iFlytek Won 210 Enterprise AI Contracts. Your Brand Visibility Dashboard Has Never Heard of It. (2026)

Here is a number almost no brand team in China tracks: 210.

That is how many enterprise large-model contracts iFlytek won in the most recent full-year Chinese public-tender tally — RMB 2.316 billion in disclosed value, first place in both project count and contract value among general-purpose LLM vendors. iFlytek is the reigning "bid king" (标王) of China's enterprise AI market.

It is also a model that appears on virtually no brand visibility dashboard. Ask a marketing team which Chinese AI models they monitor and you will hear the consumer roster: Doubao, DeepSeek, Qwen, Kimi, ERNIE, Hunyuan. That list is built from monthly active users. The procurement list is built from signed contracts. They are not the same list — and increasingly, they are not even the same market.

This matters because the answers your customers see are not all delivered by an app with a logo. A growing share arrives inside someone else's product: a bank's advisory chatbot, an insurer's claims assistant, an OTA's trip planner, a retailer's in-app shopping guide, a manufacturer's spec-lookup tool. Every one of those has a foundation model underneath it. Every one of those inherits that model's brand priors. And nobody is measuring them.

The two markets have decoupled

Look at the consumer side first. As of May 2026, China's native AI apps reached roughly 499 million monthly active users, led by Doubao at 382 million, the Qwen app at 167 million, and DeepSeek at 130 million. That is the surface brand teams audit, and rightly so — it is enormous.

Now look at the procurement side. In the financial vertical alone, one full-year tally counted 587 large-model tender projects with RMB 1.506 billion in disclosed value. Project count grew 341% year over year; disclosed value grew 527%. The top five vendors by that measure were Baidu, iFlytek, Volcano Engine, Zhongguancun Kejin, and Alibaba Cloud — a lineup that only partially overlaps with the consumer leaderboard, and that includes at least two names most international brand teams have never evaluated.

The single largest disclosed win in that period went to Alibaba Cloud: a RMB 179 million compute-and-model-application procurement for a cable operator. That is one contract worth more than most brands' entire annual China marketing budget, and it will shape how a model answers questions for a specific user population for years.

Zhipu is the sharpest illustration of the gap. The company went public in Hong Kong in January 2026 at a valuation above six billion dollars. Internationally it holds a modest share of routed developer traffic — roughly 5.6% on OpenRouter — but it commands a disproportionate share of Chinese domestic enterprise procurement. Its GLM 5.2 release saw what Vercel's agentic infrastructure team described as the fastest adoption of any model they tracked in 2026: daily token volume up roughly 27x and customer count up roughly 80x in its first full week.

A model can be nearly invisible in consumer MAU charts while quietly becoming the default reasoning engine behind hundreds of Chinese enterprise products. Zhipu, Moonshot, and MiniMax together account for something on the order of 10% of the foundation model layer. That 10% is not evenly distributed across use cases — it is concentrated in exactly the B2B and embedded contexts where purchase decisions get made.

Why enterprise deployments are worse for brand accuracy, not better

The instinct is to assume an enterprise deployment is safer for your brand: it is scoped, it is governed, someone is accountable. In practice the opposite is usually true, for three structural reasons.

Retrieval is narrowed, not widened. An enterprise RAG deployment restricts the model to an approved corpus — the client's own product database, their contracted supplier list, their internal knowledge base. If your brand is not in that corpus, the model will not fall back to the open web to find you. It will simply answer from what it has. In consumer AI search, a weak brand still has a chance of being surfaced from a third-party review site. In an enterprise deployment, absence from the corpus is absolute.

Base-model priors leak through anyway. Where the corpus is silent — comparisons, category framing, "which type of X should I choose" — the underlying model's pretraining priors fill the gap. So a retailer's shopping assistant built on Volcano Engine inherits the same ByteDance-ecosystem tilt that shapes Doubao, where roughly 60% of citations trace to ByteDance properties (Toutiao at about 35%, Douyin at about 25%). A tool built on a Baidu foundation inherits the ERNIE profile, where roughly 55% of citations come from the Baidu ecosystem. Your Douyin and Baijiahao footprints are not just consumer-facing assets. They are inputs to enterprise systems you will never be told about.

There is no visible surface to audit. You can open the Doubao app and type your category query. You cannot open a bank's internal advisory model. The failure mode is silent by construction: you lose a shortlist slot inside a procurement tool and receive no signal at all — no ranking drop, no traffic dip, nothing that shows up in a dashboard.

The routing question brands should actually be asking

The useful reframe is to stop asking "which AI apps do my customers use?" and start asking "which model sits underneath the tools that touch my buying journey?"

For a consumer brand, that chain usually runs: category question in Doubao or Qwen → product comparison inside a Taobao or Douyin surface → final check inside a retailer's own assistant. Three surfaces, potentially three different foundation models, at least one of which you cannot inspect.

For a B2B or considered-purchase brand, the chain is worse. A procurement analyst's shortlist may be generated inside an enterprise assistant running on iFlytek or Zhipu, drawing from a vendor database you were never invited into. By the time anyone opens a browser, the shortlist already exists.

This is a different problem from the one most GEO programs are built to solve. Consumer-facing GEO is about being present and accurate on a measurable surface. Enterprise-layer GEO is about being ingestible — present in the structured, machine-readable places that enterprise corpora get built from: industry directories, certification registries, standardized product feeds, supplier databases, and the Chinese encyclopedia and Q&A properties that vendors scrape when assembling a knowledge base.

A quick way to size your own exposure

You do not need a research budget to find out whether this applies to you. Run a three-column check on your top ten category queries.

Column one: the answer from the consumer app your customers actually use — Doubao for mass-market consumer questions, Qwen for anything anchored in the Taobao and Tmall ecosystem, DeepSeek or Kimi for research-shaped queries.

Column two: the answer from the same vendor's raw API with no search tool enabled. This is the closest available proxy for what an embedded deployment sees when its corpus comes up empty. It isolates pure pretraining priors from live retrieval.

Column three: the answer from an enterprise-tilted model you would not otherwise test — GLM being the practical pick.

Where columns one and two agree, your position is durable. Where column one is strong and column two collapses, your visibility is retrieval-dependent and will not survive being embedded in someone else's product. Where column three diverges from both, you have a blind spot in exactly the layer that generates B2B and considered-purchase shortlists.

Most brands that run this exercise find the same pattern: a respectable consumer score, a much weaker unassisted score, and a third-column result that looks like a different company entirely. The fix is not more content. It is denser, more structured, more machine-readable evidence in the places enterprise builders actually scrape.

Takeaway for Brand Marketers

Four concrete moves, in order of effort:

1. Extend your model list beyond consumer apps. Add at least one enterprise-tilted model to whatever you test — Zhipu's GLM is the obvious candidate given its domestic procurement position, and iFlytek's Spark is worth a baseline read given its tender lead. You are not testing them for consumer reach. You are testing them because they are the substrate of tools you cannot see.

2. Test the same query on the consumer app and the API. Doubao the app and Volcano Engine the platform are not the same product, and the gap between them is a proxy for how your brand fares once the consumer scaffolding is stripped away. If your brand score drops sharply on the raw API, your visibility depends on retrieval you will not get in an embedded deployment.

3. Audit for ingestibility, not just presence. Ask a blunt question about every asset you own in China: could a vendor building an enterprise knowledge base pull this in a batch job? A Xiaohongshu post cannot be. A structured product feed, a Baidu Baike entry, an industry association listing, and a certification registry record can be. Prioritize accordingly.

4. Follow the tender data as a leading indicator. Public procurement disclosures tell you which model is about to be embedded in which vertical, often a year before it affects your category. A 341% jump in financial-sector large-model projects is advance notice that finance-adjacent brands are about to be shortlisted by systems they have never tested.

The uncomfortable summary: the model that decides whether your brand makes a Chinese shortlist may not have an app, a logo, or a MAU figure. It has a contract number.

Related: see how international brands currently score across China's major AI models on the hubGEO brand rankings.