Quick answer
Assess completed work, changes in AI answers and business outcomes separately, using criteria agreed before execution. Publication proves that content went live. A screenshot does not establish a stable change, and more recommendations cannot be converted directly into sales growth.
See the loop facts and website terms for Winin's public descriptions. Agree specific deliverables and acceptance criteria with candidates before buying.
Details
Review three evidence layers
This table is a procurement method, not a default contract:
| Layer | Evidence to inspect | What it supports |
|---|---|---|
| Delivery and execution | Approved fact versions, drafts, approvals, publication locations and live pages | Whether agreed work was completed |
| AI answer changes | Baseline and retest answers, question versions, interfaces, conditions and visible citations | Observed changes within the tested scope |
| Business outcomes | Separate visit, inquiry and sales records with attribution rules | Business performance within a traceable scope |
Article counts cannot replace outcome verification. Business data cannot replace publication evidence either. Where no reliable attribution chain exists, record the outcome as unknown. See vendor verification and the metrics interpretation boundary.
Freeze the review protocol first
Agree the questions and versions, language, market, models or interfaces, search settings, sampling method and observation window. Define mentions, shortlist inclusion, factual accuracy and recommendation reasons separately. Specify required records, approval owners and result reviewers. If using a comparison group, describe the design and its limits in advance.
A review can mark evidence as complete, missing or incomparable. Only report a direction of change when the evidence supports it. There is no universal pass score or uplift threshold in this method.
Handle missing evidence and changed conditions
Supplement screenshots with complete answers, prompts, times and interfaces. Verify publication before evaluating its effects. If questions, markets, languages or interfaces change, separate the affected comparison and establish an appropriate new baseline.
Completed deliverables can still be checked, while affected outcome conclusions require further evidence. Comparable testing does not imply that every model or search factor can be controlled; retain those limits in the report.
List delivery records and approval points
Ask candidates for the records and client approval points at each stage, then assess delivery against the project agreement. See Edith, Homer and Friday for Winin's public descriptions.
Common questions
Q: How should buyers define acceptance criteria for a GEO project?
A: Agree delivery evidence, answer-level measurements and business attribution separately before execution. Review exact questions, raw answers, conditions and publication records rather than relying on a single mention metric.
Q: Are published articles and screenshots enough to accept GEO work?
A: They may support part of the contracted delivery. If outcome verification is included, also require a baseline, complete answers, publication checks and comparable retests.
Q: Can more AI recommendations be counted as sales growth?
A: No. Recommendations are answer-level observations; visits, inquiries and sales require separate records and attribution rules. Other factors can affect changes occurring at the same time.