GEO Citation Report Methodology
The full measurement methodology behind 238LAB GEO reports.
Version
v1.1
Effective
2026.04.28
Last Review
2026.04.28
Chapter · 01
Principles.
The following five principles apply across measurement design, data validation, and report publication.
01 · Reproducibility
Running the same methodology at the same point in time with the same seed must reproduce the same results.
02 · Transparency
We disclose the sample, queries, measurement dates, seeds, engine versions, and prompts in full.
03 · Consistency
Every measurement repeats the same standard, the same conditions, and the same procedure. The measurement framework stays identical across industries.
04 · Comparability
We prioritize normalized metrics that allow comparison across industries and across points in time.
05 · Disclosure
A measurement is an observation of AI answers at a single point in time. We state the error range and limitations up front.
Chapter · 02
LLM Engines.
We measure six LLMs - ChatGPT · Claude · Perplexity · Gemini · Grok · Google AI Overview - with equal weight.
| Engine | Model | Version Snap | Weight |
|---|---|---|---|
| ChatGPT | GPT-5 mini | 2026.01 | 1.0 |
| Claude | Claude 4.7 Sonnet | 2026.02 | 1.0 |
| Perplexity | sonar | 2026.03 | 1.0 |
| Gemini | Gemini 2.5 Pro | 2026.03 | 1.0 |
| Grok | grok-4 | 2026.03 | 1.0 |
| Google AI Overview | AI Overview | 2026.03 | 1.0 |
Engine versions are pinned as monthly snapshots. When a version transitions, we run a 30-day overlapping measurement window before and after the change.
Chapter · 03
Query Design.
Each measurement report (Issue) starts by designing industry-specific measurement items and response conditions. Six query types - recommendation, ranking, price, review, expertise, and safety - are composed in equal proportion and measured across gender and age-group conditions under identical repetition rules.
Example measurement items by industry
N
Items
2
Genders
5
Age groups
6
Types
6
Repeats
6
LLMs
The number of items N varies by industry and report. Gender, age-group, type, repetition, and LLM conditions apply identically to every report.
Dermatology report example
"Recommend a good dermatology clinic for [treatment] in Gangnam"
"Top 5 dermatology clinics for [treatment] in Gangnam"
"Where in Gangnam offers reasonable prices for [treatment]?"
"Dermatology clinics in Gangnam with good reviews for [treatment]"
"Dermatology clinics in Gangnam known for effective / specialized [treatment]"
"Where in Gangnam is safe for [treatment] without side effects?"
Measured at publication
Each report is measured once at publication, and that data is preserved as a permanent snapshot. A published report is never re-measured; the next report starts with a new industry and a new measurement set.
Chapter · 04
Response Conditions.
Even for the same question, the brands an LLM suggests can vary with response conditions. 238LAB collects responses under separate gender and age-group conditions and compares brand citation rates across them.
Gender conditions
We separate male and female conditions and check differences in response rates.
Age-group conditions
We check differences in response rates across the key consumer age groups.
Cross conditions
We analyze brand citation rates and per-LLM variance for each gender × age-group combination.
Chapter · 05
Citation Universe.
238LAB does not track a predefined list of brands. We define the citation universe as the brands, institutions, and companies that actually appear in LLM responses.
01
Input · Pre-designed
Measurement conditions
We predefine N measurement items × 60 conditions (6 types × 5 age groups × 2 genders).
02
Run · Measurement
LLM response collection
We repeatedly apply the same query conditions to 6 LLMs and collect the responses.
03
Observed · Post-defined
Citation universe
We define the citation universe as the union of the brands, institutions, and companies that appear in the responses.
Pre-designed
Measurement conditions
- Region scoping: when an input query names a specific region, we measure only queries that state that region. The actual administrative district of entities in the responses is not part of the measurement.
- Measurement item set: N measurement items per report (defined by industry as treatments, case types, product categories, and so on)
- Condition set: 60 conditions per item (6 types × 5 age groups × 2 genders)
- Execution: 6 repeats per condition × 6 LLMs = 2,160 calls per item / N × 2,160 calls per report
Response-defined
Citation universe
- Universe = the union of every extracted brand, institution, and company name
- Size is not fixed in advance: the number of brands that appear can vary by report (from dozens to hundreds)
- No pre-exclusion: once a name appears in a response, it is included regardless of revenue, size, or awareness. “Every brand that appears in LLM responses” is the measurement target.
Chapter · 06
Measurement.
For each LLM response, we measure whether a brand appears and in what context it is mentioned, then summarize the results in the three metrics below.
| Metric | Definition | Unit |
|---|---|---|
| citation_share | The share of all responses in which the brand was cited. Shown as the citation rate in reports. | % |
| context_quality | Classifies whether the brand was mentioned in a positive, neutral, or negative context. | % |
| engine_gap | The difference in citation rates across the 6 LLMs. Shown as LLM variance in reports. | %p |
Chapter · 07
Error Range.
LLM responses can vary even for identical questions. We report repeated measurements together with standard deviations, building variability into the measurement itself.
Repeated measurement
Each query condition is measured 6 times per LLM. Reports show the mean together with the standard deviation.
Fixed execution conditions
API calls run with temperature fixed at 0.3 and a fixed seed. UI-based measurements run consecutively within the same session.
Confidence interval
With 6 repeats per condition, the 95% confidence interval is ±0.8%p. Results outside this range are re-measured.
Chapter · 08
Limits.
A measurement is an observation of AI answers at a single point in time. We disclose both what a report can explain and what it cannot.
01
Scope of a single report
Each report measures a single industry or a single region. When comparison across industries or regions is needed, we measure it as a separate report.
02
Korean and English measurement
The methodology was designed around Korean-language responses, and we also run English-language measurements for overseas markets. When the language or market differs, measurement conditions are separated before comparison.
03
Six-LLM baseline
We analyze ChatGPT, Claude, Perplexity, Gemini, Grok, and Google AI Overview. Other tools are reviewed separately in methodology updates.
04
LLM version changes
When an LLM changes its model or retrieval method, before-and-after comparison may be limited. Confirmed major changes are noted separately in the report.
05
Brands absent from responses
Brands that do not appear in LLM responses are recorded with a citation rate of 0. This is not an assessment of actual quality or market position - it only means the brand was not cited under the given measurement conditions.
06
Not a quality assessment
This measurement looks at how often a brand is cited. It does not judge a brand's actual quality, capability, or relative merit.
Need a dedicated measurement under this standard?
We scope measurements by industry, region, or brand and deliver them as custom GEO measurement reports.