China's GEO agencies have started doing something their Western counterparts mostly refuse to do: they put numbers in the contract.
Provider materials circulating this year promise "AI visibility improvement of 45% or above" and "core keyword TOP3 share of at least 70%," with pro-rata refunds when targets are missed. Others go further — public claims of 6x to 60x visibility gains and an average 65% lift in top-three recommendation rate. A second tier of vendors now sells a RaaS ("results as a service") structure: a thin base fee covering operations, plus a performance bonus tied to an AI visibility index, an AI first-recommendation rate, and a core keyword mention rate. Procurement whitepapers have appeared to match, one of them organized around "seven red lines and eighteen contract clauses."
For a category that barely existed eighteen months ago, this is remarkable commercial maturity. It is also a trap, and not the obvious one. The risk is not that vendors will fail to hit the numbers. The risk is that they will hit them, invoice you, and you will have no way to determine whether anything happened at all.
Why the guarantees appeared now
The Chinese GEO market is on a trajectory that makes performance guarantees a rational sales weapon. Domestic forecasts put the market above RMB 50 billion by 2030, and spend inside brands has shifted from experimental budget lines into core marketing allocation. When a category grows that fast, differentiation stops being about method and starts being about risk transfer. The vendor who says "we'll refund you" wins the bake-off against the vendor who says "it depends."
The underlying demand is real. China's AI-native apps reached roughly 440 million monthly active users by March 2026 (QuestMobile), with Doubao at about 345 million, Qwen at 166 million, and DeepSeek at 127 million. Doubao's "帮你选" shopping feature, launched in April 2026 against Douyin's e-commerce supply chain, collapsed the old twelve-click product search into a single conversation. Qwen did the equivalent through Taobao. Brands can see the traffic moving. They want someone to be accountable for where it lands.
The problem is what "accountable" can actually mean on this surface.
The four variables a "visibility lift" number conceals
An AI visibility score is not a measurement in the way a rank-tracking number was a measurement. It is the output of a sampling procedure, and the procedure has at least four free parameters. Whoever controls those parameters controls the result — without touching your brand, your content, or the model.
The prompt set. A visibility score is computed over a list of questions. Move one question from "推荐几个高端护肤品牌" to "有哪些适合敏感肌的高端护肤品牌," and a brand with strong dermatological content jumps while a brand with strong luxury positioning falls. Neither brand changed. If the vendor authors the prompt set, refines it during the engagement, or holds it as proprietary methodology, the guaranteed 45% lift is a number the vendor can produce on demand. This is not hypothetical malice — it is what any optimizer does when a metric is also its scoreboard.
The model version. Chinese models ship on a cadence that has no Western analogue: Doubao Seed 2.0 Pro, DeepSeek-V4-Pro, Qwen3.7, Kimi K2.5 and GLM-5 have all moved through material versions inside 2026. A baseline measured on one checkpoint and a final measured on another are not comparable, and the difference routinely exceeds the guaranteed delta. We covered this in The Silent Variable Breaking Your China AI Brand Scores in 2026; a performance contract that doesn't pin model versions is a contract that pays out on someone else's release notes.
The sampling window. Chinese AI answer content decays fast — brand material that anchors a model's answer in week two is frequently gone from the retrieval set by week eight. A baseline taken in a dead news week against a final taken while your brand is in the middle of a Douyin campaign will show a lift that has nothing to do with the GEO work and everything to do with the calendar. Under a refund clause, the vendor's incentive is to run the final measurement when conditions are best, not when they're representative.
The account state. Scores differ by tier, by whether assistant memory is on, and increasingly by which surface the query is routed to — app, mini-program, preinstalled phone assistant, or in-car. Most vendor measurement runs on free-tier API access with no memory. A meaningful share of China's commercial buyers no longer live there. A guarantee delivered on the free tier is a guarantee about a population that may not include your customer.
Stack the four and the arithmetic is uncomfortable: a vendor who selects the prompt set, chooses the checkpoint, times the window, and picks the account state can manufacture a 45% lift without shipping a single piece of content. Most won't. But the contract as written cannot distinguish the ones who do from the ones who don't — which means it isn't a guarantee, it's a vibe with a refund clause attached.
Refund clauses make this worse, not better
The instinct is that a money-back provision at least caps the downside. In this category it inverts the incentives.
A refund clause turns the metric into the deliverable. The vendor's cheapest path to payment is not to move the model's behavior — an expensive, slow, partially uncontrollable outcome — but to move the measurement, which is fast, cheap, and entirely within their control. Every free parameter above becomes a lever. And because the measurement is also the audit, there is no independent record against which to check the claim.
There is a second-order problem. The fastest way to move a genuine visibility number in the Chinese ecosystem is to flood the sources models actually read — and the cheapest versions of that tactic shade into content that regulators have already priced. A Hangzhou court decision this year treated GEO-placed brand content as advertising subject to China's Advertising Law, which means your vendor's KPI shortcut can become your brand's legal exposure. A refund of the service fee does not cover that.
What to specify instead
Performance language isn't the problem. Unowned measurement is. Six clauses that convert a promise into something auditable:
-
The prompt set is yours, frozen, and escrowed. Written into an appendix at kickoff, version-controlled, changed only by mutual written agreement, with any change resetting the baseline. If a vendor calls the prompt set proprietary, that is the whole negotiation.
-
Model versions and surfaces are named. "Doubao (app, free tier, memory off), DeepSeek (web), Qwen (app)" — not "six major Chinese models." Require re-baselining on any named model's major version change, at vendor cost.
-
Measurement runs on a fixed cadence, not on demand. Same day of week, same time band, same frequency, from the engagement's first week to its last. This single clause removes most of the window-timing lever.
-
Report raw responses, not scores. The deliverable is the full text of every model answer, dated and versioned, alongside the computed metric. Scores without transcripts are unfalsifiable. Transcripts let you or a third party recompute.
-
Split the metric. "Visibility" bundles mention rate, first-recommendation rate, sentiment, and factual accuracy into one number that can be moved by whichever component is easiest. Contract each separately, and add a factual-accuracy floor — a brand recommended confidently with the wrong price or the wrong flagship property is a loss, not a win.
-
Source disclosure and a compliance warranty. Require a list of every domain, account, and platform where content was placed on your behalf, delivered monthly. Pair it with a warranty that all placements comply with advertising disclosure rules and an indemnity that survives termination. This is the clause that protects you from your own KPI.
None of this requires distrusting your vendor. It requires that the party being paid on a number is not the only party who can compute it.
Takeaway for Brand Marketers
Treat a Chinese GEO performance guarantee the way you'd treat a media agency guaranteeing its own impression counts with no third-party verification — not as fraud, but as a structure that cannot be checked. Before you sign anything with a percentage in it, establish your own baseline: your prompt set, your named model versions and surfaces, your fixed cadence, your stored transcripts. Run it for four weeks before the vendor starts. That pre-engagement baseline is the only asset in this entire process that the vendor cannot adjust, and it converts every subsequent claim into something you can test.
Then contract on inputs and disclosure rather than on outcome percentages — placements made, sources touched, corrections filed, entity records fixed — and hold the outcome metric as your own internal scorecard. In a measurement environment this unstable, the brand that owns the ruler wins regardless of who swings the hammer.
Related: The Silent Variable Breaking Your China AI Brand Scores in 2026 · Your Brand Score Was Measured on the Free Tier · Track brand scores across all six Chinese models