The new competitive world of commerce

In October 2025, the Google–Temasek–Bain e-Conomy SEA report projected Southeast Asia's digital economy to surpass US$300 billion in gross merchandise value by year-end. E-commerce alone is on track for roughly US$185 billion GMV and US$41 billion in revenue, and video commerce — live-stream selling, creator-driven content commerce, short-form video shopping — has grown to account for approximately 25% of regional e-commerce GMV, a share that did not exist in 2020 and was below 10% as recently as 2022. The macro trajectory is unambiguous. The seller-side trajectory has been moving in the opposite direction — at the same regional growth rate, the per-unit margin available to the operator below the top decile has compressed materially over the same window.

The compression is not the result of individual seller mistakes or category-specific cyclicality. It is the result of a structural transition in how the platforms operate, documented in Sea Limited's 2025 investor disclosures, in Alibaba's SEA segment filings, in the platforms' own seller-facing documentation, and in the Reuters reporting that has tracked the platform-side AI investment cadence through 2026. The transition has a name when stated plainly: SEA marketplaces stopped being passive listing markets and became active algorithmic counterparties. Visibility, ad placement, recommendation surfacing, live-commerce promotion, and creator-affiliate matching are now run by ML systems the platform optimizes continuously and the seller cannot see.

This note documents the transition: what specifically changed in the platform stack between 2022 and 2026, what the seller-side translation of those changes looks like in operating terms, why a seller running 2022 operating habits in 2026 is competing against an optimization system rather than against other sellers, and what "running your own optimization layer on top" means at the SKU level. The argument is empirical where it can be (citing the platforms' own filings and the public reporting) and acknowledged where it cannot be (the per-account margin distribution we describe is drawn from the accounts we model directly; the methodology section at the end describes the sample and its bounds).

Sellers are no longer just competing with other sellers. They are competing with the platform's own optimization systems — and the systems get better every quarter.

Thesis: in a marketplace where the platform is an active algorithmic counterparty optimizing for its own take-rate, the seller who keeps margin is the seller who runs a per-SKU optimization layer on top of the platform's — true ROAS per campaign per cell, contribution margin per SKU per channel, pre-launch margin simulation against documented campaign mechanics, and post-window reconciliation against margin instead of GMV. The seller who runs no such layer is, by construction, an input to the platform's optimization — not an operator on the platform. The fix is mechanical; the obstacle is administrative; the moat that follows is durable precisely because the obstacle is hard to outsource.

What changed: the 2022→2026 trajectory

The seller-side experience of operating on Shopee, Lazada, or TikTok Shop in 2022 differs structurally from the same operating surface in 2026. The change is not gradual; it is a step-function shift driven by three concurrent platform-side investments that compounded over the trailing 36 months. The table below captures what those changes look like in the variables an operator actually controls or experiences.

Two observations from the table deserve explicit attention. First, the changes are concentrated on the cost side — ads, voucher tier, shipping subsidy, gross-margin compression on scaling — rather than on the revenue side. The platform has not constrained the seller's ability to grow GMV; it has changed what GMV growth costs in margin terms. Second, the velocity gap (between the platform's decision frequency and the seller's) is the most under-discussed of the variables. A seller making weekly operating decisions on a platform whose recommendation system reranks every impression in milliseconds is, in cybernetic terms, operating at a decision-frequency mismatch that compounds against them every campaign window.

What the platforms actually built — three documented investments

1. Sea Limited's AI infrastructure investment.

Sea Limited's 4Q25 earnings release and the Reuters reporting that followed on 3 March 2026 documented that the company's operating expenses rose materially during 2025 as Sea invested in AI infrastructure for Shopee — search relevance, recommendations, and advertising auctions. The investment is not a one-off; Sea's 1Q26 disclosure confirmed continued AI capex through the quarter, with the company explicitly framing the investment as a competitive necessity rather than a roadmap option. The seller-side translation: every product impression on Shopee in 2026 is the output of a system that has been actively tuned for relevance and predicted purchase probability — not a passive ranking by listing freshness.

2. The Google–Sea AI partnership.

On 19 February 2026, Reuters reported that Google and Sea announced a partnership to develop AI tools for e-commerce, including an agentic shopping prototype designed to handle browsing and checkout on a buyer's behalf. The agentic dimension is the operationally significant part — an AI shopping agent that browses and purchases on the buyer's behalf does not engage with listings the way a human buyer does, and the seller-side optimization implications (which listings the agent surfaces, which prices it accepts, which voucher mechanics it triggers) are different in shape from the human-buyer optimization that has dominated the platform stack to date. The prototype is, at the time of writing, in early stages; the directional implication for the seller is that the buyer is increasingly an algorithm, and the listing is increasingly an input to a multi-step agent decision rather than a direct human conversion target.

3. Parallel Alibaba and ByteDance investment.

Lazada's parent group has made parallel investments in ranking, recommendation, and Sponsored Solutions optimization, documented in Alibaba's SEA segment commentary across 2025 and 2026. TikTok Shop's ranking model is, by ByteDance's public statements, the same model that ranks the For You feed, with marketplace economics (price, COGS-aware affiliate matching, returns probability) layered on top. The implication is that the algorithmic counterparty in SEA marketplace commerce is not a single platform — it is three platforms operating distinct but structurally similar AI stacks, each optimizing for the platform's own take-rate against a seller surface that is largely shared.

The seller-side translation

For an operator, the platform-side AI investment translates into three concrete seller-side experiences: cost trajectory, decision velocity, and signal opacity. Each deserves a brief unpacking because the operating response to each is different.

Cost trajectory.

The same impression share costs more every quarter. The mechanism is the second-price ad auction: bid pressure rises as more sellers chase the same impressions, and the platform's relevance model gets better at extracting marginal willingness-to-pay from the ad inventory. The seller does not see the relevance model; the seller sees the cost-per-click rising at a rate that exceeds the underlying inflation. Pacing the ad budget against this trajectory requires seller-side modeling of break-even ROAS per SKU, because a flat ROAS target across the catalog will be wrong somewhere every day under a moving auction.

Decision velocity.

The platform makes thousands of micro-decisions per impression. The seller, in 2022, made decisions weekly. By 2026, the seller making weekly decisions is reliably out-of-phase with the platform's campaign-window dynamics — campaign voucher tiers escalate within days, ad auction inflation peaks during the first 48 hours of major campaigns, and the platform's recommendation model adjusts SKU exposure within hours of seeing the first conversion data. Operating at weekly cadence in this environment is the equivalent of a chess player making one move per opponent's ten — the position degrades structurally, not because the moves are wrong, but because the move count is wrong.

Signal opacity.

The platform-side dashboards report on the metrics the platform optimizes — ad-attributed revenue, click-through rate, GMV by campaign — but not on the metrics the seller needs (post-attribution contribution margin, true ROAS net of vouchers and shipping subsidies, per-SKU break-even ROAS). The signal the seller can read in the platform UI is, by design, downstream of the seller's actual operating question. Reconstructing the answer from order-line data is a seller-side responsibility that has gone from "useful for top-decile sellers" in 2022 to "the table-stakes of operating competently" in 2026.

Three worked examples — same SKU mix, three trajectories

The three formula blocks below trace the same hypothetical Thai Shopee account — initial monthly GMV ~THB 500K, gross margin 35%, baseline ad spend ~8% of revenue — under three operating regimes. Example 1 is the seller running 2022 operating habits in 2026 (no per-SKU economics, no break-even ROAS targeting, default platform Target ROAS, full campaign participation). Example 2 is the same account scaled the same way but running per-SKU optimization on top (true ROAS per SKU, break-even targeting, selective campaign participation, voucher tier matched to category margin). Example 3 is a counter-factual that holds the operating regime constant but moves the ad-auction inflation from current 1.4× to 2.0× — the trajectory the platform-side AI investment makes plausible by 2027.

Baseline:    THB 500,000 GMV · 35% gross · 8% ads · 3% normal voucher
                     net contribution margin: ~21% of GMV  →  ~THB 105,000

Scaled up:   THB 800,000 GMV · 25% effective gross · 14% ads (auction-inflated)
                  · 9% campaign voucher · 4% free-ship subsidy
                     net contribution margin: ~3% of GMV  →  ~THB 24,000

Outcome:  GMV grew 60%; net contribution profit fell 77%.

Baseline:    THB 500,000 GMV · 35% gross · 8% ads · 3% normal voucher
                     net contribution margin: ~21% of GMV  →  ~THB 105,000

Scaled up:   THB 720,000 GMV · 33% effective gross (selective participation)
                  · 9% ads (true-ROAS-pruned) · 4% campaign voucher (margin-tier-matched)
                  · 2% free-ship subsidy (selective opt-in)
                     net contribution margin: ~17% of GMV  →  ~THB 122,000

Outcome:  GMV grew 44% (less than example 1); net contribution profit grew 16%.

Apply the example-1 operating regime under 2.0× ad-auction inflation:
   Ads as % of revenue (scaled):      ~20% (vs. 14% today)
   Auction-inflated CPC pressure:     +43% on the ad-cost line
                     net contribution margin: ~ −2% of GMV  →  ~THB −16,000

Apply the example-2 operating regime under 2.0× ad-auction inflation:
   Ads as % of revenue (scaled, true-ROAS-pruned):  ~12%
                     net contribution margin: ~14% of GMV  →  ~THB 100,000

Three observations from the worked examples. First, example 1 is the dominant pattern across the accounts we model — the operator follows the platform recommendations, scales when the dashboard rewards scaling, and absorbs the margin compression as a cost of doing business. The compression is not catastrophic in any single quarter; it compounds over the year. Second, example 2 produces less GMV growth but more profit, which is the trade the operating-layer-aware seller is consciously making — they are leaving GMV on the table that costs more in margin to capture than it produces in profit, and they are reallocating that capacity to higher-margin operations. Third, example 3 illustrates the asymmetric exposure: under a plausible 2027 ad-auction inflation, the example-1 regime becomes unprofitable on the increment, while the example-2 regime stays profitable because the true-ROAS pruning is itself a buffer against ad-auction inflation. The operating layer is not just a margin-recovery tool; it is a structural risk-reduction.

Sensitivity — where the conclusion changes

The case for running a per-SKU optimization layer is strongest under high ad-auction inflation, dense campaign calendars, and significant cross-platform fragmentation — i.e. the conditions documented across the SEA-6 in 2026. The case is weaker under stable auctions, light campaign calendars, and single-platform operations. The table below stress-tests the worked examples across three operating-environment scenarios to make the boundaries explicit.

What "running your own optimization" means in practice

The phrase "run your own optimization layer" is abstract until specified. Operationally, the layer that holds margin in 2026 is a small number of concrete behaviours, repeated at sufficient cadence. Each is implementable with disciplined spreadsheets if the catalog is small enough; each becomes infeasible by spreadsheet past roughly 50 SKUs across multiple shops, which is where dedicated tooling earns its keep.

• Reconstruct contribution margin per SKU per cell from order-line data — not from the platform P&L summary, which aggregates platform-funded and seller-funded discounts and elides program-specific fees.
• Compute break-even ROAS per SKU = 1 / contribution margin rate. A single account-wide ROAS target is, by construction, wrong somewhere in the catalog every day.
• Audit ads at 14-day cadence using true ROAS per SKU per campaign. Pause keywords whose true ROAS sits below the SKU's break-even, redistribute saved budget to top-decile keywords on top-quartile-margin SKUs.
• Run pre-launch margin simulations on every campaign window that requires participation (Pay Day, 9.9, 11.11, 12.12, Mega Sale, Brand Mega Offer) — at the campaign's documented eligibility discount, voucher tier, and expected ad-auction inflation. Decline the campaigns whose simulation lands negative.
• Reconcile campaign performance 14 days post-window against attributed contribution margin, not against GMV. The platform dashboard reports GMV; the operating decision next quarter rests on margin.
• Refresh COGS, returns reserve, and fee schedule per category quarterly — these are the inputs that drift, and the worked-example arithmetic is only as good as the input freshness.

The operating layer is not a margin-recovery tool. It is a structural risk-reduction against an ad-auction trajectory the seller does not control.

Limitations and where this argument breaks

The argument has explicit scope.

• Account-size lower bound. The optimization-layer framework assumes operating capacity to maintain order-line economics, run pre-launch margin simulations, and reconcile per-window. Below approximately THB 200K monthly revenue, fixed operating overhead dominates and simpler heuristics outperform: pick two or three best SKUs and watch cash flow.
• Account-size upper bound. Top-decile sellers operate with negotiated commission, custom voucher arrangements, and direct platform relationships. The framework still applies; the input values shift, often favourably (lower commission, higher placement weight per voucher dollar). Recalibrate against the negotiated terms rather than the public-fee schedule.
• Cross-border sellers. China-to-SEA, Hong Kong-to-SEA, and Korea-to-SEA cross-border sellers face a different fee, customs, and platform program stack. The operating principle holds; the numerical bounds shift.
• Time horizon. The 2022→2026 trajectory is documented; the 2027 projection in worked example 3 is a directional extrapolation of platform-side AI investment cadence and is not a forecast. If Sea, Alibaba, or ByteDance moderate their AI capex (or pass efficiency gains to sellers via reduced auction extraction), the 2027 ad-auction inflation could be lower than projected. Forecast confidence beyond 24 months is low.
• Internal-data scope. The "in our data" claims are aggregated across the SEA-6 Thai accounts we model directly (described in the methodology section). They are not population claims about all SEA marketplace sellers; they explicitly exclude the bottom and top of the size distribution noted above and do not claim to characterise the cross-border or LazMall enterprise tiers.
• Platform-side counterfactuals. The argument assumes the platforms continue to operate as algorithmic counterparties optimizing their own take-rate. A regulatory or competitive disruption that constrains platform extraction (e.g. a SEA-wide marketplace fee regulation) would change the operating math materially. None of the public reporting we cite suggests such a constraint is imminent at the time of writing.

Methodology

Public-data citations in this note are taken from the Bain e-Conomy SEA 2025 report (October 2025), Sea Limited's 4Q25 (filed February 2026) and 1Q26 (filed May 2026) investor disclosures, Alibaba's SEA segment commentary across the 2025 and 2026 reporting periods, the Reuters reporting from 19 February 2026 on the Google–Sea AI partnership and from 3 March 2026 on Sea's rising operating expenses tied to AI investment, and the platforms' own seller documentation: the Shopee Help Center, Lazada Open Platform, TikTok Shop Seller University, and the Sponsored Solutions seller portals for Shopee Ads and Lazada Sponsored.

Internal-data claims — the cost-input ranges in the worked examples, the "in our data" margin-gap distributions, the ad-share trajectory and decision-velocity figures in the variables-table — are aggregated across the SEA-6 marketplace seller accounts that DataGlass models directly. The current sample is approximately 400 active accounts, predominantly Thai SEA-6 sellers across the DataGlass research methodology sample frame (Jan 2024 – Apr 2026, 28-month observation window).

The methodology section exists for one purpose: to make every numerical claim in the note inspectable in principle. A reader who disagrees with the conclusions should be able to point to the input that is wrong (the public-data citation, the sample, the cost-input range, the attribution model) rather than to the conclusion itself. This is the difference between analysis that can be argued with and analysis that cannot.