AEO/GEO audit methodology for large B2C organisations
An audit should connect the customer's question to the observed answer, the relevant facts and an action your team can complete. Our common framework is configured for insurance, retail, ecommerce, banking and other consumer businesses. Industry knowledge determines what must be checked and who can approve a correction.
A customer comparing insurance needs different evidence from a shopper checking a product variant or a store's pickup service. The same brand mention measure can be calculated in each case, but the meaning of an accurate, useful answer depends on the question and its applicable facts.
1. Define the decision and audience
Agree the industry, customer situation, journey stage, language, geography and brand/product context. Separate informational questions from recommendations and branded queries from unbranded discovery. A supplier shortlist is a different panel from consumer product discovery.
| Sector | Illustrative customer situation | Facts required before assessment |
|---|---|---|
| Insurance | Compare cover for a particular situation | Policy/version, audience, coverage, exclusions and process |
| Retail | Find a product at a nearby store | Product identity, location, opening hours, assortment and channel-specific availability |
| Ecommerce | Compare compatible products and delivery | SKU/variant, attributes, offer timestamp, stock, destination and returns |
| Banking | Understand suitability and the next step | Product/version, eligibility, conditions, exceptions and official process |
These are planning examples, not live findings or conditions at a client.
2. Freeze observation conditions
Record surface, mode, language, market, date, session state, search availability and repetition. Retain raw answers, displayed citations and run status. Consumer-interface and model-API observations are separate evidence channels unless equivalence is established for the purpose.
Record product, variant, store or market context where it affects the question. Capture the applicable source version or timestamp. Live stock or offer data can change between an answer and a later source visit; that mismatch must remain visible.
Failed or unsupported observations receive their own status. They do not become zero visibility or disappear silently. Repeated sampling reveals variation; it does not make a defined panel representative of all customer demand.
| Measure | Count | Denominator |
|---|---|---|
| Mention rate | Valid answers naming the defined brand | Valid runs in the defined question group |
| Recommendation rate | Answers presenting the brand/product as a relevant choice | Valid recommendation-intent runs in that group |
| Citation rate | Answers displaying a link matching the source rule | Valid runs in the defined group |
| Assessable-claim accuracy | Claims matching applicable approved facts | Correct + incorrect assessable claims |
| Claim assessment coverage | Claims classified correct or incorrect | All in-scope claims submitted for assessment |
| Run success | Successfully captured valid observations | Attempted runs |
Agree aliases, matching and exclusions before analysis. Unverifiable claims are separate from incorrect claims; non-applicable questions are separate again. Vendor-native scores retain their original definitions.
Keep industry, language, surface, intent and journey stage available. Do not compare retail stock and insurance coverage as one accuracy benchmark. Aggregation needs a disclosed purpose and weighting, rather than an average of incompatible vendor scores.
4. Check the applicable source
A citation can be present without supporting the claim. Inspect the passage, version, timestamp, product and location. Contradiction, missing conditions, out-of-date information and unavailable evidence need different follow-up.
Use the responsible product, policy, category, store-operations or ecommerce owner. A source can be correct for one variant, policy version or channel and wrong for another. A wrong answer alone does not prove that the company's page caused the error.
| Recipient | Output | Acceptance |
|---|---|---|
| Marketing/content | Topic priority and evidence-linked brief | Customer need and missing answer are clear |
| Product/policy/category | Claim, applicable fact and review status | Owner accepts or corrects the interpretation |
| Ecommerce/store/web | Reproducible feed, page or channel issue | Target, context and required outcome are specified |
| Procurement | Scope, evidence and handover | Agreed outputs can be inspected and retained |
The brief names the target, persona, evidence, structure, owner, approval and remeasurement condition. Product/offer changes and publication follow the client's process. Recommendations remain distinct from implemented changes.
6. Evaluate follow-up and outcomes separately
Reuse suitable question cells after recording approved changes. Disclose interface changes, source-data changes and missing observations. Mention, citation, website visit and conversion are different outcomes.
Insurance may track quote starts; retail may track store-page engagement or pickup journeys; ecommerce may track product visits and checkout progression; banking may track application starts. Use permitted analytics and explain attribution. Prompt-panel improvements alone do not prove additional sales, policies or applications caused by AI search.
Google's current guidance retains core search requirements for its AI features. Check accessible, indexable and useful source content before speculative techniques. This is Google-specific guidance, not a guarantee across answer engines. Google AI optimisation guidance
Our completed GINOP agent is configured around client questions, facts and reports. External outputs and adapters are validated for the assignment, with expert review.
Explore the Sample report and agree your industry pilot. Its fictional records show the evidence chain and calculations.