Why you need it
The client sites you deliver carry wording risk. The risk is not that you wrote something wrong — it is three things:
- When a complaint lands, the client comes back to you. The site is yours, and the copy is what you wrote or edited.
- It is hard to tell which sentence to change. Article 9(3) of the Advertising Law enumerates terms of the "national-level" and "highest-grade" kind — and a single "and similar" clause means the boundary is not in the statute at all, but in two separate enforcement guidelines.
- Changing the wrong thing is also a risk. Verbatim regulation text, objective facts and statements that decline to promise are exactly what should stay; deleting them wholesale cuts away true information too.
What it returns
- Graded risk hits — 6 levels, from "expressly enumerated prohibited terms" to "terms disparaging competitors"; different levels carry different legal consequences.
- Article-level legal basis — the verbatim provision each hit maps to, with a sha256 content fingerprint and an official link, so you can re-check it.
- Exemption calls — every hit is run through an exemption check; anything falling into the 13 exemption categories is marked "probably may be kept".
- A "do not change" list — it proactively lists content that is outside the banned scope, or should not be remediated: the part most often deleted by mistake.
- Citation-source check — do your data, statistics and survey results state their source?
- Privacy-exposure check — output is auto-masked; the report never echoes the original numbers or account IDs.
Regulations covered
All currently in force — 7 regulations, 70 verbatim articles:
| Regulation | Version / in force |
|---|---|
| Advertising Law of the PRC | 2021 amendment |
| Enforcement Guidelines on Absolute Terms in Advertising | 2023-02-25 |
| Enforcement Guidelines on Citation Content in Advertising | 2026-06-03 |
| Anti-Unfair Competition Law of the PRC | 2025 amendment, in force 2025-10-15 |
| Measures for the Administration of Internet Advertising | in force 2023-05-01 |
| Law of the PRC on the Protection of Consumer Rights and Interests | current |
| Administrative Penalty Law of the PRC | in force 2021-07-15 |
Article text is copied verbatim, never rewritten. A query returns the verbatim text plus a content fingerprint, so you can compare it against the officially published version.
How the grading works
The same word is treated completely differently depending on context — some are expressly prohibited, others merely "need a factual basis". So we do not grade by word; we grade by legal consequence:
| Level | Meaning |
|---|---|
L1 Expressly enumerated prohibited terms | terms listed directly in the body of a regulation |
L2 Same or similar meaning | other terms with the same or similar meaning as an enumerated one |
L3 Relativised subjective evaluation | not prohibited, but needs a factual basis |
L4 Absolute promises and guarantees | a deterministic guarantee about effect or outcome |
L5 National-symbol-associated terms | wording associated with national symbols or state organs |
L6 Competitor-disparaging terms | derogatory wording about competitors |
The graded term list holds 6 levels / 180 terms; the exemption list holds 13 categories. Every hit goes through an exemption check — it never just returns a bare "compliant / not compliant".
Why a "do not change" list
Telling you only where the problems are is not enough. After a complaint, the most common mistaken fix is deleting content that was already compliant — the verbatim regulation text goes, the objective facts go, and so do statements that decline to promise. The page ends up less truthful, and easier to complain about a second time.
So the tool proactively lists the 13 categories of content that are outside the banned scope, or should not be remediated, including:
- Verbatim quotations of laws, policies and standards
- statements only of service attitude, business philosophy, corporate culture or subjective aspiration
- statements only of goals or aspirations
- self-comparison within the same brand / the same company's products
- consumer notices (how to use, timing, shelf life)
- awards and titles assessed under the relevant state provisions
- factual information (sales volume, revenue, market share) under defined time / geographic conditions
- purely chronological or spatial descriptions
- negation, refusal and disclaimer contexts
The full 13 categories are available via list_exemptions. A hit that falls into an exemption is marked "probably may be kept", with the basis for that call.
Eleven tools (all read-only)
| Tool | What it does |
|---|---|
check_text | Main tool: give it a piece of copy, get a graded self-check report (hits / basis / exemptions / citation / privacy) |
list_terms | The graded term list: 180 terms across 6 levels, each labelled with the provision it rests on |
explain_term | Look up which level a single term falls into, what it rests on, and its legal consequence |
list_exemptions | The "do not change" list: the full description of the 13 exemption categories |
check_citation | Citation check only: do data, statistics and survey results state their source? |
check_privacy | Privacy check only: scans copy for numbers, account IDs and the like, output auto-masked |
list_regulations | Lists the 7 regulations covered, with their versions and article counts |
get_article | Fetch an article's verbatim text by number, with a sha256 content fingerprint and an official link |
search_articles | Keyword search across articles |
disclaimer_template | A ready-to-use disclaimer / statement-of-scope template |
self_test | Runs a live control-group self-check and reports whether the instrument is currently usable |
Tool self-check passes 13/13; the control-group self-check passes 15/15 (healthy corpus not falsely flagged, degenerate corpus detected, negative control at zero, all 7 exemption classes recognised, old-logic regression proven).
How to use it
Hosted endpoint (Streamable HTTP, public read-only, no API key):
Drop it into the MCP config of Claude Desktop / Cursor / Cline / your own agent:
{
"mcpServers": {
"savantcat-ad": {
"type": "http",
"url": "https://savantcat.cn/mcp-ad"
}
}
}Once connected, ask the agent to "run this customer-service copy through a compliance check" — it returns graded hits, article basis and exemption calls, not a one-word "risky".
Do not want to set up MCP? Use the web version instead: advertising-compliance self-check (free, content never leaves your machine) →
Free edition and team edition
| Free (this endpoint / web version) | Team edition | |
|---|---|---|
| Scan scope | a single piece of copy | whole-site batch (a set of URLs) |
| Output | graded report + article basis + exemption calls | site-wide risk map + remediation priority |
| Remediation | gives direction | gives drop-in replacement copy |
| Citation | flags which figures lack a source | traces each source, scope and validity period into a file |
| Audit trail | none | remediation ledger (what changed, when, on what basis) |
The free edition is enough to run one check and explain the problems to a client. Batch, replacement copy and the ledger are in the team edition — that part belongs inside your project, and is not suited to a public endpoint.
Honest boundaries
- The rule engine is character-level and does no semantic understanding. It does not know what you sell, or what is customary in your industry.
- The term list is illustrative. Article 2 of the Enforcement Guidelines on Absolute Terms in Advertising makes clear that "other terms with the same or similar meaning" are in scope too, so a term-list miss does not mean there is no risk. Hand the tool the full sentence and let it judge the context.
- The output is an observation, not a prescription: whether something violates the rules is determined by the enforcement authority under the law; this tool is not a substitute for legal advice.
- It does not modify or store any of your content. Read-only, stateless.
- It does not state penalty amounts, and avoids scare framing — this is both the tool's design stance and a position we hold.
Try it on a piece of copy first
Graded hits · verbatim basis · exemption calls · citation and privacy checks
Open the free self-check pageCitation and licensing
This MCP server is public and read-only; agents and researchers are welcome to call it. When citing our determinations, credit 合尘猫 SavantCat (savantcat.cn) and keep the statement of scope with them.
Other forms of the same capability: China AI Compliance MCP (filing and generative-AI duties), GEO MCP (website AI-citability audit), all MCP endpoints.