Measure
Agree the actual services, jurisdictions and buyer situations first. Freeze a versioned prompt panel with unbranded discovery queries, separate branded accuracy queries and planned paraphrases. Collect repeated runs across ChatGPT, Gemini, Perplexity and Google AI experiences on multiple dates. This is our proposed operating method, not a published customer result.
Diagnose
Keep three fields: observed output, possible explanation and next check. A missing mention is observable; the claim that a missing biography caused it is a hypothesis. Inspect access, indexability, relevant pages and cited sources before proposing changes.
Build Evidence
Improve accurate practice descriptions, biographies, public business details, useful explanations and internal links. Fix technical barriers. Review genuine third-party information. Make changes reviewable, name an owner, and preserve a dated change log. Legal content requires the firm’s factual and professional review.
Retest
Repeat the frozen panel using the recorded conditions. Keep unfavorable outputs, errors and answers without business recommendations. Report sample sizes and engine-specific results. Add new prompts as a new panel version rather than quietly replacing questions that perform poorly.
The full discovery path
Search activation → crawling / indexing → retrieval → reranking / context → generation / citation → user attention → click / conversion. This is a conceptual map; products differ, and most intermediate stages are not directly observable.
Search activation: establish whether the experience used search. Crawling and indexing: can the source be accessed and included? Retrieval: is relevant material found? Reranking and context: what material reaches the model? Generation and citation: what is said and linked? Attention: does a person notice it? Conversion: do visits become relevant inquiries? Improving one stage cannot guarantee the next.
What we record
For each attempt: panel version, prompt ID and exact text, engine/product/mode, timestamp and timezone, stated geography, account/session context, run number, observed search activation, full answer, named firms, citations and URLs, description accuracy, relevance and any failure. Save the raw output with an audit trail.
A mention is a business name appearing in an answer. A citation is a linked source. A recommendation is a separate qualitative classification. Prominence describes placement within that answer; it is not a universal AI rank.
Denominators and uncertainty
For each engine and panel, report successful responses, non-answers and failed attempts separately. Mention rate is responses naming the firm divided by successful responses in the specified panel, including successful responses that name no firms. Citation rate uses responses citing the firm’s domain over that same denominator. Publish raw counts alongside rates. Repeated prompts are correlated; small movements are not automatically meaningful.
A practical cadence, not a promise
Start with a baseline and a scheduled retest after approved changes have had time to be discovered. Monthly observation is a planning option, not a scientifically established optimum. Repeat sooner to investigate an error or a major platform change. Fix the schedule in advance to reduce selective reporting.
What an experiment cannot establish
Before-and-after observations alone do not isolate causality. Platform updates, competition, personalization and sampling can change results. Where feasible, retain unchanged comparison pages or prompt groups and record confounders. Do not attribute revenue to AI visibility without reliable measurement, and do not treat citations as evidence of clicks.
Research informs hypotheses, not guarantees
The GEO paper studies content interventions under specified experimental conditions. Its results do not establish a universal ranking recipe or expected gains for a law firm. Google documents that existing SEO foundations apply to its AI features. Our workflow uses these as context while testing observable outputs in each product.
Sources and scope
Platform documentation supports the technical context. Our page plans and measurement protocol are proposed methods. They do not demonstrate customer results or reveal private ranking systems.
Google: AI features and your websiteOpenAI: crawler controlsAggarwal et al.: GEO — Generative Engine Optimization (2024 revision)