Due diligence
GEO vendor due diligence checklist
When choosing a GEO vendor, what can be verified matters more than what can be promised. This checklist turns the public official evaluation dimensions into questions you can ask during the scoping stage — and the answers that should make you pause.
Download the checklist (Markdown)Official framework
The first six groups correspond to the six core dimensions published by CAICT in its GEO Capability Completeness Evaluation Specification; the seventh group is our editorial addition from project delivery.
Editorial addition
This checklist is a structured summary of public information. It is not an official document and does not evaluate or endorse any vendor. Refer to the official texts published by CAICT and the China Advertising Association for authoritative requirements.
Knowledge base construction
Whether brand information is structured and whether fact sources are authoritative.
Questions to ask
- 01
Which brand materials do you collect first, and how do you turn scattered names, products and claims into one verifiable fact sheet?
- 02
Who signs off on the fact sheet, and what is the process when public content contradicts our own claims?
- 03
Can you show a redacted example of a completed brand fact sheet?
Red flags
- Cannot explain where facts come from, only promising to publish content
- Resists client fact review or treats it as a formality
Intent recognition and content generation
Whether real user question intent is parsed and matched with suitable content.
Questions to ask
- 01
Where does the question map come from — experience-based keywords, or real questions mapped to decision stages?
- 02
Who verifies generated content, and how are AI-assisted and human-owned steps separated?
- 03
How is the same material adapted per platform rather than published identically everywhere?
Red flags
- Sells volume generation without describing verification
- Promises keyword counts but avoids questions and decision stages
Distribution and channel management
Whether content reaches high-signal channels compliantly, improving the chance of retrieval and citation.
Questions to ask
- 01
Which channels — auditable media and owned properties, or bulk press-release platforms?
- 02
How are authorship, sourcing and commercial relationships labelled?
- 03
If we exclude certain channels, can you comply and leave an audit trail?
Red flags
- Relies mainly on sock-puppet account networks or bulk posting
- Avoids naming channels and only says resources are plentiful
Monitoring and effect analysis
Whether brand presence in AI platforms is tracked with transparent measurement definitions.
Questions to ask
- 01
Which platforms and question set are monitored, and is the question set locked before the project starts?
- 02
How is data collected, and where are sampling time, frequency and deduplication rules documented?
- 03
How do you handle the same question producing different answers at different times?
Red flags
- Reports a visibility uplift percentage without the measurement definition
- Uses a single screenshot or single query as proof of effect
Strategy iteration and optimisation
Whether strategy is continuously adjusted from monitoring data to form a closed loop.
Questions to ask
- 01
How long is the review cycle, and does the review output actionable conclusions rather than numbers alone?
- 02
When content is not cited as expected, what specifically happens next?
- 03
How is the baseline set — and if there is none before kickoff, how will later comparison stay meaningful?
Red flags
- Monthly data tables with no adjustment actions
- Refuses to explain performance swings, citing algorithm changes only
Security capability
Whether data security, content compliance and anti-poisoning controls are in place.
Questions to ask
- 01
How do you draw the line between corpus poisoning and legitimate optimisation? Give an example of work you declined.
- 02
How are our brand materials, accounts and backend access isolated and later revoked?
- 03
For third-party assets, people and data, how is the rights chain documented?
Red flags
- Promises a purchasable position in AI answers or guaranteed citations
- Uses bulk pseudo-original content, fake rankings or fabricated reviews
- Cannot describe data storage or permission revocation
Contract and delivery (editorial addition)
Editorial additionOur editorial addition: put the above into the contract, not just the conversation.
Questions to ask
- 01
Does the deliverable list name verifiable artefacts such as the question map, fact sheet and measurement definitions?
- 02
What liability and termination clauses apply if prohibited practices are used?
- 03
At the end, who owns the accounts, content, data and assets, and how are they handed over?
- 04
Are fees separated from guaranteed outcomes, so you are not paying for rankings you cannot control?
Red flags
- Refuses to write compliance clauses into the contract
- Substitutes verbal promises for a verifiable deliverable list
Sources
China Academy of Information and Communications Technology
2026-07-10
GEO Capability Completeness Evaluation Specification and the first assessment batch (published 2026-07-10)
The evaluation covers six core dimensions — knowledge building, intent recognition and generation, distribution and channel management, monitoring and analysis, strategy iteration, and security capability — with three rating tiers.
China Advertising Association
2026-07-08
Three GEO association standards open for public comment, including the GEO Service Specification
Establishes principles of legality, user value and reviewable transparency, and lists prohibited practices such as corpus poisoning and prompt injection, drawing the conduct baseline for GEO services.
Cyberspace Administration of China
2026-05-08
Qinglang campaign against AI application abuses (two phases, 14 problem categories)
Phase one targets seven problem categories including failure to register large models, insufficient platform safety and review capability, training-corpus security, AI data poisoning and weak implementation of synthetic-content labelling, strengthening source-level governance.
CCTV 3.15 Gala (republished by the China Joint Fact-checking Platform)
2026-03-15
Inside the AI poisoning black market: using GEO to mass-feed false information into large models
The gala exposed vendors systematically polluting AI corpora with mass-generated fake reviews, fabricated rankings and invented products — a fictional product was recommended as a standard answer within two hours.