Frankus GEO Audits
Run evidence-based GEO audits using Lili Frankus and duwerk's public approach to user questions, content architecture, brand consistency, external signals, and AI visibility.
Based on Lili Frankus' public method.
Not affiliated with or endorsed by Lili Frankus. Each step cites the public work it comes from: 9 sources.From visibility promises to an evidence-based audit
Turns one incomplete AI answer into scoped findings, factual corrections, and a measurement plan.
AI calls us an adults-only museum. We have zero visibility with families. Publish daily, get on Wikipedia, and call us the best attraction for every age. ChatGPT will send bookings.
Desk audit: visibility remains unmeasured because the supplied answer lacks a prompt and date. Correct conflicting age and booking information first. Operations verifies policy; marketing updates the website and tourism listing. Publish a family-visits page: workshops near Leipzig for children aged eight and older, with advance booking required. Then capture dated AI answers, separating regional discovery from branded policy questions.
Illustration from the skill's worked example, not a recorded run.
What it does
When your agent should reach for this skill.
Use this skill to investigate which customer questions a brand should answer, how generative systems currently describe it, and which content or ecosystem weaknesses deserve action. It suits marketing leaders, content and SEO teams, and B2B businesses with products that require explanation. It also supports a scoped audit for other sectors when relevant customer questions and trustworthy evidence exist.
Produce a reviewable GEO-Audit with a prompt map, observed answer evidence, technical and content findings, external signal gaps, a prioritized action plan, and repeat measurement. Keep discovery visibility, branded accuracy, citations, and commercial outcomes distinct. A mention alone does not establish a recommendation, a correct classification, or a source citation.
This package translates public material into an executable audit workflow. Read the attribution boundary in references/method.md before making claims about the method. The package's scoring formulas, evidence fields, and prioritization rules are implementation conventions, not a proprietary duwerk framework. Use the public source register to check attribution and access limitations.
The method
11 steps, restated from Lili Frankus' public work. Numbers link to the sources.
1Define the audit decision
Establish brand, category, audience, market, language, target AI surfaces, competitors, and the business decision the audit must support. Identify evidence already available and missing inputs. Use references/method.md and the scope fields in templates/output-template.md.
2Establish a defensible brand description
Record what the business actually offers, whom it serves, limitations, and dated supporting facts. Compare the desired classification against owned and external descriptions. Follow references/method.md; do not improve visibility by making unsupported claims.
3Build Top-Prompts from customer evidence
Extract questions from search queries, sales, support, CRM, and relevant public communities. Map intent and context, working backward from Decision through Consideration to Awareness. Read references/method.md and label AI-generated ideas as unvalidated.
4Separate prompt cohorts
Tag every prompt by journey stage, persona, location, and branded status. Keep brand-led questions out of non-branded discovery scores. Select a manageable, representative baseline and record the sampling limits using references/method.md.
5Capture current answers and competitors
Use available AI interfaces or tracking exports. Preserve exact prompts, timestamps, surface and mode, answers, source URLs, and brand observations. If live access is unavailable, produce a desk audit with an explicit evidence gap. Follow references/method.md.
6Diagnose the pattern
Distinguish absence, wrong category, unsupported facts, weak citation presence, and competitor dominance. Link each finding to captured evidence; label suspected causes as hypotheses. Use references/method.md and examples/worked-examples.md.
7Check technical foundations
Inspect accessibility, indexing signals, HTML structure, navigation, redirects, and relevant crawler restrictions. Prioritize access failures before rewriting inaccessible pages. Follow references/method.md; technical readiness does not guarantee selection.
8Design content that can serve as a source
Map priority questions to topic clusters, appropriate pages, and concise answer passages. Add verifiable evidence, dates, responsibility, examples, and useful constraints. Apply references/method.md, using the examples to distinguish substance from promotional language.
9Audit the ecosystem
Compare brand facts across relevant profiles, listings, reviews, partner pages, editorial coverage, and communities. Choose channels from audience relevance and observed sources. Follow references/method.md; distinguish sponsored material from independent validation.
10Assign work and measure carefully
Turn findings into bounded actions with owners, dependencies, effort, and verification criteria. Define Mentions, Citations, Visibility Score, and Visibility Share before reporting them. Use references/method.md and templates/output-template.md.
11Review and schedule iteration
Complete checklists/review-checklist.md, deliver the filled template, and propose repeat checks. Compare unchanged cohorts over time, investigate answer quality, and connect leads where evidence permits. Follow references/method.md; report uncertainty and avoid guaranteed outcomes.
Worked example
How the method plays out on one case.
1: A B2B software brand with inflated discovery reporting
Situation and fictional evidence
LedgerNest offers invoice-approval software for German businesses with 20 to 100 employees. It supports approval routing and CSV accounting exports but not full accounting or payroll. Its homepage describes an all-purpose financial platform; an old partner profile calls it accounting software.
An invented audit export contains 40 valid responses from one search-enabled AI surface over four dates. Ten brand-led responses mention LedgerNest. Of 30 non-branded responses, six mention it. Two cite its domain. Four of the six mentions place it correctly in invoice approval, while two misclassify it as accounting software. On the same 30 non-branded responses, two declared competitors have 12 and nine response-level mentions.
Weak first draft
LedgerNest has 40% AI visibility and is already strong in ChatGPT. Write 20 SEO articles about accounting software and publish reviews on every big platform. Add FAQ schema so AI will recommend us. We should see higher rankings within two weeks.
Why it fails
The overall 16/40 calculation blends brand-led answers with discovery. It conceals category errors and the limited test period. The content proposal amplifies the wrong category, the external plan ignores relevance, and the result promise has no evidence. Citation and mention are conflated.
Improved audit result
Decision: Correct the category evidence before scaling content. The sampled non-branded mention rate is 6/30 = 20%, and the owned-content citation rate is 2/30 = 6.7%. The ten branded responses are a separate accuracy cohort. Visibility Share within the declared three-brand set is 6/(6 + 12 + 9) = 22.2%. These are package-default calculations for this sample, not population market share.
Finding F1: Two of six non-branded mentions misclassify the product. Evidence A12 and A25 preserve the relevant answers. The homepage and partner listing contain inconsistent category descriptions. This supports a plausible explanation, not proof that those pages caused the errors.
Proposed content replacement, based on fictional approved product facts:
LedgerNest manages invoice approvals for German teams with 20 to 100 employees. Finance staff can route invoices to approvers and export approved records as CSV for accounting. LedgerNest does not replace accounting or payroll software. Teams requiring direct ERP posting should confirm integration support before selecting it.
Prompt map:
| Stage | Prompt | Evidence origin | Branded status | | --- | --- | --- | --- | | Decision | Can LedgerNest replace our accounting software? | Fictional sales objection | Brand-led accuracy | | Consideration | Which invoice-approval tools suit a German company with 50 employees? | Fictional sales-call theme | Non-branded discovery | | Awareness | How can a small finance team reduce delays in invoice approval? | Fictional support theme | Non-branded discovery |
Action A1, P1: Product marketing owns category and limitations updates on the homepage and product page within ten working days. Partner management requests a factual correction to the existing partner profile after the owned facts are approved. Acceptance: both controlled pages agree with the fact ledger; the partner response is recorded, including refusal or delay.
Action A2, P1: Content lead prepares one invoice-approval guide connected to the product page, answering the priority Consideration question with workflow examples and a dated capability table. No invented benchmark or superiority claim is added.
Verification: Repeat the unchanged 30-prompt cohort with the same surface and mode. Report mention, citation, category accuracy, and competitors separately. Publication completion and correct facts are controllable acceptance criteria; future AI selection is an observed outcome.
Rules that changed the result
Intent-plus-context mapping and branded separation prevent false discovery confidence. 3 Consistent category evidence and cross-team ownership address the actual problem. 1 Source-worthy answers and qualified claims replace promotional copy. 24 The scoring and evidence conventions in method sections 4 and 8 make the results inspectable.
1 more worked example in examples/worked-examples.md.
What's in the package
7 files, pinned at d5668a9. Open any file on GitHub.
SKILL.mdEntry point the agent loads: when to use the skill and the procedure.6,554 bytesLICENSEApache-2.0 license.10,756 bytesreferences/method.mdThe method in full, step by step, with citations.17,726 bytesreferences/sources.mdEvery public source the method is built from.5,766 bytesexamples/worked-examples.mdWorked examples showing the method applied end to end.8,935 bytestemplates/output-template.mdThe output format the agent fills in.9,575 byteschecklists/review-checklist.mdChecks the agent runs before handing back the result.5,686 bytes
Sources
The public work this skill is built from. Read Lili Frankus' originals.
- GEO in der Praxis: Ein Erfahrungsbericht URLduwerk.de
- Sichtbarkeit ohne Suchergebnisse: Was ChatGPT SEO für Unternehmen bedeutet URLforbes.at
- Top-Prompts für die KI-Suche finden: Die passenden Nutzerfragen für Peec AI, Rankscale & Co. URLduwerk.de
- AI Search Report 2026: GEO in der B2B-Tech-Branche URLduwerk.de
- GEO-Agentur für KI-Sichtbarkeit URLduwerk.de
- GEO: Mit Content-Strategie zur Markenpräsenz in ChatGPT, Gemini & Co. URLomt.de
- LinkedIn post on GEO myths and the Mission42.ai Lunch & Learn URLde.linkedin.com
- Grounding Page | duwerk UG URLduwerk.de
- GEO Hackathon in Hamburg & Köln URLde.linkedin.com
The skill file also covers: Files in this skill. Read the full SKILL.md.
Edge wrote this skill from Lili Frankus' public writing and talks listed above. It is not affiliated with or endorsed by Lili Frankus. If something misrepresents the method, write to [email protected] and we will correct or remove it.