# The Business Case for China AI Visibility

A reference for the CMO, VP Marketing or category lead asking whether to invest in China AI visibility — what is actually measurable today, what is honestly still hypothesis, three scenarios where the case is strongest, what spending typically looks like, and where the work can fail. We measure surface, not sales attribution; this page is built around that boundary.

_Last reviewed 2026-05-10. Methodology grounded in published research; per-engine numbers from Eastbound's own zh-CN consumer panels._

## The shift in how Mainland-Chinese consumers research products

The starting fact is that Mainland-Chinese consumers have moved a meaningful share of their product-research questions from search-engine queries to AI assistants. DeepSeek crossed 100 million monthly active users within weeks of its January 2025 V3 release; Qwen (Alibaba's Tongyi) and Doubao (ByteDance) have each reported user bases in the tens-of-millions range across consumer-facing surfaces in 2025–2026. The composite picture: a non-trivial fraction of category-research questions in Mainland China is now being answered by a generative engine rather than a list of blue links.

Numbers in this space drift quickly and provider-published figures vary by surface (mobile app, web, embedded in WeChat, embedded in third-party apps). Treat every adoption figure with sample-size disclosure: the share that matters for any specific brand is not "DeepSeek MAU" but "what fraction of consumers in our category use AI assistants for category research, in our markets, in our purchase window". That is a research question every business case should answer for itself; population-level adoption stats are the floor, not the answer.

Top-15 source overlap (Jaccard) between the three Chinese engines is 0.20–0.30 in our 540-call panel (May 2026), per [DeepSeek vs Qwen vs Doubao: Why Brands Look Different](https://www.eastbound.ai/blog/three-chinese-ais.html). The engines diverge enough that "we are visible on Chinese AI" is not a single claim — it is three separate questions about three separate surfaces with materially different source pools.

## What is measurable today vs what is honest hypothesis

The single most important sentence in any China AI visibility business case: **we measure selection, absorption and mention. We do not measure sales lift.** The full distinction between what the methodology can support and what it cannot:

| Outcome | Status | How we'd measure / where the limit is |
|---|---|---|
| **Selection** — does the engine pull our domain into its source pool? | Measured | Direct observation per engine, per prompt panel; reliability stats reported (κ_top-5, κ_top-15) |
| **Absorption** — does our content shape the answer language? | Measured | Direct observation; recall depth and language-reuse scoring |
| **Mention** — does the answer name our brand? | Measured | Direct observation; mention rate and positioning frame captured per response |
| **Source-graph density** — how many third-party surfaces reference us? | Measured | Top-15 source extraction per engine, per category |
| **Sales lift attributable to AI mention** | Hypothesis | No published peer-reviewed study attributing sales lift to GEO interventions in Mainland China. Plausibility argument only; would require brand-side click-tracking + uplift testing the brand controls itself. |
| **Consideration-set entry** — does AI mention drive the brand into the buyer's consideration set? | Hypothesis | Reasonable from search-funnel parallels; not directly measured. Brand-tracking surveys would be the right instrument and require a multi-quarter cohort. |
| **Brand-recall persistence over time** | Hypothesis | Engine source pools drift; we observe drift across re-runs but cannot predict 12-month persistence from a single panel. |

> **Why this matters in a board deck.** Vendors who quote sales-lift numbers from GEO interventions are almost always conflating measured signals (surface) with hypothesised downstream effects (revenue). The peer-reviewed literature — Aggarwal et al. KDD 2024, Zhang Kai & Yao Jingang 2026 — measures citation lift inside generative engines, not sales attribution. If a number arrives without a named study and a sample size, it is marketing.

The honest framing for a CFO conversation: AI visibility is a brand-presence investment with the same epistemic status as paid PR or trade-press coverage. Mention rates are observable; downstream conversion is plausible by analogy to other brand-presence channels but is not directly measured by any current methodology, ours included. Aggarwal et al.'s headline finding — third-party citation roughly **6.5× more effective** than self-citation alone — describes citation lift inside the engine, not revenue.

## Three scenarios where the case is strongest

The business case is materially stronger in some shapes than others. Three patterns where we see the case land:

**Scenario 1: brand entering or re-entering the Mainland market.** A brand opening a Mainland presence is, by default, absent from the source pools of all three engines. The first prompt-panel run will return zero or near-zero brand mentions — and a non-trivial share of category-research conversations are happening on these engines now. The cost of being absent from those conversations during the launch window is the strongest version of the business case, because the comparator is "category prompts that name competitors and not us." Source-graph build-out runs in parallel with the rest of launch marketing; the discipline is the same as PR or trade-press seeding for any new market.

**Scenario 2: brand defending against a price-aggressive competitor surfaced more often.** A competitor with a lower-priced tier appears in 60–80% of category prompts where the established brand appears in 20–30% (these ratios are illustrative of the gap shape, not a specific category number). The competitor is benefiting from a denser third-party source-graph — community discussion, comparison reviews, deal-aggregator coverage. The defensive case is not about removing the competitor's mentions; it is about anchoring the established brand in surfaces with stronger framing (premium, established, durable) so the engine's framing of the category does not collapse to a price comparison. Our category-level work in [Where Chinese AI Engines Source Luxury Brand Information](https://www.eastbound.ai/blog/luxury-ai-sources.html) shows how source-graph composition shifts framing — at the ultra-luxury price tier (¥30K+ handbags), the heaviest-weight sources on DeepSeek collapse from SMZDM to The Purse Forum, Vogue Business, and auction-house archives. The framing of the category follows the source mix.

**Scenario 3: white-space window — competitors absent from these engines entirely.** The category panel returns no reliable brand mentions for the established competitor set. None of the named players have built source-graph density on the Mainland surface. This is a land-grab window — the first brand to invest in source-graph density typically establishes a default association that takes a long time for later entrants to dislodge. The window is rarest and shortest-lived; when present, it is the highest-ROI version of the case. Our research has surfaced a few categories in this shape; they tend to be specialist consumer categories where the global incumbents have not historically invested in Chinese-language brand presence.

## What spending typically looks like

The engagement shape we see most often, ordered by commitment:

| Stage | Cost | Time | Output |
|---|---|---|---|
| [Free audit](https://www.eastbound.ai/ai-visibility-audit/) | Zero | Days | Per-engine selection / absorption / mention scores plus top-5 cited sources for one URL across DeepSeek + Qwen + Doubao on a stratified zh-CN consumer prompt panel |
| Three-month measurement engagement | Mid five-figure to low six-figure (varies with category breadth and competitor set size) | ~12 weeks | Full panel (hundreds of prompts) with reliability stats; competitor matrix; per-stage gap analysis; 12-month source-graph plan; baseline + one re-run for trend signal |
| Ongoing source-graph work | Variable; in line with PR / content retainers | Quarterly cycles, multi-year | Pitched coverage, contributed analysis, community presence on the surfaces the engine repeatedly names; quarterly re-measurement against the baseline |

The reason the structure is "free audit → measurement engagement → ongoing source-graph work" rather than "single project" is that the highest-leverage work is the multi-quarter source-graph build, and the source-graph build is only worth doing once the measurement layer is in place to read whether it is moving the dial. Ordering matters: skipping the measurement layer turns the source-graph spend into untracked PR.

## Risks — what could go wrong

The honest risk list, ordered by how often we see each show up:

**Engine source pools shift.** What worked on DeepSeek six months ago may not work today. Engines re-train, source-pool weights shift, and a brand's previously-stable mention rate can move. We re-measure quarterly precisely because the surface drifts. A business case that assumes static engine behaviour is the most common cause of disappointed expectations.

**Cross-engine non-portability.** Top-15 source overlap (Jaccard) between any two Chinese engines was 0.20–0.30 in our 540-call panel. Work that lifts mention rate on one engine does not automatically transfer. The plan must address each engine separately or accept that visibility wins will be uneven across engines.

**Sales-attribution overpromise from vendors.** Other vendors in the AI-visibility category will quote sales-lift figures. Those figures are not measured by any current methodology, ours or theirs, against published peer-reviewed evidence. A business case anchored on a vendor-claimed conversion rate is anchored on something that has not been demonstrated. The defensible case is anchored on measured surface metrics plus a labelled hypothesis about downstream effect, the same way a paid-PR business case is anchored.

**Wrong engine prioritisation.** The three engines are not equally important for every category. Travel and hospitality lean toward Doubao and Qwen for consumer queries; certain technical and B2B categories lean toward DeepSeek. A plan that runs all three at equal intensity wastes spend; a plan that runs only one misses a meaningful share of the conversation. The audit answers the prioritisation question.

**Market-side risks outside the engine.** Regulatory changes affecting AI services in Mainland China, new entrants in the engine market (Yuanbao, Kimi, ERNIE Bot have all moved share), and changes in consumer adoption patterns can all reshape the surface mid-engagement. Quarterly re-measurement is the only mitigation.

## Decision-maker checklist

The questions a CMO or VP Marketing should be able to answer before committing:

1. **What share of category-research conversations in our market are happening on AI engines today?** If the share is below 5%, the case is weak; above 20%, the case is structurally strong; in between, the audit produces the data.
2. **Are our top-3 competitors named in the prompt panel and we are not?** If yes, the defensive case is strongest. If we are named and competitors are not, the offensive case is strongest. If none of us are named, the white-space case applies.
3. **Do we have the operational muscle for a multi-quarter source-graph build?** The infrastructure layer is hours; the third-party source-graph layer is the multi-quarter work where the lift compounds. Without operational commitment to the slow layer, the fast layer alone is not enough.
4. **Are we comfortable with the boundary between measured and hypothesised?** The case rests on measured surface plus labelled hypothesis on downstream effect. Stakeholders who require directly-measured sales attribution before approving spend will find this case incomplete — and they should be told that honestly upfront.
5. **Have we set the re-measurement cadence?** Engine drift is real. A business case that does not include quarterly re-measurement understates the ongoing cost.

## Related reading

- [Methodology](https://www.eastbound.ai/methodology/) — the two-stage selection-vs-absorption framework, prompt-panel design, reliability discipline, what we deliberately do not claim
- [China AI visibility for global brands](https://www.eastbound.ai/china-ai-visibility/) — the pillar reference for the Mainland-China surface
- [Free AI visibility audit](https://www.eastbound.ai/ai-visibility-audit/) — the audit that anchors stage one of the engagement shape above
- [Competitor AI visibility audit — how to benchmark](https://www.eastbound.ai/competitor-ai-visibility-audit/) — the how-to reference for the competitor matrix referenced above
- [DeepSeek vs Qwen vs Doubao: Why Brands Look Different](https://www.eastbound.ai/blog/three-chinese-ais.html) — the 540-call source-overlap study (Jaccard 0.20–0.30)
- [Where Chinese AI Engines Source Luxury Brand Information](https://www.eastbound.ai/blog/luxury-ai-sources.html) — category-level source-mix patterns at handbag / watch / luggage tiers
- Aggarwal et al., "GEO: Generative Engine Optimization" (KDD 2024) — [arxiv.org/abs/2311.09735](https://arxiv.org/abs/2311.09735)
- Zhang Kai & Yao Jingang, "From Citation Selection to Citation Absorption" (2026) — [arxiv.org/abs/2604.25707](https://arxiv.org/abs/2604.25707)

## Start with the free audit

Stage one of any business case is the data. The free Eastbound audit runs your URL against a stratified zh-CN consumer prompt panel across DeepSeek, Qwen and Doubao, and reports per-engine selection, absorption and brand-mention scores with the top cited sources per engine. Bring the output to your category review.

[Run AI visibility audit](https://www.eastbound.ai/ai-visibility-audit/) or [book a 30-minute fit check](https://www.eastbound.ai/book-consultation/).