Home → Insights → Merger review data

China Merger Review Data: SAMR Publications and Court Records

China merger review data starts with the State Administration for Market Regulation, not a search across court judgments. SAMR publishes separate routes for simplified-case notices, unconditional clearances and conditional approvals or prohibitions. A judicial corpus adds litigation context, but its documents do not form a register of notified transactions. For a legal AI product, preserve the administrative record and the court record as different evidence objects.

This guide maps those public sources and reports a small, reproducible court-text query experiment performed on October 15, 2026. It is for data teams deciding what to procure, connect and validate. It does not determine whether a transaction needs notification, whether a decision remains effective, or whether a proposed transaction can proceed. Those questions require current primary evidence and qualified legal review.

1. Find the source that matches the event

SAMR's simplified-case notice section identifies a publication category distinct from final clearance lists. Use it to discover the notice and its associated transaction description. In a database, label the object as a notice and retain its publication date. The existence of that notice alone should not cause your product to display “approved.” That status would require separate evidence of the outcome.

The unconditional-clearance section is a different route. Its lists can group transactions by a stated period. A list page and each transaction row therefore have different identities. Preserve both the publication date of the page and the period named in the list, rather than assigning the page date to every underlying event. Do not assume a weekly list is a reasoned decision for every transaction.

SAMR also maintains a conditional-approval and prohibition section. Keep the disposition and any published reasoning or obligations attached to the particular administrative object. A conditions document is not interchangeable with an unconditional-clearance row. These source categories were checked through official public search results on October 15; this exercise did not authenticate to a filing system, download a complete archive, or verify a transaction's current status.

2. Keep judicial coverage separate from merger clearance coverage

China's competition-law materials extend beyond merger review. The Supreme People's Court's 2024 interpretation on monopoly-related civil disputes defines its civil-litigation subject and discusses connections with administrative enforcement. It is a judicial interpretation, not a list of cleared concentrations. A retrieval system that treats every antitrust judgment as a merger-review decision mixes different legal and evidentiary objects before a user asks a question.

For procurement, translate that distinction into separate coverage questions. Ask the administrative-data supplier which publication categories and periods were collected. Ask the court-data supplier which query and document classes were evaluated. A statement that a corpus contains competition-related language does not answer either supplier's completeness question. The broader antitrust court-data guide addresses adjacent judicial material; this page focuses on concentration-review source selection.

Missing judicial results are also easy to misread. A transaction may have an administrative publication without a matching court dispute in the supplied corpus. A court may mention an acquisition while deciding an unrelated contract or procedure issue. Neither situation proves that the administrative source or judicial source is defective. Record the intended coverage of each track before calculating a missing-record rate.

3. What our October 15 court queries actually returned

We ran three mention searches against the current judgments index. Every query combines ik_smart and ik_max_word phrase matching with a logical OR and excludes the bench test dataset. The downloadable queries and count responses preserve the request, retrieval time and shard status. Each count request completed without failed shards.

These are overlapping record pools. We did not calculate their intersections, normalize case numbers, deduplicate documents, or manually code relevance. In particular, 10,594 does not mean that many merger-notification violations: the declaration phrase is deliberately broad and can occur in other contexts. The four title hits do not establish that China has only four relevant cases. A small corpus result is a property of the query and corpus, not a national incidence statistic.

Phrase matching here follows analyzed token positions; it is not a literal-substring audit. Combining analyzers reduces a known retrieval blind spot, but does not prove semantic accuracy or exhaustive recall. We did not sample records, calculate a relevance percentage, or measure the largest single claimant's share. The figures should help a buyer design an evaluation slice, not estimate enforcement risk, approval probability or transaction volume.

4. Model transactions, publications and court documents separately

A practical evaluation schema can start with three object tables: transaction candidates, administrative publications and judicial documents. Add explicit relations between them only when evidence supports the connection. This is a proposed data design, not a claim that the delivered court corpus already includes a complete merger-transaction graph. The distinction matters when comparing a supplier's existing fields with work that would require a separate research project.

For an administrative publication, retain the original title, issuing authority, source URL, observation time, publication category, document date where stated and any available administrative identifier. For a transaction candidate, retain the names and roles as written before adding normalized corporate identities. A buyer, target and joint-venture participant are different roles; a similar company name is not sufficient evidence that two notices describe the same transaction.

For a court document, preserve its own case number, court, date field, document classification and available text. Do not reuse a supplier's ingestion identifier as a regulator's decision number. Our general corpus does not provide reliable original-judgment source URLs for every record. When an official judicial source link is unavailable, keep that field unavailable and use the court's case-number citation as a verification coordinate, with independent checking where possible.

Relations need their own provenance. A match based on an explicit administrative decision number is different from one based on party names and a date range. Store the relation method, supporting passage or coordinate, reviewer status and uncertainty. If a target name is masked or a corporate group has changed, an unresolved relation is preferable to silently joining two unrelated events. Preserve multiple documents for one proceeding instead of deleting them as duplicate transactions.

5. Separate the clocks before drawing a timeline

Useful dates may include a transaction announcement, notice publication, administrative decision, court filing, judgment signature, website publication and your own retrieval time. They answer different questions. A public list's coverage period should remain separate from its release date. A date copied from a court-data index should retain its field definition until it has been compared with the underlying document.

This is particularly important when someone requests “cases after a rule change.” Some ingestion batches use publication or collection dates in fields that appear to be decision dates. Grouping those records by year can move an older judicial event into a newer bucket. Our date and text coverage audit explains that risk. A year filter alone does not construct a legally valid before-and-after research design.

Define missing dates explicitly. An absent decision date is not the publication date, and an unknown transaction completion date is not evidence that completion never happened. Keep the timestamp of each status observation so that a later change does not overwrite what the system previously knew. A useful timeline distinguishes the event claimed by a document from the date on which your team obtained that evidence.

6. Evaluate text and identifiers before building an AI answer layer

Start with a small, scoped set of known source objects supplied by the buyer or independently verified by researchers. Establish which track each belongs to and the question it can answer. If the pilot is intended to retrieve published conditions, evaluate that task against the administrative materials. If it is intended to retrieve judicial reasoning, evaluate the court documents. Combining both into a single pass rate conceals different failure modes.

For each judicial result, check whether the text is usable for the intended task. A populated field may contain a non-publication notice, an incomplete passage or content that does not include the relevant reasoning. The presence of a body field is not a completeness measure. Record the missing or unusable-text cases in the denominator rather than excluding them after the supplier has shown a successful demonstration.

When a sample is needed, disclose its construction. Stratify by ingestion batch and publication category where applicable, retain the selection method and seed, and examine repeated proceedings or corporate groups. A batch of near-identical documents can make a retrieval test look strong while offering little coverage of diverse questions. No such sample audit was performed for the three counts on this page.

Have the answer layer distinguish facts explicitly recorded in a source, a party's allegation, quoted administrative material and the court's conclusion. Ask it to state the source track and observation date. A model should not convert “this court text mentions a concentration” into “SAMR approved the transaction.” Test that precise error in the pilot, alongside ambiguous names, absent dates and a notice without a corresponding outcome document.

7. Put the boundary in the delivery and acceptance plan

A useful pilot request names the geography, time period, document categories, languages and intended output. Specify whether the buyer needs source discovery, structured court retrieval, administrative decision extraction, entity matching or a current-status workflow. Those are different deliverables. Ask suppliers to label what is already available, what would be newly collected and what remains unavailable after the pilot.

For an API or bulk court-data delivery, inspect actual returned fields and the limitations of the selected slice. Our vendor coverage checklist helps separate a successful request from verified coverage. A working endpoint is evidence of access, not evidence that every relevant administrative publication or court judgment has been collected.

Agree a correction process before production: how a mistaken relation is reversed, how a changed publication is versioned, and how an unavailable source is represented. Retain the original Chinese text alongside any English summaries and identify translations as derivative outputs. Verify commercial reuse, personal-data handling and retention terms independently; public accessibility alone does not decide every downstream use permission.

SinoVerdict's enterprise court-data service can be evaluated for the judicial component through a trial API and a scoped corpus coverage report. The delivery overview explains available delivery routes. This page makes no claim to a complete SAMR filing database, live merger-status monitoring or comparative superiority in coverage. Keep those requirements separate when requesting a trial.

Frequently asked questions