Eco-GEO: Legal-service AI citations—comparing facts, coverage and feasible effort

I. Mechanism Start Point: The reference is an intervenable channel, but the channel does not equal the result.

Google Search documentation says AI Overviews and AI Mode show relevant links and may use query fan-out to search across related subtopics and data sources [S1]. To appear as a supporting link, a page must be indexed and eligible for a search snippet; existing SEO fundamentals remain applicable and no additional technical requirement is imposed [S1]. Bing’s public-preview documentation describes total citations, average cited pages, grounding queries and page-level citation activity, while noting that these figures do not represent rank, authority or placement within an individual answer [S4].

Together these documents support a limited inference: whether a law firm’s content is cited can be observed and influenced. They cover different platforms; Bing aggregates supported AI surfaces, and grounding queries represent only a sample of citation activity [S4]. They are measurement entry points, not proof of business impact: none establishes that citations produce engagements, leads or revenue, or establishes purchasing cycles or adoption rates in China’s legal-services sector.

II. Three directions of controlled comparison prompts, and what it cannot prove

A controlled two-source RAG experiment compared candidates differing in one content factor across six models and recorded which source was cited first [S2]. In its original Table 2, finite odds ratios for evidence range from 2.09 to 46.3, with another value above 10000; finite price-factor values range from 6.26 to 36.1, with two more above 10000; finite depth-of-coverage values range from 3.98 to 1480, with another above 10000. These large values must not be omitted, the finite range must not be presented as the range of the entire row, and “above 10000” is not a precise estimate. The table has six model columns plus grouped headings; this article does not assign values to individual models from flattened headings [S2].

These are odds ratios, not citation probabilities or percentage-point gains. The experiment used two injected sources, anonymized brands and B2C product material; some fits have convergence warnings and effect sizes vary substantially across models [S2]. Evidence, price and coverage are selected here as relevant discussion prompts, not as the only effective factors. The structure row contains values below 1, but organization and consistency also have values clearly above 1; they cannot all be described as near 1 or reversed. A point estimate above 1 does not automatically establish significance, and selecting a few rows does not prove real benefits for legal services.

The experiment compared content attributes of injected sources; it did not test the effect of publication volume. Evidence, specific facts and coverage offer prompts for designing a pilot, not causal conclusions for legal services. This article proposes two hypotheses for testing, not findings established by S2. First, expanding content that contains serious factual errors may create more routes for those errors to be cited. Second, sparse evidence may increase staff-hours per additional qualifying correct citation even when more questions are covered. Section 6 and the assumptions list explain how these hypotheses could be tested.

A regulatory example also needs its period and jurisdiction attached. In its 2020 review concerning England and Wales, the Competition and Markets Authority reported increased disclosure of price, service, redress and regulatory information, but only limited effects on competition and sector outcomes, and called for further quality-information disclosure [S3]. This is a 2020 regulatory progress statement from another jurisdiction. It cannot be extrapolated to China, GEO effectiveness or AI citation behavior; it only provides background on the continuing importance of consistent factual disclosure in professional services.

III. The same people, the same window: Three approaches and real choices

For this illustration, assume roughly three people can provide 40 equivalent staff-hours in total for baseline and intervention; the intervention observation window is two weeks. Shared preliminary work takes 10 hours: 6 for baseline sampling and 4 for monitoring and review. A costs 30 hours in total; B’s base cost is 20 hours. Both totals already include those 10 hours. In the base scenario x = 0, doing both with shared preliminary work costs 10 + 20 + 10 = 40 hours and fits capacity. If extra verification x is also needed, combined implementation costs 40 + x hours and needs a separate capacity check. This article compares a pilot with one content intervention at a time to distinguish the paths’ effects. Doing both is a different intervention design; its benefit cannot be obtained by simply adding the assumed single-path increments. This restriction is an explicit experimental-design choice, not an inevitable consequence of the time budget.

Apply the conditions in order. Condition 1: if directional agreement across two rounds is below 70%, or key indicators cannot be measured, choose C. Condition 2: if e ≥ 3/25, exclude B; choose A if it fits capacity and has a positive expected increment, otherwise C. Condition 3: if e < 3/25 but Ev is below 8 or coverage gap g < 0.5, again retain only an A that fits capacity and has a positive increment, otherwise C. Condition 4: in all other cases first exclude any path costing more than 40 hours or producing no positive increment; compare unit incremental labor cost among the survivors. Choose the sole survivor, C if none survives, and C with additional measurement if costs are equal. The values 3/25, 8, 0.5 and 70% are experimental policies requiring baseline calibration, not industry standards. With 25 questions, g ≥ 0.5 means at least 13 questions lack adequate coverage.

The table turns “which comes first” into a conditional choice: if the error condition fails, prioritize A; if errors are acceptable but evidence is thin or the coverage gap is small, prioritize A; B becomes a candidate only when errors, evidence and the coverage gap all meet their conditions. A still needs sufficient capacity and a positive increment. Choosing A delays growth in question coverage and leaves some valuable long-tail questions uncovered. Choosing B delays further error reduction and evidence improvement, and new content may later need revision. C defers both paths, retaining only baseline and monitoring information, whose value is limited if the question set is unstable.

Rounding the thresholds: 10% of 25 questions is 2.5 questions, but an observed error count must be an integer. This example therefore excludes B when e ≥ 3/25 (12%), including equality; e < 3/25 is below that exclusion threshold. The illustrative evidence ratio of 3 items per 10 questions becomes 7.5 items for 25 questions and is rounded upward to at least 8. Ev counts evidence items, not questions.

4. Economics and the feasibility boundary: one-variable sensitivity

All table values are illustrative assumptions for showing formulas and decision conditions, not industry benchmarks or measurements, and cannot directly justify a budget. Here e = 2/25 (8%, < 3/25), Ev = 8 items and g = 0.6 (≥ 0.5), so B may enter the effort and increment comparison. Both paths use the same fixed 25 questions, two-week observation window and review procedure. Q counts questions whose AI answer cites this brand’s content and whose cited claim agrees with the firm’s official facts. The increment is final Q minus baseline Q. Q is a count; only a separately reported rate uses 25 as its denominator.

Table 2 varies only one input: B’s extra verification hours x, set to 0, 8 and 20. Every column fixes A’s total at 30 hours, A’s increment at 5, B’s increment at 7, e = 2/25, Ev = 8, g = 0.6 and capacity at 40 hours. All increments are comparable hypothetical expected values; one column must not silently replace them with observed counts. Changes in other parameters require separate scenarios, not a hidden change in denominator or admission conditions.

Calculate the algebraic intersection, then check feasibility. From 30 ÷ 5 = (20 + x) ÷ 7, x = 22 and both unit costs equal 6.00 hours per additional qualifying question. But B then requires 42 hours, exceeding the 40-hour capacity. This intersection is outside the feasible set and cannot trigger the equal-cost C rule in this pilot. B requires 20 + x ≤ 40, so 0 ≤ x ≤ 20. Extra work x cannot be negative.

Check the actionable boundary: at x = 0, B costs 20 ÷ 7 ≈ 2.86, below A’s 6.00. At x = 20, B costs 40 ÷ 7 ≈ 5.71, remains cheaper and exactly fills capacity, so choose B. At x = 21, B needs 41 hours and is excluded by capacity first; A still costs 30 hours for an assumed increment of 5, so choose A. B must also be excluded at x = 22 or 30; algebraic cost or equality alone cannot justify C. The action switches when x exceeds the 20-hour capacity boundary, not at the x = 22 cost intersection. These results hold only while e, Ev, g and the increment assumptions remain unchanged.

Two distinctions remain essential. First, infeasibility differs from poor value: choose C if both paths exceed capacity, neither has a positive increment, or a verifiable baseline cannot be established. B’s infeasibility does not make every option infeasible. Second, keep expected values separate from observed counts: the increments of 5 for A and 7 for B are illustrative expectations. After execution, recalculate using actual integer counts and complete actual effort. Do not round percentages before deciding, or directly rank A’s observed result against B’s expected result.

V. Measurement and Deviation Control: Each indicator must be able to be recalculated independently.

Define Q, e, Ev and g consistently. Q is a count: among 25 fixed questions, it counts those whose AI answer cites the brand and agrees with an official factual source. Its increment is endline Q minus baseline Q, measured in questions, not a proportion. Only a separately reported correct-citation rate uses Q ÷ 25. Record each platform separately; do not mix other platforms into its denominator. e is seriously incorrect questions ÷ 25. Ev counts independently verifiable evidence items for those 25 questions, with an experimental threshold of 8. g is inadequately covered questions ÷ 25. The content team collects the data, a partner or compliance reviewer checks the facts, and an independent reviewer resolves disagreements. Baseline and post-intervention observation each last two weeks.

Build the fixed question set from service scope, fees, lawyer qualifications and public performance statements, then freeze it rather than changing it to suit new content. Repeat sampling by platform, record sources and claims for each question, and retain all rounds rather than the best result. Reviewers compare questions, answers and official facts without seeing the executor’s desired outcome; an independent reviewer records how disagreements were resolved. This article does not claim statistical significance: the source table marks some estimates as significant, but effects differ and some fits have warnings [S2]. Odds ratios are not probabilities, and this pilot’s sample is not claimed to be sufficient for significance testing. The proposed schedule is two weeks of baseline, two weeks of intervention observation, and one week of ledger review; it is a pilot plan, not a research finding.

6. Opposition’s Explanation and Conditions for Rejection

Objection 1: “Fixing facts is low-return compliance work; broad AI visibility matters more for financing, so B should come first.” Test that claim conditionally: a two-week baseline must show e < 3/25, Ev ≥ 8 and g ≥ 0.5, and subsequent comparable intervention observation must show a larger qualifying correct-citation increment for B at the same complete effort. The increment cannot be inferred from the pre-intervention baseline alone. Conversely, if errors already reach the threshold that excludes B and A cannot reduce them within available capacity, consider C rather than choosing A merely because it sounds steadier.

Counterargument: “References in this industry come from a lawyer’s personal reputation and recommendations, not from official website content. Therefore, investing resources on content alone is ineffective overall.” This explanation can also be interpreted as follows: If the baseline shows that most brand references point to third-party directories or media, and the number of references on official website pages is close to zero, then resources should be shifted from content to third-party directories and verification with public facts. In this case, neither A nor B should be used as the default option. This approach aligns with S4’s methodology—grounding queries and page-level reference activities can help determine which pages are actually used and which ones are indexed but rarely referenced [S4]. However, this tool only covers the AI interface supported by Bing, and cannot represent all engines.

The strongest objection is that the extrapolation may fail. An experiment on source attributes in anonymized B2C material does not establish a decision rule for legal services. This proposal remains a pilot. To test the first hypothesis, compare changes in errors under A and B from comparable baselines and record Q and e for the same question pool. A reduction in errors after A alone does not prove that B would spread errors. To test the second, track complete effort and correct-citation increments before and after adding evidence while keeping questions, platform and window comparable. A two-week baseline establishes a starting point; intervention comparisons require later observations. Insufficient or unstable evidence leaves both hypotheses unverified.

VII. Visible gaps and next steps

The gaps are explicit. The actual baseline of AI citations and factual errors for Chinese legal services is unknown; all example values here are illustrative. Whether citations lead to engagements, consultations or customers is unknown, so no conversion coefficient is introduced. Differences and volatility across engines require repeated sampling separately by platform. Financing or IPO preparation may add constraints on public performance and fee disclosures, limiting evidence available to A. Finally, the exclusion threshold e ≥ 3/25, the Ev threshold of 8 evidence items for 25 questions, and g = 0.5 are proposed trial settings, not validated thresholds.