China Social Commerce Data: The Agent Opportunity

Key takeaway (TL;DR) — China’s social commerce runs on Douyin, Xiaohongshu, Weibo and a few other platforms where content, engagement and purchase sit in the same app. AI agents could help brands, creators and agencies work inside it, but only if they can read structured public data from those platforms at machine speed. The data is there. The bottleneck is access architecture.

The market

China social commerce data matters because discovery, content and purchase often happen inside the same platform. Market-size estimates vary a lot by source and definition, so this piece doesn’t lean on a GMV figure; the argument rests on how the platforms work.

The key platforms:

Platform	Primary commerce model
Douyin	Live commerce + short video
Xiaohongshu	Content-driven discovery
Kuaishou	Live commerce
Weibo	Celebrity/brand endorsement
WeChat	Mini-programs, private traffic

These aren’t just social platforms with a “buy” button. They are integrated commerce ecosystems where:

Content is the storefront
Engagement metrics are purchase signals
Creator influence is distribution
Algorithmic recommendation is merchandising

Why agents need this data

Traditional analytics tools for Chinese social commerce assume a human will review dashboards, interpret trends and make decisions. That creates a bottleneck that scales linearly with team size.

AI agents loosen this constraint. An agent can:

Track public product notes and reviews on Xiaohongshu, flagging shifts in review sentiment and price mentions
Screen many KOL candidates for campaign fit based on public content, engagement counts and marketplace listings
Detect trending products before they peak by cross-referencing Xiaohongshu mentions, Douyin video engagement and Weibo discussion
Plan live commerce schedules using public live-room and competitor activity

Each of these needs structured, programmatic access to public platform data. Not screenshots. Not CSV exports. Not weekly reports. JSON at agent speed.

The data landscape: What’s available

Scope for everything below: public, read-only data through one SandBase API key. SandBase is not an official partner of these platforms, and these endpoints don’t give access to private accounts, DMs, owner-only analytics (such as a shop’s own conversion or cart data), publishing or account actions. Where an item is something you compute yourself, it’s marked derived. Endpoint names are from the catalog as of 2026-10-01. Field names are stated only for the three endpoints we called for this post (douyin/app-v3/user-post-videos, douyin/app-v3/multi-video-statistics, douyin/app-v3/user-profile); the rest are described by capability, so check each endpoint reference for what it returns.

Douyin commerce data

Creator/influencer data:

Public profile and follower count (douyin/app-v3/user-profile; user.follower_count observed on 2026-10-01)
Per-video like, comment, share and save counts in a public account’s video list (douyin/app-v3/user-post-videos; see the tested example below)
Batch statistics for a list of video IDs (douyin/app-v3/multi-video-statistics): on 2026-10-01 it returned like, play and share counts, plus a download count for one of the two videos; no comment or save counts (example below)
Engagement rate and growth trends (derived from the counts above, stored over time)
Xingtu marketplace listings: listed service prices and the aggregate audience portrait the marketplace shows for a creator (douyin/xingtu/kol-service-price-v1, douyin/xingtu/kol-audience-portrait-v1)
Public live-room info (douyin/web/user-live-info-by-uid)

Product/shop data: The SandBase Douyin catalog had no dedicated Douyin Shop endpoints as of 2026-10-01. Product signals come indirectly, from public videos and comments (douyin/app-v3/video-comments) and Xingtu’s creator–product views (douyin/xingtu-v2/author-convert-videos-or-products).

Trend/discovery data:

Billboard rankings across categories (douyin/billboard/*, e.g. hot-total-list)
Hot search and related keywords (douyin/app-v3/hot-search-list, douyin/search/search-suggest)
Keyword search across videos and users (douyin/search/*)

As of 2026-10-01 the SandBase Douyin catalog shows 262 Douyin endpoints across the index, web, app-v3, billboard, Xingtu, creator, DOU+ and search channels.

The catalog page itself carries that count, and each endpoint card shows its status and price.

Caption: The SandBase Douyin catalog with the “262 SHOWN” endpoint count; the selected multi-video-statistics card reads Available · Free (captured 2026-10-02).

Xiaohongshu commerce data

Xiaohongshu is the “discovery engine” of Chinese social commerce. Users research products here before buying anywhere:

Content data:

Note search, note details and note comments (xiaohongshu/app-v2/search-notes, xiaohongshu/app-v2/image-note-detail, xiaohongshu/app-v2/note-comments)
Product search, details and public reviews (xiaohongshu/app-v2/search-products, xiaohongshu/app-v2/product-detail, xiaohongshu/app-v2/product-reviews)
Topic feeds for seasonal patterns (xiaohongshu/app-v2/topic-feed)

Creator data:

Public user info and posted notes (xiaohongshu/app-v2/user-info, xiaohongshu/app-v2/user-posted-notes)
Pugongying creator-marketplace detail and fan-profile endpoints (xiaohongshu/pgy/blogger-detail, xiaohongshu/pgy/blogger-fans-profile)
Content category mix (derived from a creator’s notes)

Commerce signals (all derived from public notes and comments):

Product mention frequency and growth
Price sensitivity and purchase-intent language in comments

As of 2026-10-01 the SandBase Xiaohongshu catalog shows 34 endpoints.

Weibo commerce data

Weibo serves as the sentiment layer, where public opinion forms and brand crises emerge:

Discussion data:

Real-time search and hot search (weibo/web-v2/realtime-search, weibo/web-v2/hot-search)
Reposts, comments and likes on a public post (weibo/app/status-reposts, weibo/app/status-comments, weibo/app/status-likes)
Sentiment polarity and volume spikes (derived with your own model over search results)

Celebrity/brand data:

Public timelines of brand and celebrity accounts (weibo/app/user-timeline)
Share of voice versus competitors (derived from search volume)

As of 2026-10-01 the SandBase Weibo catalog shows 52 endpoints.

The bottleneck: Data access architecture

The core problem: the data is public on the platforms, but there’s no general programmatic route to it for third-party agents.

What doesn’t work

Manual export: Analysts copying data from platform pages. Scales at O(n) with team size. Introduces hours or days of latency. Agents can’t use it.

Browser automation: Headless browsers scraping platform UIs. Fragile (breaks when the UI changes), slow, hard to maintain across platform updates, and it can raise terms-of-service and legal questions depending on what you collect. If you process personal information, check your obligations under China’s Personal Information Protection Law for your specific use case.

Official APIs (scoped to authorized accounts): The Douyin Open Platform serves authorized, account-linked workflows, and Ocean Engine gives advertisers their own campaign data. They’re the right tool when you act for accounts that authorized you, not for market-level public research.

Need	Official platform APIs	Public-data API (e.g. SandBase)
Data scope	Accounts that authorized your app	Public, read-only data
Auth	Platform developer app + user OAuth	One SandBase API key
Publishing / account actions	Yes, for authorized accounts	No
Owner-only analytics	Yes, for authorized accounts	No
Fits	Managing your own or clients’ accounts	Market research, monitoring, agent tools

What works

Structured public-data APIs: services that expose documented REST endpoints for public platform data. This is what makes agent-speed access possible.

The example calls douyin/app-v3/user-post-videos (endpoint reference). You need a SandBase API key.

import os
import httpx

SANDBASE_API_KEY = os.environ["SANDBASE_API_KEY"]


async def latest_videos(sec_user_id: str, max_cursor: int = 0) -> dict:
    """One page (up to 20) of a public account's videos."""
    async with httpx.AsyncClient(timeout=60) as client:
        resp = await client.post(
            "https://api.sandbase.ai/v1/api/douyin/app-v3/user-post-videos",
            headers={"Authorization": f"Bearer {SANDBASE_API_KEY}"},
            json={"sec_user_id": sec_user_id, "count": 20, "max_cursor": max_cursor},
        )
        resp.raise_for_status()
        body = resp.json()
    outputs = body.get("outputs") or []
    if body.get("status") != "completed" or not outputs:
        raise RuntimeError(
            f"SandBase call did not complete: {body.get('status')} {body.get('error')}"
        )
    data = outputs[0].get("data") or {}
    return {
        "videos": data.get("aweme_list") or [],
        "has_more": bool(data.get("has_more")),
        "next_cursor": data.get("max_cursor"),
    }
# Parse, reason, decide: all in one agent loop iteration

This is the Model API route (POST /v1/api/douyin/app-v3/user-post-videos, Bearer auth, JSON body with only the endpoint’s parameters). The reference documents the completed payload at outputs[0].data; the function raises when a call doesn’t complete.

The endpoint reference shows the same route and the three body fields the function sends.

Caption: The douyin/app-v3/user-post-videos reference: POST /v1/api/douyin/app-v3/user-post-videos with count (max 20), max_cursor for paging and required sec_user_id (captured 2026-10-02).

Tested on 2026-10-01 (UTC) with the public official account of People’s Daily. Request body:

{"sec_user_id": "MS4wLjABAAAA8U_l6rBzmy7bcy6xOJel4v0RzoR_wfAubGPeJimN__4", "count": 20, "max_cursor": 0}

Trimmed response (one of 20 videos shown, most fields removed):

{
  "id": "ae988129-adb7-4393-a10d-46484cfbb5ed",
  "status": "completed",
  "model": "douyin/app-v3/user-post-videos",
  "outputs": [{
    "data": {
      "has_more": 1,
      "max_cursor": 1790762105000,
      "aweme_list": [{
        "aweme_id": "7691636135411420452",
        "desc": "国庆前夕，人民日报记者来到集市拍下这组照片……",
        "create_time": 1790848603,
        "statistics": {"digg_count": 14620, "comment_count": 82, "share_count": 254, "collect_count": 360}
      }]
    }
  }]
}

aweme_list, has_more, max_cursor and the statistics counts come from that call; they are observed-only, not a documented guarantee. To page, send the returned max_cursor as the next request’s max_cursor while has_more is truthy.

Tested on 2026-10-01 (UTC): douyin/app-v3/multi-video-statistics with two public video IDs (the People’s Daily video above and one more public video). Request body {"aweme_ids": "7691636135411420452,7687107502068026650"}; trimmed response (run id af0f2b72-968d-4475-b7ee-d77a5dde545a):

{"statistics_list": [
  {"aweme_id": "7691636135411420452", "digg_count": 54636, "download_count": 30, "play_count": 2191932, "share_count": 57},
  {"aweme_id": "7687107502068026650", "digg_count": 195, "play_count": 24922, "share_count": 4}
]}

These fields are observed-only. download_count was missing for the second video, so read every field with .get(); for comment and save counts, use the user-post-videos list, which returned comment_count and collect_count in the test above.

Together, the four catalogs most relevant here show 493 endpoints as of 2026-10-01: Douyin 262, TikTok 145, Weibo 52 and Xiaohongshu 34, all behind the same key and calling pattern. For a broader survey, see Top 5 Social Media Data APIs for AI Agents.

Opportunity 1: Product trend detection agent

Problem: Brands discover trending products after they peak. By the time a human analyst spots a Douyin trend, it may already be saturated.

Agent architecture:

Every 15 minutes:
  1. Scan Douyin billboards for rising items
  2. Cross-reference with Xiaohongshu mention velocity
  3. Check Weibo discussion for early negative signals
  4. Score trend strength (volume × acceleration × sentiment)
  5. Alert if score exceeds threshold AND market isn't saturated

Data requirements:

Douyin billboard endpoints (douyin/billboard/*)
Xiaohongshu note search (xiaohongshu/app-v2/search-notes)
Weibo real-time search (weibo/web-v2/realtime-search)
Historical baselines you store for scoring

Opportunity 2: KOL selection and vetting agent

Problem: Brands evaluate influencers manually from headline metrics that don’t show audience fit or commerce potential. Fake followers inflate perceived value.

Agent architecture:

Given: Brand brief (product, target audience, budget)
For each candidate KOL:
  1. Pull the Xingtu marketplace listing and audience portrait
  2. Analyze content-commerce alignment from public videos
  3. Flag suspicious engagement (ratio heuristics on public counts)
  4. Estimate cost per view from listed prices and public view counts
  5. Score and rank against brief requirements

Data requirements:

Xingtu endpoints (douyin/xingtu/*, douyin/xingtu-v2/*)
Public video engagement (douyin/app-v3/user-post-videos)
Xiaohongshu Pugongying listings (xiaohongshu/pgy/*) for cross-platform creators

Steps 3 and 4 are your agent’s own estimates, not platform-reported metrics. See our KOL screening agent tutorial for implementation patterns.

Opportunity 3: Live commerce optimization agent

Problem: Live stream scheduling and product ordering are often decided by gut feel.

Agent architecture:

Pre-stream:
  1. Review public live history of similar creators
  2. Identify competitor streams in the same time slots
  3. Recommend a start time and duration
  4. Order the product list by predicted engagement

During stream (near-real-time):
  5. Watch public live-room signals of competing streams
  6. Suggest product switches when your own team sees engagement drop

Data requirements:

Public live-room info and live videos (douyin/web/user-live-info-by-uid, douyin/web/user-live-videos)
Public video history of comparable creators
Your own stream’s sales data stays in your own seller tools; it isn’t in the public-data API

Opportunity 4: Cross-platform price intelligence agent

Problem: The same product appears on Douyin, Xiaohongshu, Tmall and JD at different prices. Brands lack cross-platform price monitoring.

Agent architecture:

For each monitored SKU:
  1. Search Xiaohongshu product listings and notes for price references
  2. Scan public Douyin videos and comments for price mentions
  3. Flag possible unauthorized resellers
  4. Alert on price undercutting
  5. Recommend price adjustments

Data requirements:

Xiaohongshu product search and details (xiaohongshu/app-v2/search-products, xiaohongshu/app-v2/product-detail)
Douyin video search and comments (douyin/search/*, douyin/app-v3/video-comments)
Cross-platform product matching (by name or image); Tmall/JD data is outside these catalogs

Problem: Brand crises on Chinese social media can escalate within hours. Manual monitoring can’t keep pace.

Agent architecture:

Continuous monitoring:
  1. Track brand/product mentions across Weibo, Douyin, Xiaohongshu
  2. Sentiment analysis on new mentions
  3. Volume anomaly detection (spike = potential crisis)
  4. Escalation scoring (KOL amplification + negative sentiment + volume)
  5. Immediate alert if escalation score crosses threshold

Data requirements:

Weibo real-time search and hot search (weibo/web-v2/realtime-search, weibo/web-v2/hot-search)
Douyin comment and video search
Xiaohongshu note search
Cross-platform volume aggregation (your own)

Cost to run a competitive intelligence agent

An estimate for one example workload. Data endpoints: as of 2026-10-01 the catalog shows all four used here as Free (douyin/app-v3/user-profile, douyin/app-v3/multi-video-statistics, xiaohongshu/app-v2/search-notes, weibo/web-v2/realtime-search). Prices can change; check each model card before budgeting, and don’t assume quotas beyond what’s documented. One inconsistency to know about: the multi-video-statistics reference description still contains an upstream line, “Price: 0.025$ each time”, while the catalog and the model page (“Base price Free”) list it as Free as of 2026-10-01.

Until SandBase reconciles the two, treat the data-API cost of this workload as unverified rather than zero. The workload below makes 144,000 multi-video-statistics calls a month; if the reference’s upstream price applied, that line alone would be 144,000 × $0.025 = $3,600/month. The other three endpoints showed no such conflict when checked on 2026-10-01. The LLM estimate below is the only part with a verified price.

Caption: The douyin/app-v3/multi-video-statistics model page, showing Base price Free, sync execution and one input field (aweme_ids, up to 50 IDs) (captured 2026-10-02).

Hourly monitoring cycle:
  - 50 Douyin profile checks       (douyin/app-v3/user-profile)
  - 200 video stats calls          (douyin/app-v3/multi-video-statistics)
  - 30 Xiaohongshu searches        (xiaohongshu/app-v2/search-notes)
  - 20 Weibo searches              (weibo/web-v2/realtime-search)
  Total per cycle: 300 data calls

Daily (24 cycles): 7,200 data calls
Monthly (30 days): 216,000 data calls (data-API cost: unverified, see above)

LLM reasoning (assumption: 100 calls/day, 3,000 input + 500 output tokens each,
30 days = 9M input + 1.5M output tokens per month), SandBase prices as of 2026-10-01:
  - openai/gpt-4.1-mini ($0.40/M in, $1.60/M out): 9 × 0.40 + 1.5 × 1.60 = $6.00/month
  - openai/gpt-4o       ($2.50/M in, $10/M out):   9 × 2.50 + 1.5 × 10   = $37.50/month

Compared with a human analyst doing the same checks, the agent covers 50 accounts every hour around the clock. It doesn’t replace judgment on what to do about a signal; it removes the collecting and first-pass sorting.

What’s needed to unlock the opportunity

1. Broader commerce coverage

Shop analytics, live-commerce metrics and product catalog data are still underserved by public-data APIs; the Douyin catalog, for example, has no dedicated shop endpoints today.

2. Better cross-platform identity resolution

Agents need to correlate creators and products across platforms. A KOL who’s big on Douyin may also be on Xiaohongshu. Products appear under different names. Identity resolution would unlock cross-platform analytics.

3. Historical data access

Most APIs return current state. Agents making strategic decisions need 30/60/90-day trajectories for engagement, follower growth and content output, which today usually means storing your own snapshots.

4. Event-driven data delivery

Beyond polling, agents benefit from push-based data: notifications when monitored metrics cross thresholds, reducing unnecessary calls.

5. Compliance infrastructure

Agents need clear data provenance, usage logging and, where personal information is involved, a lawful basis for processing. This is infrastructure, not a feature.

Ready to try it: create a SandBase API key, browse the Douyin catalog, or open the user-post-videos endpoint reference.

Conclusion

China’s social commerce is a large commerce opportunity that AI agents haven’t yet reached at scale. The public signals exist: content interactions, creator listings, rankings and discussion on every product and creator. The bottleneck is not intelligence (LLMs are capable enough) or demand. It’s structured data access at agent speed.

Platforms that solve this access layer, with reliable, documented, normalized public-data APIs across Chinese social platforms, will enable the next generation of commerce agents. The window is open now because:

Agent frameworks reached production use in 2025–2026
Structured public-data endpoints now cover the main Chinese social platforms behind one key
Teams that start collecting history early get compounding data advantages

For developers and entrepreneurs, the agent opportunity in China social commerce is architectural, not algorithmic. The hard problem isn’t “build a smarter model”. It’s “get structured data from Douyin, Xiaohongshu and Weibo into an agent’s decision loop.” Solve that, and much of this commerce activity becomes programmable.

The market

Why agents need this data

The data landscape: What’s available

Douyin commerce data

Xiaohongshu commerce data

Weibo commerce data

The bottleneck: Data access architecture

What doesn’t work

What works

Five agent opportunities in China social commerce

Opportunity 1: Product trend detection agent

Opportunity 2: KOL selection and vetting agent

Opportunity 3: Live commerce optimization agent

Opportunity 4: Cross-platform price intelligence agent

Opportunity 5: Social sentiment early warning agent

The economics of agent-powered social commerce

Cost to run a competitive intelligence agent

What’s needed to unlock the opportunity

1. Broader commerce coverage

2. Better cross-platform identity resolution

3. Historical data access

4. Event-driven data delivery

5. Compliance infrastructure

Related Reading

Conclusion