Eco-GEO: Medical Aesthetics Branded GEO: Clear Errors First, Then Compare Unit Correct Citation Incremental Cost

When a medical aesthetics brand discusses "branded GEO", the most common mistake is treating the number of AI mentions as the goal. The real decision problem is: within a limited 80-hour product team workload, should you prioritize building an AI mention and factual error monitoring system (A), or prioritize expanding citable compliant content assets (B), or postpone GEO investment (C)? The core judgment of this article is: first meet the feasibility threshold of a zero severe factual error rate, then compare the incremental unit cost of qualified correct citation question counts under the same intervention workload. This judgment changes resource allocation—not blindly pursuing exposure in AI answers, but choosing the intervention with a lower unit cost after errors are cleared. This framework does not apply to two types of situations: first, when on the target question set the brand's correct mentions are almost zero and platform citation monitoring is unstable; second, when content cannot pass review due to medical advertising regulations. In these two situations, GEO investment should be postponed.

Why "AI mentions" are not the goal: from citation counts to qualified correct citations

Google Search Console and Bing Webmaster Tools already provide AI citation monitoring capabilities. Google includes links from AI Overviews and AI Mode in Search Console's Performance report and notes that click quality may be higher [S1]. Bing's AI Performance dashboard shows total citation count, average cited page and citation queries, but explicitly states that citation count does not indicate ranking, authority, or the page's role in the answer [S5]. These tools can answer "whether the brand is mentioned", but cannot answer "whether the mention is accurate, whether it helps users make safe decisions, or whether it drives consultations." In a medical aesthetics scenario, an incorrect safety message cited by AI not only fails to generate leads but may also trigger compliance risks. Therefore, we should not use platform-reported raw citation volume as a proxy indicator for GEO effectiveness, but should instead turn to controlled repeated sampling and human evaluation to measure the "qualified correct citation question count."

Generative AI answers are stochastic. A study on arXiv sampled four categories in the German-speaking part of Switzerland and found that AI search visibility varies with prompts, time, and engines; a single measurement is unreliable, requiring multiple samples and depicting visibility as a distribution rather than a single-point result [S3]. Although this study is not about Chinese medical aesthetics, the stochasticity mechanism is universal: for the same question, an AI answer may cite official indications one time and unapproved uses the next. Therefore, when defining the "qualified correct citation question count Q", we must sample each question 5 times; a question is qualified only if at least 3 times the brand is correctly cited and there is no severe error. Q is an integer from 0 to 20, directly observable.

Using severe factual error rate as a hard threshold: feasibility under compliance constraints

China's State Administration for Market Regulation's "Medical Aesthetic Advertising Enforcement Guidelines" prohibits advertising treatment effects or making guarantee commitments about safety and efficacy, prohibits using patient names or images as proof, and prohibits using advertising spokespersons to recommend [S4]. This means that when optimizing AI citations, medical aesthetics brands cannot use patient before/after images, influencer experience officers, or effect promises to attract AI citations. Any such content, even if cited by AI, will trigger regulatory risk. Therefore, we define the "severe factual error rate E_severe" as: among the 20 target questions, the proportion of questions for which the brand's citable pages and AI answers contain serious errors in safety, efficacy, qualifications, etc. This indicator must be zero—it is a non-negotiable feasibility constraint, not an optimization objective. If E_severe>0, the plan should be excluded no matter how high Q is.

FDA information on dermal fillers' approved uses, unapproved uses, and risk information can serve as a reference for fact-checking, such as unapproved injection sites and serious risks of silicone injections [S2]. However, FDA information cannot replace Chinese approval information and cannot be used for clinical efficacy claims. When constructing citable content, brands need to cross-validate FDA facts with Chinese approval information to ensure that any content cited by AI passes compliance review.

Conditional decision framework: error rate, qualified citation count, and evidence gap

Our decision framework includes three observable variables: severe factual error rate E_severe (%), qualified correct citation question count Q (0-20), and number of citable third-party evidence items G (items). First check E_severe: if any plan has final E_severe>0, exclude it directly. If both A and B are qualified, compare the unit qualified correct citation incremental workload cost C=64/(Q1-Q0). The one with lower C takes priority. When Q0≤2 or monitoring is unstable, default to C (postpone).

The number of citable third-party evidence items G is the key to deciding whether A or B is better. G represents the number of real, compliant, verifiable third-party sources (such as regulatory documents, academic guidelines, institutional certifications) on the official website and citable platforms. Low G means the brand lacks authoritative materials that can be cited by AI; at this time, prioritizing correcting errors in existing content may be more effective, because content expansion lacks material support; high G means there are already sufficient compliant materials, and expanding content can directly increase correct citations. Therefore, we divide enterprises into four groups according to E_severe and G:

• Group 1: E_severe>0 (any error): regardless of G, A must be prioritized, because B's unrepaired errors may be excluded and bring compliance risk.
• Group 2: E_severe=0 and G<8: prioritize A, but A's marginal increment is limited and may shift to maintenance monitoring.
• Group 3: E_severe=0 and G≥8: prioritize B, because content expansion has high marginal returns.
• Group 4: Q0≤2 and multiple sampling fluctuations are large: prioritize C (postpone), first verify the use of AI search among target users.

These thresholds (G=8, Q0=2) are suggested trial values, not industry benchmarks. Calibration method: first audit the current number of citable sources G, then use a 20-question pilot to measure the relationship between Q increment and G, and finally determine the threshold.

Table 1: Available actions and real trade-offs

The following table shows the real trade-offs of the three actions under the shared 80-hour constraint. Note that in "resource intensity", A and B are both 64 hours of intervention + 16 hours of measurement, and C is 0-8 hours. The deferred work explicitly lists opportunity costs.

The value of this table is that it does not simply present three options, but selects an action under given trigger indicators. For example, if E_severe>0, A must be selected; only if E_severe=0 and G≥8 should B be considered. The stop/expand conditions provide a clear exit mechanism, avoiding empty talk of "continuous optimization".

Same-basis economic calculation: unit cost and inversion threshold under 64-hour intervention

To compare A and B, we use a shared constraint: total workload 80 hours, of which 16 hours are common measurement workload (baseline 8h + final 8h), and 64 hours are intervention workload. The fixed question set is 20 representative questions, each sampled 5 times; qualified correct citation is defined as at least 3 of 5 samples where the brand is correctly cited and there is no severe error. Q is the number of questions satisfying the condition (0-20). E_severe must be 0. Unit cost C=64/(Q1-Q0); if Q1≤Q0, the plan is unqualified.

The following calculation is hypothetical/illustrative, not an industry benchmark. We fix E_severe=0, Q0=5, intervention workload 64h, and vary the number of citable third-party evidence items G. Assume ΔQ_A is negatively correlated with G (the more G, the smaller the increment from monitoring and correction), and ΔQ_B is positively correlated with G (the more G, the larger the increment from content expansion). The specific values are shown in the table below. These relationships need to be calibrated through a pilot in actual business and cannot be directly used for budget decisions.

We verify the inversion threshold: when G=7, ΔQ_A=7, ΔQ_B=3, C_A=64/7≈9.1h, C_B=64/3≈21.3h, A better; when G=8, ΔQ_A=5, ΔQ_B=5, C_A=C_B=12.8h, equal; when G=12, ΔQ_A=3, ΔQ_B=9, C_A≈21.3h, C_B=64/9≈7.1h, B better. The inequality reverses at G=8, and the results on both sides are consistent with the table. Note: these numbers only demonstrate the decision logic and do not represent the actual situation of any brand. The actual ΔQ-G relationship must be obtained through pilot measurement.

The decision value of this calculation lies in that it provides the team with a same-basis comparison method, rather than a feeling that "monitoring is important" or "content is important". In the A/B comparison, we uniformly use the increment ΔQ as the numerator and 64 hours as the denominator, and both deduct the same 16 measurement hours, ensuring fairness. If B requires A to be completed first, B's cost must include A's prerequisite workload, but in this model A and B are mutually exclusive intervention options, so there is no prerequisite dependency.

Who should postpone: when monitoring is unstable or correct mentions are near zero

The decision framework assumes that AI search is already in the information path of target users. But this premise may not hold. The proportion of Chinese medical aesthetics users obtaining information through AI search is unknown, and the citation preferences and filtering mechanisms of AI platforms for medical aesthetics content are not publicly disclosed. If baseline measurement of 20 target questions shows Q0≤2, and multiple sampling fluctuations are large (for example, the same question yields completely different results in two samples, making it impossible to determine whether it is random noise or systematic difference), then the unit cost of continuing to invest in A or B may be extremely high, and even no observable increment can be produced. At this time, choose C (postpone), complete only minimal verification (8-hour baseline measurement), and allocate remaining workload to verified customer acquisition channels. Postponing is not giving up, but waiting for verifiable trigger conditions: for example, customer service records showing users mentioning AI search sources, or some platforms providing click data indicating that users enter the official website from AI answers.

Strongest counterargument explanation: when AI mentions cannot be attributed, reallocate budget

A counterargument explanation that would change resource allocation is: AI mentions cannot be traced to offline consultations or consumption, and the investment is not as good as verified advertising channels. If the brand cannot establish a credible attribution chain from AI mentions to official website visits, online consultations, and offline store visits, then the return on GEO investment cannot be proven. At this point, GEO should be paused and budget assigned to traditional customer acquisition. This article's response is: we do not require proving final ROI, but set an intermediate condition—if monitoring shows that AI answers on the target question set are already being used by target users (which can be verified through customer service source inquiries, some platform click data, user interviews), and there is an observable gap in the brand's correct citations on these questions, then GEO investment is conditionally valid; otherwise, the counterargument recommendation to postpone should be adopted. If for two consecutive periods the qualified correct citation question count shows no increment or severe factual error rate >0, then continuing investment should be overturned.

This counterargument explanation is powerful because it questions the causal chain between GEO and business results. Our framework does not claim that AI mentions necessarily bring leads, but only compares which plan can produce correct brand mentions at a lower cost under a given workload. If correct mentions themselves cannot be proven valuable, then the entire GEO investment should stop. This is falsifiable and is an opportunity cost that decision-makers must face.

Executable path: responsible persons, time windows, indicators, and continue/stop thresholds

The following action path assumes the team has decided to start A or B based on preliminary evidence, not to postpone. Responsible persons and time windows are estimated based on the 80-hour constraint, not research results.

• Week 0: Product owner, content owner, compliance personnel form a group, determine 20 target questions, audit the number of citable third-party evidence items G. Responsible: product owner.
• Week 1: Complete baseline measurement (Q0, E_severe, sampling stability), record baseline unit cost. Responsible: independent analyst.
• Weeks 2-4: Choose A or B according to the decision model, record workload consumption and mid-term Q changes weekly. Responsible: content owner (B) or product owner (A), with compliance personnel involved throughout.
• End of Week 4: Final measurement Q1, E_severe, calculate unit cost C. Responsible: independent analyst.
• Week 5: Review. Expansion condition: Q increment ≥5 and E_severe=0 and C<8h/question, continue and expand to new question group. Stop condition: Q increment <2 for two consecutive periods or E_severe>0, stop GEO investment and shift to traditional channel verification.

Indicator collection requirements: Q is determined manually by an independent analyst through 20 questions × 5 samples; target questions should cover dimensions such as safety, efficacy, qualifications, and price transparency of the brand's core services. E_severe is checked by compliance personnel on pages and AI answers involved in the 20 questions; severe errors include but are not limited to unapproved indications, fabricated safety data, and incorrect qualification statements. Workload is recorded in the project management system. All sampling should be completed within the same time window (e.g., 3 consecutive days) to control time drift. If the difference between two sampling results is extremely large (change in qualified question count exceeding ±3), increase sampling frequency or suspend judgment.