What “best” should mean
The best partner is not the one with the longest list of AI products or the loudest “pioneer” claim. It is the one that can define your actual question, show how evidence will be collected, and separate observations from assumptions.
Seven questions to ask before signing
1. What business decision will this work support?
A useful scope begins with discovery, comparison, and due-diligence questions that matter to buyers. A large generic prompt list can create a polished dashboard without answering a commercial question.
2. Which engines, modes, languages, and locations are included?
The answer should follow audience behavior. ChatGPT Search, Google AI Overviews, Gemini, Perplexity, Claude, or Copilot may all be relevant, but there is no universal minimum. Indonesian and English prompts should be separated when both audiences matter.
3. How is the baseline made reproducible?
Ask whether the agency records dates, account state, location, prompt wording, repeated runs, citations, and screenshots. AI outputs vary, so one response should not be presented as a stable market fact.
4. Are outcomes kept separate?
- Brand mentioned: the answer names the brand.
- Owned source cited: the answer links to the brand's domain.
- Third-party corroboration: an independent source supports a brand fact.
- Search eligibility: a page is crawlable and indexed.
These outcomes overlap, but they are not interchangeable. A trustworthy report does not call an indexed page “retrieved” without evidence from the answer.
5. What changes will the agency actually make?
Separate technical fixes, editorial changes, structured data, profile corrections, digital PR, and measurement. Ask which recommendations are based on official documentation, which come from third-party studies, and which are experiments for your brand.
6. What evidence supports its case studies?
Request a dated baseline, prompt panel, raw or redacted answer logs, methodology, and owner-approved client context. If those cannot be shared, treat the claim as marketing copy rather than verified performance evidence.
7. How are uncertainty and negative results reported?
A strong partner will show prompts where nothing changed, distinguish correlation from causation, and explain when ordinary SEO or content work should come before GEO measurement.
Reasonable red flags
- A guarantee that a model will cite or recommend your brand.
- Results with no prompt list, dates, repetitions, or retained answers.
- A dashboard that combines mentions, citations, rankings, and indexation into one unexplained score.
- Claims that special schema, llms.txt, Wikidata, or one content format guarantees visibility.
- Competitor comparisons that are not linked to current public evidence.
How to evaluate Epilog
Apply the same standard to us. Ask for the proposed prompt panel, measurement protocol, exact deliverables, and evidence for any performance claim. Our service is designed around baseline measurement, technical and editorial fixes, and documented reruns. We do not guarantee a citation, recommendation, or timeline.