Eco-GEO: Supply-chain AI visibility from zero to one—check eligibility before comparing fact repair with coverage expansion
With fixed hours, a fixed question set and independent review, first check error rates and the net increase in qualifying citations, then compare the full costs of feasible options. In the illustrative m=20 case, E_base>0.10 excludes B. Holding other baseline inputs fixed, qA=0.60 gives A an E_A of 0.20, so choose C; qA≥0.80 gives E_A≤0.10 and ΔQ_A=8, so choose A. At m=2, reducing hB alone to≤1.625 hours or increasing pB alone to≥4/13 lets B reach the minimum increment. All numbers are illustrative; final acceptance uses new error counts, net increments and actual full hours.
Core judgment: exclude options that fail the thresholds before comparing fact repair with coverage expansion.
Condition: During a cold start period of 0-1, the channel manager has 80 hours and 8 weeks at his disposal (a indicative budget). Brand-related facts are scattered across old quotation sheets, expired certification pages, and various dealer statements. Additionally, competitors’ references can be found in the AI answers. The choices are: A) First unify the facts and fix expired and incorrect information; B) First conduct issue coverage and content expansion; C) Postpone expanding GEO. All three options share the same budget, observation window, and set of issues.
- Thresholds come before ranking, but they do not automatically favor A. In the illustrative case m=20 (E_base=0.50), B is excluded because its baseline error rate exceeds 0.10. Holding the other baseline inputs fixed, qA=0.60 gives E_A=0.20, so A also fails and the decision is C. With qA≥0.80, E_A≤0.10 and ΔQ_A=8, making A the only feasible option. Equality holds only at qA=0.80; for example, qA=0.90 gives E_A=0.05.
- The upper limit of the number of repaired objects is determined by both the number of errors and the unqualified spaces, not by man-hours alone. nA=min(floor(52/hA), m, Q_cap). High error rates increase the difficulty of meeting the A standard (more errors need to be eliminated to stay within the threshold), and they may also prevent available objects from being used by m.
- At N=40, A and B cannot both qualify. If pA=0, A has zero increment and is excluded. For pA>0, A needs m≥8/pA to reach ΔQ≥8, while B requires m≤0.10N at baseline. Thus N≥80/pA is necessary for joint feasibility; N≥200 follows only when pA=0.40. All other conditions must still pass. The example therefore mainly asks whether to select the sole feasible option or pause.
Not applicable: For enterprises with no versioned fact ledger, no independent review personnel, unstable issue extraction (N<20), or insufficient qualified space (Q_cap<8), this calculation does not apply. These enterprises should directly proceed to step C and provide additional evidence first. All numerical values in the text are hypothetical and not industry benchmarks or actual measurement results for any enterprise.
1. Mechanism assumptions and falsification methods: Evidence only reaches “qualification and relevance”; it does not reach “effect”.
Google’s search documentation states that AI Overviews and AI Mode may use “query fan-out,” initiating multiple related searches across sub-topics and multiple data sources. This can result in a broader set of links compared to traditional searches. A page must be indexed and eligible to display summaries in searches to become a supported link, with no additional technical requirements. [S1] This document describes eligibility criteria and general practices, but does not promise any brand-specific growth in citations.
This suggests a hypothesis: if a brand parameter has conflicting versions across pages, different subqueries may retrieve different versions and produce inconsistent answers to the same intent. The proposed mediator is multiple public versions of one fact. The observable test is whether answers to the fixed question set repeat outdated or incorrect versions. The explanation fails if unifying all versions does not reduce incorrect answers. Propagation across answer paths is a mechanism to test, not an established engineering fact.
A v1 preprint on arXiv conducted extensive measurements: covering 55,936 queries, six large-model-based search systems, and two traditional search systems. It reported that 37% of the domain names in its sample appeared only in the large-model search side. The domain names favored by these systems were more structured in terms of features (hierarchical HTML), with easier-to-read text, lower domain popularity, and more external links pointing to high-reputation sources. [S3]These are feature associations, not causal interventions. It cannot be inferred that “more external links pointing to high-reputation sources will increase the probability of being cited.” The check direction it suggests is: if the bottleneck of this brand lies in page structure and readability, there is limited room for further investment in error correction.
The same preprint also cited two existing studies in the literature review: one study suggested that incorrect information in the retrieved documents might spread to the generated answers (Deng et al., 2025); another study argued that users would be influenced by the number and type of citations, even if the attribution was incorrect (Miroyan et al., 2025). [S3] are existing papers cited by S3, not experimental conclusions of S3 itself. This paper does not consider “the existence of citations” as evidence of effectiveness.
Microsoft’s Copilot Search announcement says the product highlights sources, links complete sentences or paragraphs to their sources, and lists all links used to generate the answer. It also states that availability excludes China and Russia. [S5] This is a vendor feature claim, not proof of citation accuracy or conversion effectiveness. Availability of the same intervention may differ across platforms for readers in China.
The OECD policy report distinguishes routine disruptions, addressed through ordinary enterprise management, from extreme events that call for government facilitation and emergency resources. Applying ordinary tools to major disruptions can be wasteful and ineffective. [S2] A public consulting commentary describes explicit trade-offs: concentrated sourcing can reduce complexity and cost, while diversification can improve resilience at greater cost and complexity. [S4] Neither source studies GEO. Their contribution is a methodological analogy—segment decisions by conditions and include all costs—not evidence of AI citation effectiveness.
2. Caliber: Problem sets, two rounds of baseline, and differences between N and Q_cap targets
For example, consider a fixed set of 40 questions (for illustration), with each question representing a sample unit. The questions are selected by channel managers from nearly 12 months of inquiry and sales interactions for this brand, after deduplication, and then frozen in terms of platform, version, language, question type, and sampling period. Operations staff complete two rounds of baseline sampling, and final rechecks are conducted using the same procedures. Each question retains its answer and source link. Independent reviewers determine the answers based on version-specific facts, with disagreements resolved by a third party. Different platforms report separately, without combining denominators.
“Qualified references” require traceable references in both rounds, and the brand-related facts involved in the question must be correct within the agreed verification range. Questions with incorrect brand facts in any round are counted as errors and are not excluded based on severity—severity grading requires separate criteria, which are not used in this article. Error questions do not overlap with qualified questions. The number of baseline error questions is denoted as m, E_base = m/N; Q_base is the integer count of qualified questions (illustrated as 6), Q_cap = N - Q_base = 34, representing the remaining unqualified space. Here, the two quantities must be separated: N is the total number of questions in the pool, with a threshold of 20 questions; Q_cap is the potential improvement space in this round, with a threshold of 8 questions. If either condition is not met, choose C directly for (0), the action is the same, with no difference in order.
- E: Numerator = number of questions containing brand-related factual errors; Denominator = N = 40; Sampling unit = question; Collector = channel operation; Window = two consecutive rounds before intervention; Deviation control = same collector, same question format, same time period; The question set is not modified between the two rounds.
- Q is the integer count of qualifying questions. The qualifying proportion is Q/N; a proportion must not be confused with a count. An independent third-party reviewer checks the start and end of the same observation window. Retain both answers and sources. The disagreement rate is the number of disputed questions divided by the number reviewed.
- ΔQ=Q_end−Q_base is the end-minus-start change in the count of qualifying questions within the same question set and observation window, measured in questions. The change in qualifying proportion is ΔQ/N. Do not substitute the end-of-period total for the increment.
B’s eligible pool is U=Q_cap−m only if every question has undergone fact verification, every in-scope brand factual error is included in m, and erroneous questions do not overlap with qualifying questions. If these conditions fail, do not use the formula; establish the evidence first. This article promises consistent definitions and reviewable calculations, not statistical significance. No significance markers were retained in the read excerpt, so it cannot establish whether any effect is statistically significant.
3. Thresholds and sorting rules: First, eliminate those that do not meet the criteria. Then, compare them. Finally, sort them accordingly.
The pre-frozen indicative threshold (calibration method: first determine the stable level that this brand can achieve through two rounds of baseline testing, then review the results. The threshold must not be changed afterward): N≥20, Q_cap≥8, total planned hours for the project ≤80 hours, E_post≤0.10, ΔQ≥8, review disagreement rate <20%. 0.10, 8 questions, 20 questions are all test values to be calibrated, not industry parameters given by the source.
Apply the decision rules in order. (0) Choose C if the data conditions fail: N<20, Q_cap<8, no versioned fact ledger, no independent review, any unverified in-scope brand fact, or a disagreement rate≥20%. (1) Exclude any option with total hours>80 or expected ΔQ≤0 without calculating its unit cost. A requires ΔQ_A≥8 and expected E_A≤0.10. B requires E_base≤0.10, expected E_B≤0.10 and ΔQ_B≥8. The forecast E_B=m/N holds only under the illustrative assumption that B introduces and removes no errors; actual acceptance must recalculate E_post. (2) Select the sole feasible option; if neither is feasible, choose C. (3) If both are feasible, compare full unit costs H/ΔQ and prefer the lower cost. If costs tie, prefer the higher increment. If both costs and increments tie, collect further observations before deciding.
Keep two uses of equality separate. Equality at an eligibility threshold: E_post=0.10, ΔQ=8, N=20, Q_cap=8 and total hours=80 all pass. The disagreement rate must be<20%; a rate≥20% means C, consistent with the fixed criteria and step (0). Equality in ranking: when unit costs tie, choose the higher ΔQ; when both tie, gather more observations before deciding. These are different rules, not interchangeable uses of equality.
Step (3) cannot occur in this example: at N=40, B requires m≤4, while A has ΔQ_A≤m×pA≤m≤4<8. In general, exclude pA=0 first because A cannot meet the minimum increment. For pA>0, A requires m≥8/pA and B requires m≤0.10N, giving the necessary condition N≥80/pA. Only fixing pA=0.40 gives N≥200. This is not sufficient: eligible counts, hours, error rates and B’s increment must also pass.
Consider A’s feasible range with illustrative inputs pA=0.40, hA=1.6, N=40 and Q_cap=34. Available hours cap nA at floor(52/1.6)=32. For m≤32, nA=m, and ΔQ_A≥8 requires m≥20. If qA=0.80, E_A=m(1−0.80)/40≤0.10 requires m≤20, so the only feasible integer m is 20. For m>32, nA=32 and E_A=(m−25.6)/40 still fails. If qA=0.90, integers m=20 through 32 pass; for m>32, E_A=(m−28.8)/40≤0.10 requires m≤32.8, so integers 33 and 34 fail. A higher error rate therefore does not automatically favor repair: too few erroneous questions cannot meet the increment threshold, while too many cannot meet the residual-error threshold.
| Situations and Conditions of Application (including verifiable indicator definitions) | This scenario is preferred. | Opportunity costs forgone under the same budget | Conditions for triggering stop or turning around |
|---|---|---|---|
| Scenario one: m=20, E_base=0.50 (numerator=error question 20, denominator=40); N=40, Q_cap=34; hA=1.6, pA=0.40, qA=0.60 (simplified). | C Pause expansion: B was excluded first due to E_base>0.10 and its ΔQ is not calculated. A is expected to have E_A=0.20 and ΔQ_A=8, but it cannot pass the error rate threshold. | Abandoned B’s coverage work and A’s second round of fixes; the 28 hours of joint work spent cannot be recovered | If the pilot supports qA≥0.80, with m=20 and all other inputs fixed, E_A≤0.10 and A is the only feasible option. Otherwise retain C. Measured E_post>0.10 or ΔQ<8 means C. |
| Scenario 2: m=20, qA=0.80 (after pilot calibration); otherwise the same as above | A: The only feasible option. nA = min(32, 20, 34) = 20, ΔQ_A = 8, E_A = 0.10, H_A = 60, unit cost: 7.5 hours/certified increment | Surrender the 52-hour capacity available to B during the same period; surrender the efforts devoted to new coverage outside of the ledger. | Measured E_post>0.10 or ΔQ<8 → C; unit cost rises above alternative uses recorded by the team → converted to C |
| Scenario three: m=2, E_base=0.05; hB=2.0, pB=0.25 (simplified) → nB=26, ΔQ_B=6.5 | C: Both ΔQ_A=0.8 and ΔQ_B=6.5 are below the threshold of 8, so neither solution is feasible. | Abandon the GEO increment in both directions of this window; the unit costs of 39 and 12.31 are provided for completeness only, not as a basis for selection. | First, reduce hB to ≤1.625 hours or calibrate pB to ≥4/13, then reevaluate the decision for B; otherwise, keep C unchanged. |
| Scenario Four: m=2, but hB=1.625 or pB=4/13; the rest is the same as above. | B: The only viable option. nB=32 (or 26), ΔQ_B=8, H_B=80, unit cost 10 hours/certified increment | Abandoned 2 inventory error issues fixing (nA=2, ΔQ_A=0.8<8); abandoned addition of working hours to the fact ledger | Measured E_post>0.10 or ΔQ<8 means C. Recalculate E_post as the final number of erroneous questions divided by N. If B introduces new errors, stop using the forecast E_B=m/N and compare the new measured value with the threshold. |
| Scenario Five: m=0, E_base=0 (Number of incorrect questions=0, Denominator=40) | A has no erroneous questions to repair and is excluded. B still requires ΔQ_B≥8; the baseline nB=26 and ΔQ_B=6.5 therefore mean C. | Abandon the only possible override action; work hours are recorded as unused or set aside in a separate item and ledger | With pB=0.25, hB≤1.625 permits at least 32 questions, meeting the minimum increment. If hB=2.0, pB≥4/13 is required. All other conditions must also pass; otherwise choose C. |
| Scenario Six: N<20 or Q_cap<8, or no versioned ledger, no independent review, disagreement rate ≥ 20% | C and further proof: All calculation conditions in this article do not hold true. | Abandon all GEO interventions in this window; expenses are non-recoverable. | Complete the ledger, review the manpower or issue pool, and then run two more rounds of the baseline. Do not skip the baseline and proceed with immediate intervention. |
Write the rules as executable statements, corresponding to (0)(1)(2)(3) item by item, without any additional prerequisites:
- Step 0: If N<20, Q_cap<8, there is no versioned fact ledger or independent review, any in-scope brand fact remains unverified, or the review disagreement rate is≥20%, choose C and end the procedure.
- Step 1: First exclude options with total hours>80 or expected ΔQ≤0. For the rest, A is feasible if and only if ΔQ_A≥8 and expected E_A≤0.10; B is feasible if and only if E_base≤0.10, ΔQ_B≥8 and expected E_B≤0.10. Actual acceptance always uses a newly sampled E_post, not the baseline or forecast.
- Step 2: Neither A nor B feasible → C; only A feasible → A; only B feasible → B. B may fail because E_base>0.10, or because its error-rate or increment condition fails. A may fail its error-rate or increment condition. Do not add B’s baseline condition as a prerequisite for selecting the only feasible option A.
- Step 3: If both options are feasible, prefer the lower full unit cost H/ΔQ. If costs tie, prefer the higher ΔQ. If both cost and ΔQ tie, gather further observations before deciding.
For the m=20 rows qA, pA, hA and S in Table 2, E_base=0.50 excludes B at step 1. A is the sole feasible option only in columns that pass all of that row’s conditions; a column in which A fails means C. Selecting the sole feasible A at step 2 does not require B’s E_base≤0.10 condition.
4. Simplified ledger: How 80 hours can be closed in two scenarios
Shared preliminary work is required for either A or B and is included in each option’s full cost: 10 hours for the fact ledger +12 hours for the two baseline rounds and final sampling +6 hours for independent review and adjudication =28 hours, leaving 52 intervention hours. This shared work organizes evidence and measurement without changing public pages; the model assigns it no citation increment. A and B are alternative trials, not two separate allocations of 80 hours.
A: hA=1.6 hours per question, illustratively split into 0.8 for checking and 0.8 for rewriting and submission; nA=min(floor((80−S)/hA),m,Q_cap); ΔQ_A=nA×pA; E_A=(m−nA×qA)/N; H_A=S+hA×nA; 0≤pA≤qA≤1. B handles only questions whose facts are verified as correct but which do not yet have qualifying citations; hB=2.0 hours per question; U=Q_cap−m; nB=min(floor((80−S)/hB),U); ΔQ_B=nB×pB; H_B=S+hB×nB; 0≤pB≤1. Baseline shared work S=28 leaves 52 intervention hours; if S changes, recompute 80−S. Require 0≤S≤80 and positive hours per question. The illustrative forecast E_B=m/N assumes B neither removes existing errors nor adds new ones. At the end, calculate E_post from the new error count. For example, starting at m=2, one new erroneous question with none removed gives E_post=3/40=0.075, which still passes≤0.10; three new erroneous questions give 5/40=0.125 and fail. New errors do not automatically mean exceeding 10%, and the old baseline cannot establish a pass. If B must first complete A, include the repair hours and reset the shared baseline; do not use this independent-B ledger.
High error scenario (example): m=20, Q_cap=34, pA=0.40, qA=0.60, hA=1.6 → nA=min(32,20,34)=20; ΔQ_A=8; Q_end=6+8=14; E_A=(20−12)/40=0.20; H_A=28+32=60. Budget closure: 10+12+6+32=60. B is excluded in step 1 due to E_base=0.50; ΔQ_B and unit cost are not calculated. Conclusion C: A also fails because its error rate exceeds the allowed maximum. If qA is calibrated to 0.80, then E_A=(20−16)/40=0.10, ΔQ_A=8, and A becomes the only viable option; the action is set to A.
Low error scenario (example): m=2, U=32, hB=2.0, pB=0.25, pA=0.40, qA=0.60 → A: nA=min(32,2,34)=2, ΔQ_A=0.8, E_A=(2−1.2)/40=0.02 (meets error threshold but increment is insufficient), H_A=28+3.2=31.2, unit cost = 31.2/0.8=39; B: nB=min(26,32)=26, ΔQ_B=6.5, Q_end=12.5, E_B=0.05 (hypothetical, requires re-measurement), H_B=28+52=80, unit cost = 80/6.5≈12.31. Both ΔQ<8 are not feasible → C. The values 39 and 12.31 here are for completeness only; the threshold has excluded both options, so cost comparison no longer affects the decision.
Integer execution: nA and nB are integer counts of interventions. ΔQ is a model expectation and may be fractional; it is not an observed count. Final acceptance uses actual integer counts of erroneous and qualifying questions among the 40. Q_end must exceed the baseline by at least 8 (Q_base=6 → Q_end≥14). If previously qualifying questions are lost, subtract those losses: measure the net increment, not only newly qualifying questions.
The remaining working hours are accounted for using conditional rules: if a plan uses less than 52 hours of intervention capacity (for example, nA=20 uses only 32 hours), the remaining working hours are filled in according to pre-established rules by filling in missing fields in the fact ledger (each field is assumed to be 1.0 hours). If there are no missing fields to fill, they are recorded as unused and reassigned in the next budget cycle. It is not allowed to calculate costs based on all 52 hours while ignoring the remaining capacity. The budget closure sentence in Section 7 follows the same criteria as this section.
5. Table 2: How the single-variable critical values change the behavior
In Table II, only one input per row should be changed; the rest should remain unchanged. The low and high values represent demonstration ranges, not measurement intervals. The result column shows the model’s expected values and corresponding actions (A/B/C).
| Variable changes (inputs that are fixed and not listed) | Low value → Result | Baseline → Result | Critical value → Result (equal sign handling) | High values → Results | Calibration method and boundary meaning |
|---|---|---|---|---|---|
| qA; fixed pA=0.40, m=20, nA=20, N=40, hA=1.6, S=28 | 0.50 → E=(20−10)/40=0.25; C | 0.60 (Section 4 baseline) → E=0.20; C | 0.80 → E=(20−16)/40=0.10; A (equal sign passed through) | 0.90 → E=0.05; A | Estimate qA as erroneous questions corrected/erroneous questions treated. Below 0.80, A fails the error-rate threshold. B is excluded by E_base=0.50>0.10. A is the sole feasible option only in columns meeting all row conditions; otherwise C. |
| pA; fixed qA=0.80, m=20, nA=20, N=40, S=28 (under this combination, E_A is always 0.10) | 0.20 → ΔQ=4;C | 0.40 → ΔQ=8; A | 0.40 (aligned with the baseline, 8/20) → A (equal sign passed) | 0.60 → ΔQ=12; A | Estimate pA as questions becoming qualifying/questions treated for errors, with pA≤qA. B is excluded by E_base=0.50>0.10. A is the sole feasible option only in columns meeting all row conditions; otherwise C. |
| pB; fixed m=2, U=32, nB=26, hB=2.0, S=28, E_B=0.05 | 0.15 → ΔQ=3.9;C | 0.25 → ΔQ=6.5; C | 4/13≈0.3077 → ΔQ=8, Q_end=14, H_B=80, unit cost 10; B (equal sign passed) | 0.55 → ΔQ=14.3; B | Estimate pB as questions becoming qualifying/questions covered. The baseline 0.25 differs from the critical value 4/13. B passes its baseline error-rate condition because E_base=0.05≤0.10, while A fails ΔQ_A=0.8<8. B is the sole feasible option only in columns meeting all row conditions; otherwise C. |
| hA; fixed m=20, pA=0.40, qA=0.80, S=28 (ΔQ_A≥8 requires nA≥20) | 1.2 → nA=20, ΔQ=8, E=0.10, H_A=52; A | 1.6 → nA=20, H_A=60; A | 2.60 → nA=floor(52/2.6)=20, H_A=80; A (equal sign passed) | 2.61 → nA=19, ΔQ=7.6; C | Record actual repair hours per question. Above 2.60, A cannot meet the increment threshold. B is excluded by E_base=0.50>0.10. A is the sole feasible option only in columns meeting all row conditions; otherwise C. |
| hB; fixed m=2, pB=0.25, U=32, S=28 (ΔQ_B≥8 requires nB≥32) | 1.0 → nB=32, ΔQ=8, H_B=60, unit cost 7.5; B | 2.0 → nB=26, ΔQ=6.5; C | 1.625 → nB=32, ΔQ=8, H_B=80, unit cost 10; B (equal sign passed) | 3.0 → nB=17, ΔQ=4.25; C | Record actual coverage hours per question. The baseline 2.0 already exceeds the critical value, so B fails there. A fails ΔQ_A=0.8<8. B is the sole feasible option only in columns meeting all row conditions; otherwise C. |
| S (common pre-work time); fixed values: hA=1.6, m=20, pA=0.40, qA=0.80 (when nA≥20, S must be ≤48) | 20 → nA=20, H_A=52; A | 28 → nA=20, H_A=60; A | 48 → nA=floor(32/1.6)=20, H_A=80; A (equal sign passed) | 49 → nA=19, ΔQ=7.6; C | Record actual ledger, sampling and review hours. Above 48 shared hours, A cannot meet the threshold. B is excluded by E_base=0.50>0.10. A is the sole feasible option only in columns meeting all row conditions; otherwise C. |
For each item, modify only one input at a time: qA is derived from (20−20qA)/40≤0.10, resulting in qA≥0.80. Taking 0.79 gives E=4.2/40=0.105, which fails. Taking 0.80 gives 0.10, which passes. Taking 0.90 gives 0.05, which also passes. pA is derived from 20pA≥8, resulting in pA≥0.40. With 0.39, ΔQ=7.8 fails. With 0.40, it’s 8, which passes. With 0.60, it’s 12, which also passes. Under this combination, E is always 0.10 (since qA is fixed at 0.80). pB is derived from 26pB≥8, resulting in pB≥4/13≈0.3077. When equal, ΔQ_B=8, Q_end=14, H_B=80, unit cost 80/8=10, E_B=0.05 → B. With 0.30, ΔQ=7.8 fails → C. On the capacity side, also modify only one input: hA=2.60 yields nA=20, which passes. hA=2.61 yields 19, ΔQ=7.6 fails. hB=1.625 yields nB=32, which passes. hB=1.63 yields 31, ΔQ=7.75 fails. S=48 yields nA=20, H_A=80, which passes. S=49 yields nA=19, ΔQ=7.6 fails.
When combining the high and low sides, the action does indeed flip. For combination one (m=20 side), taking qA=0.50 and pA=0.20, with everything else fixed: A’s E_A=(20−10)/40=0.25>0.10, ΔQ_A=4<8. A is not feasible; B is excluded by E_base=0.50 → C. In the same scenario, setting qA to 0.90 and pA to 0.60: E_A=(20−18)/40=0.05≤0.10, ΔQ_A=12≥8 → A. For combination two (m=2 side), taking pB=0.15: ΔQ_B=3.9<8 → C; taking pB=4/13: ΔQ_B=8 → B (at this point, A is not feasible due to ΔQ_A=0.8<8). In other words, there are only two scenarios that truly change A/B/C: when the conversion rate exceeds limits (qA, pA, pB rows) causing C to become A or B; when the capacity exceeds limits (hA, hB, S rows) causing A or B to become C.
The algebraic intersection of the two expected increments in the low-error example is pB=0.8/26=2/65≈0.030769. For pB=0.02, 2/65 and 0.05, ΔQ_B is respectively 0.52, 0.8 and 1.3. All are below 8, so all three cases choose C. Use the exact fraction for equality; the rounded 0.0308 is not the exact intersection. The intersection is below the illustrative range 0.15–0.55 and has no actionable significance. The actionable low-error threshold is pB≥4/13, holding nB=26 and hB=2.0 fixed. The unit costs of 39 and 12.31 hours per additional qualifying question are shown only for completeness: both options already fail the increment threshold, so comparing costs does not change the decision.
In this N=40 example, changing the inputs listed in the table cannot make both options feasible while B still requires m≤4 and A repairs only erroneous questions with pA≤1. Both cannot reach ΔQ≥8, so there is no ranking of jointly feasible A and B. In general, pA=0 excludes A. For pA>0, a necessary condition for joint feasibility is N≥80/pA; only fixing pA=0.40 yields N≥200. This is not sufficient: check hours, error rates, eligible object counts and B’s increment separately. N≥200 alone does not establish that both options are feasible.
6. The most threatening opposing interpretations and invalidation conditions
An alternative explanation is that the brand’s citation shortfall comes mainly from readability and structure rather than factual errors. S3 includes structured HTML and readable text among its observed features, [S3] so this explanation is compatible with the cited evidence. For every nonqualifying question, record possible causes—missing facts, incorrect facts, structure/readability or lack of third-party evidence—allowing multiple causes, and report their shares across the two baseline rounds. This diagnosis adds no threshold and does not override steps (0)–(3). If A’s pilot still gives ΔQ_A<8, step (1) rejects A regardless of the labels assigned to causes. The labels guide the next pilot toward B or structural changes; they do not waive eligibility. If errors decline but net citation growth remains 0, the proposed correction-to-growth mechanism fails and expansion stops. A predominance of factual causes offers partial support for A’s mechanism in this brand’s setting, but the decision still follows (0)–(3), without a separate diagnostic cutoff.
Opposition Explanation 2: The real bottleneck is that the problem pool is too small (N<20 or Q_cap<8). In this case, neither A nor B is reliable. Any sorting is based on noise, so it should be switched to C. First, prove the evidence before discussing investment. Opposition Explanation 3: The opportunity cost of offline actions is lower. This article does not provide a conversion rate between “yuan/effective clues” and “hours/qualified references.” It also does not argue that offline methods are inherently better. If the team has separate records of effective clues for offline actions, then set separate targets and compare them separately, without mixing them with the citation increment mentioned in this article. C is implemented according to the sequential rules of steps 0, 1, and 2 in this article, rather than through cross-dimensional comparisons of unit costs.
Observations that change the illustrative decision include a pilot supporting qA and pA beyond their respective thresholds, with the same eligible counts and working-hour constraints, making A the only feasible option; or separately reducing hB or increasing pB, making B feasible instead of C. If the question pool expands, first exclude A when pA=0. For pA>0, the general necessary condition is N≥80/pA, or N≥200 when pA=0.40. All eligibility conditions for both options must still be checked before comparing full costs under section 3(3). The read excerpt retains no significance markers; S3 is a v1 preprint. Its full methods and statistical significance cannot be confirmed from this excerpt, and this does not establish that any effect is statistically insignificant.
7. Execution path, responsible persons, and budget closure
Week 1: the channel lead is responsible, with one product and one compliance reviewer signing off. Freeze the question set, fields, thresholds and ledger template. Field completion—verified fields/planned fields—is only a readiness indicator. All in-scope brand facts must be verified; any unverified question means C under section 3(0), without adding an 80% decision threshold. Weeks 1–3: one operations collector, one independent reviewer and a third adjudicator complete two baseline rounds, reporting E_base, Q_base and disagreement rates separately by platform. If the rounds lead to different sides of a threshold, or disagreement is≥20%, first calibrate the review procedure. Weeks 3–7: run a pilot of at most 10 questions within the remaining hours to calibrate pA, qA, pB, hA, hB and S. Count pilot interventions within nA or nB and sampling/review within shared hours; do not count output twice. Without usable estimates, continue calibration and do not authorize expansion. Weeks 7–8: repeat the same sampling procedure and settle ΔQ, E_post, actual H and full unit cost.
Before expanding, rerun the complete sequence in section 3 using measured results: valid data conditions, actual H≤80, E_post≤0.10 and ΔQ≥8. Select the only feasible option, or compare full unit costs if both are feasible. A cost tie favors the higher increment; a tie on both requires more observations, not automatic expansion. Diagnose failures separately. For A’s error rate, check qA and the number treated: nA×qA≥m−0.10N is the forecast condition only while the illustrative prediction model holds. Actual acceptance always uses newly sampled E_post counts. For A’s increment, check nA×pA≥8 and available hours. Only fixing m=20 and nA=20 gives qA≥0.80 and pA≥0.40. The bounds hA≤2.60 and S≤48 each depend on the other inputs fixed in their table rows. For B’s error-rate failure, examine added, removed and retained erroneous questions, repair and resample; do not apply A’s qA threshold. For B’s increment, check nB×pB≥8: pB≥4/13 assumes nB=26, while hB≤1.625 assumes pB=0.25, U=32 and S=28. Recalculate all conditions after any input changes. Exclude any option that still fails and select the sole option that passes; choose C only if neither passes. Improvement in one metric cannot offset failure of another condition for the same option.
The stopping condition is independent of the expansion condition. Check in order: (a) If the team records the unit cost for alternative uses under the same unit (hours/correctly referenced), and the selected option is higher than that, then stop and proceed to C; (b) If alternative uses cannot be converted to the same unit (for example, offline actions are only recorded in yuan/effective leads), no conversion rate will be provided in this case. This comparison record is a qualitative judgment, which will be determined in writing by the responsible person during the next budget review round. It does not participate in the automatic determination of this round.
Budget reconciliation: ledger 10 +sampling 12 +review 6 +actual intervention hours +supplementary ledger hours +unused hours =80. Supplementary ledger hours count only actual additional work beyond the original 10 hours; do not count those original hours twice. The high-error example uses 10+12+6+32=60 hours. Account for the remaining 20 under section 4: spend them on missing ledger fields if needed, illustratively at 1.0 hour per field, or record them as unused for reassignment in the next budget. If C is selected before intervention, the 28 shared hours already spent cannot be recovered and the remaining 52 hours are not GEO intervention spending. If C is selected after a pilot, retain all actual pilot costs and calculate the unused remainder; do not erase completed work. This timetable is an arrangement based on the illustrative hours, not a source finding. Insufficient capacity calls for postponement or a smaller trial, never removal of independent review.
8. Assumptions, boundaries, and non-movable parts
Time and version: The data packet was read on 2026-09-29T11:45Z (UTC; S1 11:45:18, S2 and S3 11:45:22, S4 and S5 11:45:24). This text does not convert to local dates in other time zones. S3 is an arXiv v1 preprint, and the reading window was truncated, making it impossible to confirm the complete method and significance. The “published” fields for S1, S2, S3, and S5 are empty. S4 is recorded as 2022-10-04, and its publication date is not inferred from links or titles. Platform documentation may be updated, and conclusions are based on its current version.
Industry evidence is absent: the sources provide no Chinese supply-chain buyer AI-search behavior data, no measured GEO conversion results for this brand or another company, and no interviews or client cases. The inputs m, pA, qA, pB, hA, hB, S and the thresholds 0.10, 8 questions, 20 questions and 80 hours/8 weeks are all illustrative assumptions. Calibration uses two baseline rounds and a pilot of at most 10 questions. If calibration cannot be completed, retain only the sequential decision rules and make no numerical recommendation.
An explicit assumption: The buyer’s target issues in this scenario are focused on parametric facts such as specifications, certifications, delivery timelines, MOQ, origin, and payment conditions. This comes from the scenario setup of this article; it is not evidence from sources and must be verified with a real list of procurement problems. What can be transferred is the order of calculation—first identifying the total amount and the gap from standards, then limiting work hours based on actual quantities, checking error rates and incremental thresholds, and comparing complete unit costs. If insufficient, pause and clarify where the budget goes. What cannot be transferred are any specific numbers, industry parameters, or conclusions about effectiveness. S2 and S4 only provide a method of “segmenting by conditions and listing all costs,” which does not constitute a basis for GEO investment.
A low share of factual causes calls for reassessing the mechanism and the direction of the next pilot; it adds no threshold and does not automatically select C. If errors do not decline after facts are unified, or errors decline but the net citation increment remains 0, the corresponding explanation of A’s mechanism fails. This does not automatically exclude B if B still meets every condition. Invalid data conditions mean C under step (0). Otherwise, assess each option using actual hours, E_post and the net increment; exclude any option that fails, select the sole option that passes, and select C only if neither passes. If both pass, use the cost and tie-breaking rules in step (3).
Sources and Methodology
This analysis draws on the retrieved source text below. External facts, analytical inferences and illustrative assumptions are distinguished in the article; findings are bounded by their market, sample and date.
- [S1] AI Features and Your Website | Google Search Central | Documentation | Google for Developers — Google Search Central · Retrieved 2026-09-29
- [S2] Promoting resilience and preparedness in supply chains (EN) — OECD · Retrieved 2026-09-29
- [S3] Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines — arXiv authors · Retrieved 2026-09-29
- [S4] Emphasize Resilience in Supply Chains — BCG · Published 2022-10-04 · Retrieved 2026-09-29
- [S5] Introducing Copilot Search in Bing — Microsoft Bing · Retrieved 2026-09-29
Related reading
Eco-GEO: Not a bunch of keywords: Making the supply chain brandable GEO, what really needs to be done is to let AI describe your capabilities correctly.
The sales managers of supply chain companies during this growth period are starting to face a new problem: customers first hand over supplier comparisons to AI assistants. The star
Eco-GEO: Game companies engage in brand-related GEO: Let AI repeat the correct sentence
Growth leader for financing/IPO preparation: Brand-based GEO is not about competing with AI answers in rankings. Instead, it involves unifying versions, regions, fees, ratings, and
Eco-GEO: What to do when the opponent is mentioned by AI first: A guide to branding intellectual property services and tool selection for GEO
During periods of increased competition, the fact that competitors appear first in AI responses does not mean they should take over rankings. This article explains how to turn qual
Eco-GEO: Branding for the 0-1 cold-start phase of medical aesthetics GEO: Let AI verify first before discussing citation
The first summary given by AI might be the potential customer’s first impression of medical beauty institutions. This article breaks down how to unify brand facts, settle terminolo