish raw research data?

Client and confidential data stay private. Public research may share anonymized aggregates, query definitions, logging fields, and enough method detail to interpret results.

Why use a fixed query set?

Fixed wording reduces comparison noise so teams can see whether mentions, citations, and descriptions change as work lands over time.

How many times should one question be sampled?

Enough for the research goal and budget. Important findings should be retested across independent sessions with platform, date, region, login state, and prompt logged.

Can large overseas studies represent Hong Kong?

No. English/global source mixes are directional only. Hong Kong Traditional Chinese queries need separate sampling; percentages from different studies should not be added together.

Can methodology remove AI answer volatility?

No. Fixed conditions and repeats reduce misreads; they cannot erase model, index, personalization, or time-driven change.

Want a baseline of how AI answers describe your brand?

Request a free Audit