02 / GEO · Historical measurement · 4 August 2026
Banking GEO Visibility Report
When AI discusses this banking need, does the institution appear—and on what basis?
Anonymized historical baseline · Current GEO workflow and proposed human-reviewed work · Follow-up not measured
90 mentions in 128 valid answers—a 70.3% mention rate. Next: complete coverage and review recommendations and sources.
Chapter 1 / 5 · Historical evidence & method
Establish the brand’s position in valid answers
Separate attempts, valid answers, mentions and placement. Then determine what the historical baseline supports and what evidence the next measurement must restore.
A mention starts an inspection
Use historical mentions to select questions for inspection. Map every appearance back to question text, recommendation position and source evidence so a manager can see where the brand enters a decision and which answer requires business review.
Keep failed attempts in the coverage record
Separate attempts, valid answers and failures to identify coverage gaps before interpreting mentions. Display complete and partial cohorts together. Restore missing observations, then investigate content gaps so task priority rests on inspectable evidence.
Preserve the original metric meanings
The original visibility, top-three and rank metrics provide different views. Read them with the answers to produce three work lists: check appearance, inspect recommendation context and verify that the source supports the banking claim.
Keep attempts, valid answers and mentions separate.
52 Failed attempts
Bar widths use all 180 attempts. The original rounded mention rate uses valid answers: 90/128 = 70.3125%.
Original reported visibility and placement
- Visibility score48/100
- Top-three rate36%
- Average rank3.3
Original report values. The composite score’s full inputs and all answer logs were not supplied; these figures do not show a measured improvement.
Historical measurement: 4 August 2026. Monitoring, edits and retesting below are proposed work, not completed improvements.
Chapter 2 / 5 · Historical evidence & method
Connect questions, engine samples and sources
A brand can appear differently across questions and engines. A monitoring asset needs question text, version, intent, scope and an evidence page—not just an aggregate score.
Recover the panel or establish a new baseline
The available material contains 10 question records and 5 answer excerpts. Inventory recoverable wording, versions and timing, marking the gaps. A recovered complete panel supports matching; otherwise establish a versioned new panel and baseline.
Ask around a concrete banking decision
For opening, fees and cross-border use, prepare direct questions, condition comparisons and process questions. Attach market, business intent and exclusions. A reviewer can then identify omitted conditions or a source from a different jurisdiction.
Inspect coverage before interpreting expression
Read valid counts with mention rates for each engine. Inspect partial samples and acquisition failures first, then the answer’s wording. Retain the original answer, position, claim and source so the aggregate can lead to a specific review task.
Read measurement coverage and brand presence together.
Original rounded mention rates. Partial samples are 2/20 and 6/20; 0/20 has no rate. This is a historical report roster, not a list of current product integrations.
| Engine | Coverage · valid / attempted | Mentions · among valid answers |
|---|---|---|
| 智谱GLMPartial sample | 10%2/20 valid | 50% |
| 千问 | 100%20/20 valid | 85% |
| DeepSeek | 100%20/20 valid | 95% |
| MiniMax | 100%20/20 valid | 80% |
| 小米MiMo | 100%20/20 valid | 75% |
| 文心一言 | 100%20/20 valid | 50% |
| 腾讯混元 | 100%20/20 valid | 55% |
| KimiPartial sample | 30%6/20 valid | 17% |
| 豆包 | 0%0/20 valid | No valid answer |
Blue compares valid-answer coverage with all attempts; pink compares the original rounded mention rate within valid answers. The two measures have different denominators. Resolve the 2/20, 6/20 and 0/20 coverage gaps before treating engines as comparable.
Connect each measurement question to maintainable evidence
What should an applicant check before opening an account?
Evidence: official eligibility, documents, channels and exceptions.
Maintainer: account-opening business ownerWhich conditions matter when comparing account costs?
Evidence: current tariff, waivers and applicable market.
Maintainer: product and fee reviewerWhere can a user verify cross-border channels and costs?
Evidence: official transfer, conversion and cross-border service guidance.
Maintainer: relevant service owner
Original rates remain unchanged: partial samples are 2/20 and 6/20; 0/20 has no rate. Missing measurement is not evidence of poor brand performance.
Chapter 3 / 5 · Current capabilities & proposed handoff
Measure with MIA. Give the next action a clear owner.
MIA can run a baseline or repeat a measurement, compare platforms and sources, and develop sourced recommendations from an evidence gap. The execution plan below is proposed work for the customer team to review and carry out.
Measure and compare a defined question set
Define the brand, questions and competitor scope, run or repeat the measurement, and inspect its status. MIA reports visibility and platform comparisons, with sources and content gaps to investigate. Keep the panel and valid-answer denominators with the report; this historical nine-engine roster does not promise today’s integration coverage.
Turn an evidence gap into sourced research
Select a gap and carry the measurement baseline and diagnosis into Cowork research. The recommendation package brings supporting sources into the working view so the team can inspect a proposed answer and its basis. A recommendation is a handoff for review, not a completed edit or publication.
Keep approval and publication with the team
The customer team verifies facts, approves the version and carries out publication. The analyst separately initiates a measurement with the same scope. Keep publication records and measurement results distinct, and report a change only after comparable results are available.
Three priorities
- 01
Restore the missing measurement first.
Check failed and partial engine results, then repeat an equivalent question set with recorded versions and a fixed period.
- 02
Review the official answers people need.
Check eligibility, application documents and cross-border conditions on pages the bank can maintain.
- 03
Inspect the source behind the mention.
Separate recommendation position from citation quality. Record which claim each source supports before changing content.
A four-week execution plan
| Week | Deliverable | Responsible role | Definition of done |
|---|---|---|---|
| Week 1 | Recover questions and failure records; decide whether to reset the baseline | Analyst | Question, version, time and validity records are complete |
| Week 2 | Review eligibility, fee and cross-border evidence pages | Business owner | Each claim has a current official source and maintaining owner |
| Week 3 | Review and approve the FAQ and condition drafts | Editor + business reviewer | Source map, scope and approved version are inspectable |
| Week 4 | Run comparable measurement; record failures and mentions separately | Analyst + project owner | Raw answers, denominators and unmeasured items are disclosed |
The plan does not imply that edits, publication or retesting have occurred. Acceptance targets are operating standards, not growth forecasts.
Chapter 4 / 5 · Proposed work · Not yet executed
Write the answer rather than repeating the brand name
The two drafts demonstrate inspectable content depth: a direct question, the conditions to verify and the maintaining source. They require approval and are not current product terms.
Map one claim to one piece of evidence
Map each eligibility, fee or channel statement to an official passage, retaining product scope, version and review date. Give uncertain sentences to the business owner as individual verification tasks so approval can be specific.
Keep the question and its boundary together
Use the reader’s question as the FAQ heading. Explain preparation first, then the conditions and exceptions requiring official confirmation. A fee explainer follows a consistent sequence: situation, charging component, trigger, review date and confirmation route.
Make an approved, maintainable asset
Business reviewers confirm facts, editors refine expression and the approver signs off the public version. Keep the source map, publication state and review date with the handoff. After a reviewed measurement identifies a gap, the team can assign an answer update and retain the change for a later comparable measurement.
Content draft · Source check and approval required
FAQ: what should I confirm before submitting an account application?
Confirm the intended applicant and market, then check current official requirements for documents, application channels and appointments. A different status, residence or intended use may require a business-owner check. Attach the official passage, review date and maintaining role. Leave uncertain conditions marked for verification. Approval depends on the actual application; do not promise an outcome or fixed processing time in the article.
Content draft · Source check and approval required
Explain fees and cross-border use by condition
Separate holding an account, transfers, currency conversion and cross-border use. For each, identify the charges, conditions, possible waivers and channel limits to verify in the current official tariff and service explanation. Do not use one promotional offer to describe every situation or import another market’s terms. Show the review date and confirmation route so readers can verify the conditions that apply to them.
Proposed evidence map: from draft claim to review task
| Draft claim | Official evidence to verify | Responsible role | Definition of done |
|---|---|---|---|
| Eligible applicant and account scope | Current eligibility rules for the relevant market and account | Account-opening owner | Scope, exceptions and review date confirmed |
| Documents and application channel | Current document list, booking and channel instructions | Channel / operations reviewer | Documents, channel and exceptions map to source passages |
| Charging triggers and waivers | Current tariff, waiver terms and effective date | Product and fee reviewer | Each charge, trigger and exception is verifiable |
| Cross-border transfer, conversion and use | Official service channels, charges and limits | Cross-border service owner | Market, channel and charging components have valid evidence |
These drafts contain no real bank name, rate, offer or unverified eligibility requirement. Publication requires current official evidence and approval.
Chapter 5 / 5 · Proposed work · Not yet executed
Retest on a comparable basis and leave unmeasured results blank
Ask whether content is more complete, answers have changed and evidence is more traceable. Without new data, label the follow-up unmeasured; neither zero nor a target is a measured result.
Secure comparability first
Align question wording/version, market, engines and collection window, then check validity per attempt. If the complete historical panel cannot be recovered, establish a new baseline and compare from that version, keeping answers and failure logs with the results.
Accept coverage and mentions separately
Read coverage, mentions, position, condition completeness and citation quality in that order. Explain changes in the valid cohort before interpreting a rate difference. Record what changed and which questions still need a business check.
Make targets support the next decision
Give managers historical baseline, follow-up and acceptance-target columns. Each cycle records edit scope, panel version and open work to guide the next maintenance task. Actual retest records answer whether and by how much the result changed.
Baseline, unmeasured follow-up and acceptance targets
| Measure | Historical baseline | Follow-up measurement | Proposed acceptance target |
|---|---|---|---|
| Sample coverage | 128/180 valid; 52 failed | Not measured | Reviewable questions; valid and failed counts per engine |
| Mentions among valid answers | 90/128 = 70.3% | Not measured | Same denominator; no assumed uplift |
| Citation and recommendation evidence | Only selected historical answer excerpts | Not measured | Every judgment maps to an answer, position and source |
Only the historical baseline is available. There is no matched T1 or verified uplift. The targets on this page are proposed acceptance criteria; follow-up is unmeasured.