Thai marketplace benchmarks

How this dataset was produced

A figure derived from customer data has to say exactly what it is and what it is not. This section is not an appendix — it is part of the dataset.

1. 1Every figure is an aggregate. There is no per-shop figure anywhere, and no shop name, shop id, campaign id or user id appears in any form.
2. 2Any bucket — a window, a product category, a platform — backed by fewer than 8 distinct shops is suppressed rather than published with a warning.
3. 3Panel sizes are published as bands ("80+ shops"), never as exact counts, and per-shop counts are converted to shares of the panel.
4. 4Every figure is observational. Nothing here is the measured effect of an intervention, a counterfactual replay, or a modelled gain.
5. 5Every table names the panel it came from, and every panel states its window, its grain, how a shop got into it, and what was excluded and why.

A figure that fails any of these is suppressed rather than softened and published. The full suppression register is at the foot of this page.

Measurement panels

Every figure on this page names one of these. Here is all of it.

Tables

Standalone findings

Figures that are not naturally a table row. Each names its panel and carries its own limits.

1. The paid channel captured LESS of the midnight burst than the order book did

   2.502 against 2.802

   Median lift in midnight’s share of the day, campaign days divided by ordinary days, measured two ways on the same shops: 2.502 counting only ads-attributed orders, and 2.802 counting all live orders including organic. The ads figure exceeds the attribution-free figure in only 20.7% of the panel, paired Wilcoxon p = 0.00002. The paid channel rode a burst it did not cause, and was credited for it.

   • This is a comparison of two observed series, not an experiment. It bounds how much of a demand-peak hour the attributed series can honestly claim; it does not measure the incremental effect of the ads.
   • The mechanism is the platform’s, not the seller’s: a double-date sale opens at 00:00 and buyers check out carts assembled beforehand.

   Panel:Shopee campaign-day panel · 70+

2. Almost all ad-attributed conversion lands in the click’s own hour

   ≈97% same hour; 2.9–3.3% delayed

   Estimated in dark spells — campaign-hours with zero clicks, zero impressions and zero spend — where an attributed order can only have come from an earlier click, so within-day co-movement cannot masquerade as a delay. The delayed tail exceeds an exposure-exact permutation null by 0.0019–0.0022 with a contrast of 3.26× to 5.25× and an empirical p of 0.0000; the range on the headline share is the two defensible conventions for handling gap rows.

   • This is a fact about the platform’s attribution system, not about buyers. The outcome is the platform’s own attributed-order column, and no order-to-campaign link exists, so human deferral is not observable here at all.
   • It is a no-interference upper bound. While a campaign is live, a delayed conversion from an old click is liable to be re-credited to a newer click. State it as “for campaigns that go dark, measured in no-interference windows”, never as a fleet-wide delayed share.

   Panel:Shopee dark-spell panel (delayed conversion) · 70+

3. The midnight CPC premium is near-universal

   1.46× in 87% of shops

   Hour 0 is the dearest hour of the day on the panel, at a median 1.46× the evening rate, and the premium holds in 87% of shops. The pooled best-to-worst cost-per-order ratio across the six windows is 1.75×.

   Panel:Shopee intraday ad panel · 80+

4. The clock does not vary by day of the week

   20–23 best on 7 of 7 weekdays

   On ordinary days, 20–23 is the cheapest window on all seven weekdays and 00–01 the dearest on all seven. The best-to-worst spread by weekday runs 1.74× (Monday) to 2.14× (Sunday). At shop grain the weekday argmax is still noise — a shop’s best window is identical on all seven weekdays in about one in nine shops — so a weekday-varying schedule is not supported by this panel.

   Panel:Shopee intraday ad panel · 80+

5. “Dear clicks convert worse” is not the pattern

   median ρ = +0.263, negative in 38% of shops

   Spearman correlation of CPC against CVR across the 24 hours of each shop. The median is positive: dear hours more often convert better, so paying up is more often right than wrong. But it is weak and it reverses in more than a third of shops, which is why it works as a per-shop diagnostic and not as a rule.

   Panel:Shopee intraday ad panel · 80+

6. Most shops spend more in their own dearest hours than in their own cheapest

   26.4% vs 22.5% of budget; 63% of shops

   Median share of a shop’s ad budget landing in its own six dearest hours is 26.4%, against 22.5% in its own six cheapest. 63% of shops spend more in their dearest six than in their cheapest six, paired Wilcoxon p = 0.002, and the median dearest-to-cheapest cost-per-order ratio is 2.07×.

   • This is a statement about where the money currently sits, and nothing more. It is NOT a statement about what moving that money would earn. The price of moving budget between hours is not identified on this panel, so no gain figure follows from this line and none is published anywhere in this dataset.

   Panel:Shopee intraday ad panel · 80+

7. Two categories inside one shop order at measurably different times

   2.61× the sampling-noise null

   Total-variation distance between two categories’ hourly order profiles inside the same shop: median 0.0735 observed against 0.0354 under a multinomial null drawing both categories from that shop’s own hourly distribution. Every testable shop reaches p < 0.05; Wilcoxon p = 4.8e−07. Two categories in one shop differ by about 7.4% of probability mass in when their orders arrive, and for scale two random shops differ by 10.4% — so the shop effect is only 1.4× the category effect.

   • The order clock and the bidding clock are different objects. The same panel finds no detectable category difference in the bidding clock at shop grain: every between-shop test statistic sits essentially at its label-permutation null.

   Panel:Shopee order-book clock by product category · 150+

8. What separates the daytime categories from the evening ones is basket structure, not price

   units per line ρ = −0.542; price per unit ρ = +0.172

   The intuitive story — cheap impulse buys at night, expensive considered purchases by day — does not hold. Price per unit spans a 33× range across the categories published here with no relationship to the evening share (ρ = +0.172, p = 0.48); the most expensive category has the highest evening share. Units per line is the one significant correlate (ρ = −0.542, p = 0.017; shop-cluster bootstrap 95% CI [−0.727, −0.289]). Categories bought in multiples skew to the working day; categories bought one at a time skew to the evening and midnight.

   • Four separators were tested, so the p of 0.017 survives a Bonferroni correction across them only marginally. Treat it as a strong lead rather than a settled separator, and note that the reading “restocking versus wanting” is our interpretation, not a measurement.

   Panel:Shopee order-book clock by product category · 150+

Platform mechanics

These are not benchmarks of seller performance. They are properties of the platforms themselves, established by reading their reporting and their APIs directly, and they are published because they change how every other number on this page must be read.

1. Shopee TH

   Shopee credits a conversion to the hour the ORDER is placed, not the hour of the click

   The seven-day attribution window is an eligibility window, not a timing one: a click on day t followed by a purchase on t+4 lands entirely on t+4. A past hour’s ad metrics therefore do not grow later as conversions arrive. The practical consequence is that in any hour where the platform concentrates checkout, ads-attributed orders spike and cost per attributed order falls — and the hour looks like the best money the seller spent, whether or not the ads did anything.

   Evidence: 100% of attributed-GMV rows checked (180,695 rows across every broad-attributed hour on the audited shops) sit in an hour that also carries a shop order in that same shop-local hour.

2. Shopee TH

   Shopee’s broad attribution is multi-touch and cannot be de-duplicated

   Broad credits any purchase within seven days of a click to every campaign the buyer touched. It over-counts distinct orders by roughly 1.1× to 2.9×, and because ads reporting carries no order key anywhere, the duplication cannot be removed from campaign aggregates. Any ROAS computed on this basis is an upper bound. Spend, by contrast, is exact.

   Evidence: Audited on production shops against the order tables; the over-count range is measured, and the absence of an order key is a property of the reporting surface itself.

3. Shopee TH

   A Shopee daily budget of 0 means UNLIMITED, not “off”

   The legal domain of a daily-budget write is {0} ∪ [100, ∞) in THB, and 0 is a trap: it removes the cap rather than the campaign. Anything strictly between 0 and the 100 THB platform minimum is rejected outright with a range error, so a 40 THB budget is not a small budget — it is an API error. “Stop spending” has no budget representation at all; the only way to express it is a pause.

   Evidence: Established against the live campaign API, including the specific rejection error returned for the (0, 100) interval, and confirmed by a production incident in which a “stop this campaign” push written as a budget of 0 let the campaign run uncapped.

4. Shopee TH

   An accepted daily budget is a floor, not a cap

   On a shop with Campaign Surge enabled, Shopee raises a campaign’s daily budget by 20% each time the campaign exhausts it, up to five times — a hard ceiling of 1.2⁵ = 2.48832× the number the seller wrote. A second family exists at 30% a step, 1.3⁵ = 3.71293×. The toggle is shop-level and is distinct from wallet auto-recharge, which never touches a campaign budget.

   Evidence: Read from the shop toggle endpoint and measured against realised spend. The toggle is enabled on roughly six in ten of the 60+ shops whose account state we ingest.

5. Shopee TH

   Shopee stores a daily budget at coarser precision than you send it

   A fractional budget is a number you can send but can never read back unchanged. Anything that pushes a budget and then reads the campaign back to confirm the write will see its own fractional intent against the platform’s rounded copy, and an exact-equality check will call that difference drift on a write that in fact succeeded.

   Evidence: Measured on a live campaign write: a budget pushed with four decimal places was stored rounded to two, and the read-back returned the platform’s rounded copy rather than the value sent.

6. Shopee TH

   Thai Shopee ad spend is almost entirely auto-bid, all-placement

   99.4% of campaign spend on the panel is auto-bidding and 99.6% is all-placement. This is a product constraint rather than a fleet preference: sellers cannot set a CPC bid any more. The steering lever that remains is a target ROAS, which 86.7% of spend carries.

   Evidence: THB 42.9M of campaign spend across the intraday panel, read from campaign settings.

7. Shopee TH

   Shopee’s bundle-deal redemption flag tags only about a third of qualifying orders

   The flag marks about 33% of the orders that actually meet a deal’s qualification threshold. Comparing a units-based baseline against a flag-based during-rate therefore makes a bundle deal look as though it reduced qualification. Measure both sides in units, or the sign of the comparison flips.

   Evidence: Historical deals across 60+ shops, comparing orders that meet the threshold against orders the platform tags as redeemed.

8. TikTok Shop TH

   TikTok Shop hourly ad reporting cannot support intraday analysis

   Two independent defects, either of which alone is disqualifying. Clicks are zero on all 1,669,959 hourly rows, although the sibling daily table carries 3.3M clicks over the same period — so CPC and CVR are simply unavailable at hour grain. And 94.1% of all expense is stamped at hour 0, while hours 17–23 carry 0.0% of spend and 5.2% of attributed orders. Seven consecutive hours cannot produce orders on zero spend: the hour dimension is not a real clock. Lazada publishes no hourly ad fact at all, so the cross-platform question cannot be answered there either.

   Evidence: The tell is row concentration: TikTok’s hour-0 row count is 10.80× the median hour, against 1.01× for the equivalent Shopee table, whose every hour carries the same row count and whose spend concentration at hour 0 is 9.4%. The TikTok daily totals are correct — hourly expense reconciles to daily at exactly 1.000 per shop-day — so the table is safe at day grain and unsafe at hour grain.

Definitions and arithmetic

This section is definitions, not measurements — but you need it to read the tables correctly.

Raw ROAS overstates profit-adjusted ROAS by roughly 1 ÷ margin

Profit-adjusted ROAS is margin × realised ROAS − 1, so it is centred on zero rather than on one and break-even sits at a realised ROAS of 1 ÷ margin. At a 20% margin that is 5.0: a campaign reading a raw ROAS of 4.0 is losing money on ads, at a profit-adjusted −0.2. This is an identity, not a measurement — but it is the arithmetic that decides whether any ROAS figure in this dataset means what a reader assumes it means.

Attributed order

An order the platform credited to an ad campaign, stamped at the hour the order was placed. It is not a count of orders the ad caused, and on the broad basis it is not even a count of distinct orders.

Day-window

The six blocks the day is cut into throughout this dataset: 00–01, 02–07, 08–11, 12–16, 17–19 and 20–23, on the shop’s own local clock. Hours are inclusive at both ends.

Ordinary day

Any day that is not a platform double-date campaign day. Every pooled window table in this dataset is computed on ordinary days only.

What we hold and did not publish

A dataset that shows only what survived is not auditable. Every item below is a figure we hold and chose not to publish, with the reason.

1. Every reallocation gain from moving ad budget between hours

   Reason: The gain is a pure function of a supply elasticity that is not identified. The fitted value turned out to be the share of spend variance that is price variance, and a panel with a literally zero supply curve reproduces it; alternative estimators disagree in sign. Across the plausible estimator range the implied gain spans roughly +3% to +110%, so no point on that curve may be quoted. The internal status is UNMEASURED and only a randomised paced-release test settles it.

2. Every within-model counterfactual and shadow-mode replay result

   Reason: We hold a number of headline-looking percentage improvements from model bake-offs and replays. Not one of them is an observed production outcome; each is a comparison of one model’s prediction against another’s on historical data, and our own notes label them as not causal. Publishing them as marketplace benchmarks would be dishonest whatever the disclaimer said.

3. Every per-shop result, worked case and named example

   Reason: Rule 1 of this dataset. The underlying research is full of single-shop worked cases with their own tables; none of them appears here in any form, including as an anonymised illustration, because a shop with a distinctive spend pattern is identifiable from its own numbers.

4. Everything computed on the 20 largest advertisers

   Reason: That panel was measured to be unrepresentative on three of four tested properties — the largest advertisers have the LEAST intraday variation, not the most. Every fleet claim on this page is recomputed on the full eligible population instead.

5. The evening spend-to-order pass-through slope

   Reason: A cross-sectional slope relating a shop’s evening spend share to its evening order share. It is descriptive, but its only use is to price a reallocation, and a reader would reasonably take it as the return on moving budget into the evening. It is not that, so it is withheld.

6. The bundle-deal qualification lift

   Reason: A before-and-during comparison on historical deals with no control group. Only the platform’s flag under-tagging rate — which is a property of the platform and not an outcome — is published from that panel.

7. Product categories backed by fewer than 8 distinct shops

   Reason: One L1 category clears the 20,000-order-line floor but sits below the minimum cell size, so its row is withheld entirely rather than shown with a warning. Several published categories are carried by 8 to 15 shops and their rows are therefore partly a handful of sellers’ clocks; the shop band beside each row is there so a reader can weight it accordingly.

8. Price per unit by product category

   Reason: Known to be understated by an unequal amount across categories, because bundle and free-gift lines carry a zero unit price on 100% of rows and their share ranges 0.02% to 18.0% by category. The null result that price does not separate the clock is published; the level figures are not.

9. The shape of the delayed-conversion tail beyond its headline share

   Reason: The share of the delayed tail arriving within 24 hours is reported against two different mechanical floors in two places in our own notes, and we have not reconciled them. Only the headline same-hour share, which is stable across both gap conventions, is published.

10. The minimum viable daily budget below which a campaign stops pacing

    Reason: We operate a delivery floor as a constraint on our own solvers, but it is an operating rule with an asserted rationale, not a measurement from any panel here — and in the one place it was tested against a re-solve it bound on nothing. A rule we impose is not an observation of the marketplace, so it does not belong in this dataset.