Most GEO tools show a score. A score alone does not tell you whether the tool can prove what AI actually reads, which sources it cites, or whether its fixes are safe to ship. Use this checklist before you buy.
The seven checks
1. Evidence quality: can every finding be traced to a rendered page, JSON-LD, raw-md or FAQ anchor — with a timestamp? 2. Prompt coverage: are prompts fixed and versioned, split by market and branded vs unbranded? 3. Citation analysis: does it show owned / earned / third-party sources, not just "mentions"? 4. Market scope: does it separate overseas models from mainland-China models? 5. Safety: are changes gated by approval, snapshot and rollback? 6. Reporting: does it show N and intervals, not a single score? 7. Exit: can you export data and revoke access cleanly?
Red flags
A single "GEO score" with no evidence or sample size;
Claims of guaranteed AI rankings or citations;
Mixing overseas and mainland engines in one trend line;
Counting branded prompts as brand visibility;
No rollback plan for theme/content changes.
Where GEO-Sentry fits
GEO-Sentry is one option: it focuses on first-party page evidence, prompt-set measurement (overseas and mainland), citations/sentiment, approval-gated publishing, and an optional edge gateway. Compare it against the checklist above; do not take our word for it.
Ask whether the vendor collects from a real browser (rendered JSON-LD, FAQ anchors, raw-md) or only from source HTML. AI crawlers do not execute JavaScript; source-only audits miss the rendered surface.
1.2 Prompt methodology
Require a fixed prompt set with a version hash, a market split (overseas vs mainland-China), and branded/unbranded separation. If the vendor changes prompts between reports, trends are not comparable.
1.3 Citation attribution
Demand owned / earned / third-party / dead citation classification, plus top cited pages. "Mention rate" without citations hides whether AI trusts your pages.
1.4 Measurement integrity
Every rate should carry N and a confidence interval. Single-run screenshots are anecdotes. Repeated runs and medians are the minimum bar.
Chapter 2: Procurement and risk
2.1 Data handling
Ask where evidence is stored, retention period, whether prompts/data train public models, and how deletion works. Get it in writing (DPA).
2.2 Change safety
Theme/content changes need approval, snapshot, read-back and rollback. A tool that edits production without these is a liability.
2.3 Commercial terms
Check seat/site limits, engine coverage, overage pricing, export rights and termination. Prefer transparent scope over "unlimited" claims.