Daily campaign report — 2026-08-02, against 2026-08-01

1Overview

On the three revenue rows the change column is the change in ROAS, not in dollars. Dollars fell on all three, roughly with the spend.

The account bought less at the same price. Spend fell 62.1% and installs 62.0%, so cost per install landed within a tenth of a percent of the day before. Underneath that, CPM rose 21.1% and CTR rose 21.2% and cancelled exactly.

Booked rose while the day's own traffic earned less. 88.3% of the money arriving on 08-01 was earned by that day's installs against 48.7% on 08-02, so booked is measuring earlier buyers still paying. Within-day is sealed at 0.98 and cohort@24h agrees on direction at −18.2%.

The fall is real; its size depends on the measure. Within-day puts it at 26.8% and cohort@24h at 18.2%. The gap is the day's shape — delivery ended at 17:59 UTC, so late installs had hours rather than a day to pay, and within-day charges them for it.

No per-asset revenue row clears the payer bar, at 13, 10 and 9 payers, with one person holding 27% to 42% of each. §7 ranks them anyway, as directions.

2Hourly, 2026-08-02

Spend, impressions, clicks and installs sit on the delivery hour. Booked sits on the hour the money arrived. Within-day and cohort sit on the install hour: what that hour's traffic went on to earn. So only the within-day and cohort ratios divide an hour's revenue by that same hour's spend; the booked column is here so the day's total is traceable to where in the day it landed.

Meta's own booked total for the day is $610.49 · 2.19, on its conversion clock. Installs in hours with zero delivery are conversion-clock residue from earlier hours; there are 38 of them.

Analysis. The UTC day contains two dark blocks with two different causes. Hours 01–07 are the tail of Meta day 08-01, when the CBO ad set had already spent $505.93 against a $500/day budget by 17:59 on the account clock and stopped for the rest of its own day. Hours 18–23 are not that: the budget was unchanged at $500/day, only $271.15 of it had been spent by then, and the campaign delivered again later on the same account day, which a cap never does. Between the two blocks sits one contiguous run of ten delivering hours, 08:00 to 17:59, plus a single hour at 00:00 carried in from the end of the previous account day.

The thirteen hours that bought nothing booked $209.64, against $15.96 of within-day revenue. That is 37.5% of the day's booked total arriving while delivery was at zero, and 5.9% of its within-day total. Hour 02 alone booked $36.50 on no spend.

Within the delivering hours the buying is steady. CPM runs $8.46 to $11.95 and CTR 8.51% to 11.12%, with no trend across the block. Cost per install drifts from $0.458 down to $0.313. What the traffic did afterwards does not hold steady: within-day ROAS by hour runs 0.00 to 2.27 on materially identical delivery, and hour 11 reads 0.58 within-day against 2.97 at a 24-hour horizon on the same traffic.

That last pair is the censoring, visible in one row. Compare hours to each other on the cohort column, and one day to another on within-day.

3Country

⚠ This table is a META-day cut on its spend and install columns. Meta serves no hour × country grid, so per-country delivery exists only on the ad account's own UTC-7 day and cannot be restitched. The revenue columns are on the report's UTC clock, so every ratio in this table crosses seven hours. Read the table for mix, and read Meta's own bracketed figure when a level is wanted, because it is Meta's value over Meta's spend on Meta's day. Do not compare any level here against §1 or §4.

08-02, top by spend:

The same countries a Meta day earlier:

Analysis. The mix moved toward the rest of the world. India took $159.52 of the ad account's $505.93 on the earlier Meta day and $82.03 of $321.07 on the later one, 31.5% falling to 25.5% on spend that itself fell 48.6%. India is also the most expensive install of any country with real volume, $0.453 against Brazil's $0.207 and the Philippines' $0.163.

A single day cannot carry a per-country verdict on this account. The loudest ratios sit on the smallest cells: Australia's 3.84 rests on two installs, and Great Britain's and Italy's zeroes on five each. The whole-run estimator answers the same question with a usable sample, on Meta's clock across 07-30 to 08-03 and 4,527 installs at a 12-hour horizon. Against a pooled account value of $0.463 per install it puts India at 1,283 installs and 0.87× the account's return, Brazil at 370 installs and 0.71×, Indonesia at 279 and 0.57×, the Philippines at 198 and 0.66×, Malaysia at 116 and 0.65×. The two countries that clear the account on that sample are the United States at 120 installs and 2.32× and Turkey at 119 and 1.78×.

That reframes the day's two loudest cells. Brazil's 7.33 in the cohort column is 77 installs of a country that returns 0.71× over five times the sample; Australia's 3.84 is two installs. Neither is a market fact.

4Aggregate, both days

Pooled over the 10 hours both days delivered in, on the UTC clock. Meta's instrument on both sides of every ratio.

• click → install 25.89% → 24.50% (-5.4%, p = 0.228) — did not separate.
• CTR 7.55% → 9.90% (+31.2%, p = 0.000) — rose.

Pooling across hours is a mix comparison. It weights each hour by how much it delivered, so a day that bought differently across the clock can move these rows without any hour changing. The hour-by-hour pairing in the hourly section is the unweighted read of the same two days; where the two disagree, the mix moved.

On this pair of days the two reads agree. Pooled, CTR goes 7.55% to 9.90% at p = 0.000 and click-to-install 25.89% to 24.50% at p = 0.228; counted hour by hour in §2, CTR runs higher in nine of ten at p = 0.021 and click-to-install lower in seven of ten at p = 0.344. Both readings put CTR up and both leave click-to-install unresolved, and the pooled CPM and cost-per-install rows move in the direction their sign tests already gave. No single hour on either side of the pairing dominates the compared impressions, so the weighted and the unweighted read have nothing to pull apart.

Held to a single creative that ran on both days, the comparison can only be made on the ad account's own clock: the matched-hours work keys every cell on (campaign, ad set, ad, Meta day, Meta hour), and its shared-hours logic means something only against Meta's own delivery days. What follows is Meta day 08-02 against Meta day 08-01, a seven-hour-offset pair of windows against every other table in this report, and its levels do not reconcile against them. Only the nine hours both Meta days delivered are used, because the diurnal shape is the largest confound: 01, 02, 03, 05, 06, 07, 08, 09, 10.

Install counts here are cell-attributed with carry-back, which is a third basis from either Meta's Leads or a raw registration timestamp. Cohort coverage at H=24 on this subset is 100% on both days.

bikini's shared window is four hours, 02, 03, 05 and 06; bodysuit's is all nine.

Nothing in that pooled funnel separates the days either. Click-to-install falls 6.0% at p = 0.156 and the payer rate falls 7.3% at p = 0.767. Held to bodysuit, click-to-install rises 7.4% at p = 0.230 and the payer rate falls 39.4% at p = 0.242, on seven payers.

bikini's click-to-install drop is basis-dependent and should not be quoted as settled. On the cell-attributed basis above it reads 31.36% → 22.64%, a 27.8% fall at p = 0.004. Recomputed over the same four hours on Meta's own clicks and installs it reads 25.72% → 20.75%, a 19.3% fall at p = 0.083. One creative, one funnel step, two defensible bases, two verdicts either side of the 0.05 line, on 72 installs. The sample is too thin to settle it, and bikini's payer rate moved 99% the other way over the same cells at p = 0.162.

So the funnel after the click is not meaningfully different between the two Meta days. What moved is the price of the impression.

Back on the report clock, the two days on all three measures:

Booked revenue split by the UTC day its payer actually installed:

Analysis. The top row of that split table is within-day. $980.99 and $272.13 appear in both places. Within-day is the day's own share of its booked revenue, so the gap between the two measures is the inherited cohort and nothing else.

On 08-01 the two nearly agree, 1.51 against 1.34, because 88.3% of what arrived was earned that day. On 08-02 they diverge hard, 2.01 against 0.98, because less than half was. A day that spends less while the previous days' buyers keep paying will always read well on booked revenue, and better still on Meta's version of it. That is the measure behaving as defined.

Two things bound how far the 0.98 can be pushed:

• Within-day censors by hour of day, and 08-02's shape makes that bite. Its delivery ends at 17:59 UTC, so its last installs had seven hours or fewer to pay before midnight. The 24-hour cohort corrects exactly this, and it lifts the day 32.5% above within-day against only 18.6% on 08-01. Corrected, the fall is 18.2% rather than 26.8%. It does not disappear.
• Payer counts are small. 32 within-day payers on 08-02 against 83 on 08-01, and the per-asset rows are all under 16. Every asset-level ratio here is a direction.

Maturity is not available as an explanation on either side: both cohorts are 100% covered at a 24-hour horizon, the youngest 08-02 install being 61.4 hours old at the payment cut.

5Per ad set, hourly

This campaign runs one CBO ad set holding three ads, so the ad-set row would be the whole account and the level that separates is the ad; the rows here are labelled by creative.

Delivering hours only, single-clock metrics, on the UTC day. bodysuit delivered in 11 hours, pool5 in 10 and bikini in 9; all three ran together in the nine hours 09:00 to 17:59.

The two creatives that ran on both UTC days:

Analysis. pool5 buys the most expensive impressions and converts them best; bodysuit the reverse. Against bodysuit, pool5 pays 31.8% more per impression, earns 62.7% more clicks from them, and turns 21.2% fewer of those clicks into installs. The three effects very nearly cancel and it arrives 2.8% behind on cost per install. bikini is a clear third at $0.437, paying nearly pool5's impression price for two thirds of its click rate.

Day over day, the two returning creatives moved differently. bodysuit's impression price rose 9.2% and its click rate slipped 2.4%, but it converted 11.3% more of its clicks, so its cost per install landed within 0.6% of the day before. bikini's impression price rose 30.0% and its click rate rose 28.8%, cancelling, and it paid 5.6% more per install. The account's flat pooled cost per install is therefore two creatives whose own moved 0.6% and 5.6%, plus pool5 entering at $0.362 on 27.7% of the spend. It is not an average concealing movement underneath.

6Per ad set × country

⚠ Spend and installs are the META-day cut described in §3; revenue is on the UTC day. Cells above $3 of spend, in spend order.

Analysis. India is the biggest line for all three creatives and the three read 0.30, 0.48 and 1.73 in it at a 24-hour horizon. Brazil reads 11.18, 6.46 and 1.41. Both spreads are as wide as the spreads between countries, on one or two payers per cell, so at this cell size neither asset nor country is separable and the table is here to locate revenue, not to rank.

Locating it is what it does well. Each creative's within-day figure leans on one person. bodysuit's largest single payer is 42.3% of its $100.67, which is $42.58, the entire within-day total of hour 09. pool5's largest is 33.6% of its $105.65, which is $35.53, its entire Brazil cell. bikini's largest is 27.3% of its $65.81, which is $17.98, its entire United States cell. Three creatives, three single baskets sitting under three headline numbers.

bodysuit's Brazil row also shows the censoring directly: $10.26 within-day against $76.37 at 24 hours. Most of what that cell earned was paid after UTC midnight and inside a day of the buyer arriving.

7Per ad set, aggregate, all three measures

The three tables below are the standing per-asset block, emitted by report_tables.py and pasted unedited. They carry the same columns in the same order in every report in this series, so any two days can be read against each other.

A. Per asset — delivery

Meta's instrument on both sides of every ratio. Our registrations never enter this table.

B. Per asset — the three revenue measures

Ours on every column. Meta can express only booked, so it rides bracketed beside that column and never as a row. Payers are distinct people, never transactions. Cohort horizon H=24h, 100% covered on this day. Report clock UTC.

The 16-payer bar (#37): any row above whose payer count is under 16 is a direction and never a magnitude.

C. Per asset — country mix

⚠ This table is a META-day cut, not a UTC one. Meta serves no hour × country grid, so per-country delivery exists only on the ad account's own UTC-7 day and cannot be restitched. It is a seven-hour-offset window against tables A and B above; read it for mix, never for level against them.

Why a row in B is readable or not. The last column is the share of that asset's own within-day revenue that came from the United States: where it far exceeds the US spend share, the asset bought a geography rather than earning a result.

Analysis. Which creative is best depends on which measure is read, and the answers do not agree.

Booked revenue on the day puts bodysuit far ahead, 2.71 against 1.37 and 1.43, and Meta reads it higher still at 3.14. Its own traffic puts it last on within-day at 0.78. pool5 shows why: its booked and within-day figures are the same number, $105.65. It bought nothing at all on 08-01 UTC, and none of the buyers it picked up on 07-30 and 07-31 paid inside this UTC day, so every dollar credited to it came from someone who installed that day. bodysuit and bikini both delivered on the two UTC days before this one, and 71.3% of what bodysuit was credited with, against 36.2% of bikini's, was paid by people who installed earlier. On any day the spend falls, booked revenue promotes whichever creative holds the largest stock of earlier buyers still paying.

The hour-neutral measure narrows the gap rather than confirming it. At a 24-hour horizon pool5 leads at 1.46, bodysuit follows at 1.36 and bikini trails at 1.01. bodysuit's jump from 0.78 to 1.36 is the same censoring §4 describes: it delivered latest into the UTC day, and its Brazil cell alone goes from $10.26 within-day to $76.37 at 24 hours.

8Caveats

• Payer counts are the binding constraint. 32 within-day payers on 08-02 against 83 on 08-01. Every per-asset row is under the 16-payer bar, and in each case one person is 27% to 42% of the creative's whole within-day revenue. Country and asset×country cells routinely carry a single payment: $4.99, $5.13 and $5.89 appear as whole cell totals throughout §6.
• Maturity is not a caveat on this cut. Both days' 24-hour cohorts are 100% covered at the 2026-08-05 10:59 UTC payment cut, the youngest 08-02 install being 61.4 hours old.
• Within-day is censored by hour of day and 08-02's shape makes it bite, since delivery ends at 17:59 UTC. The 24-hour cohort removes the censoring and the day still falls 18.2%.
• The country tables and the region charts are a Meta-day cut on their spend and install columns, seven hours offset from every other table here, with UTC revenue over Meta-day spend. They carry mix, not level.
• §4's per-creative tests and the regime table are on Meta's clock by construction and describe Meta days and Meta regimes, not the UTC day the rest of this report covers.
• Revenue cannot settle a creative test at this volume. Revenue per install has a coefficient of variation of 6.49 across the run, so a 20% resolution needs 16,526 installs per arm. The largest regime-matched arm is bodysuit in R5 at 1,430 installs, 9% of that. Every ROAS here is descriptive.
• Two dark blocks split the UTC day's delivery into a single hour at 00:00 and a ten-hour run from 08:00, so every day-level figure here pools a block that had run out of budget on the previous account day, a block with no delivery at all, and one full delivering block.
• pool5 did not deliver on 08-01 UTC, so it has no day-over-day row.

9What would settle the open questions

10What this hands to E2

The CBO ad set stopped delivering at 04:20 on 08-03, Meta clock, and was replaced by E2, a randomised creative test running one creative per ABO ad set.

The baseline to beat is within-day ROAS 0.98 at $0.374 cost per install on the UTC clock, or 1.30 at a 24-hour horizon. It is not the 2.01 booked or the 2.19 Meta books. Every E2 arm is brand new and therefore has no inherited cohort, exactly as pool5 collected none on 08-02, so comparing an E2 arm against a booked figure from this campaign compares a flow against a stock and will make E2 look worse than it is. Compare within-day to within-day, and cohort at 24 h to cohort at 24 h. India stays a quarter of the buy at 0.87× the account's value, and no creative test addresses that.

11Appendix — the clock, the measures, and regime

The report day is a UTC day, to line up with Adjust. Meta stamps its hourly exports in the ad account's own zone, UTC-7, so every delivery figure above is restitched from two account days: 08-01 17:00–23:59 and 08-02 00:00–16:59.

Three revenue measures run through every table and are never mixed.

All three are read on our own payment records, with Meta's figure bracketed beside booked and never given a row of its own. The instrument check between the two runs on Meta's clock, where its conversion value can be read against the day it belongs to; there our records are 0.94 and 0.91 of Meta's booked value on the two account days, inside the documented 0.91–1.06 band.

The two measures count different install populations. Meta attributes 1,963 and 745 installs; our own records tie 2,134 and 795 registrations to these ads. Every ROAS in this report is revenue over spend and uses neither count. The install and cost-per-install rows are Meta's.

A regime is a stretch of time over which the ad set's daily budget did not change. It matters because the budget is a geography dial on this account: raising it buys wider, cheaper countries, so two creatives measured across a budget change are being compared on different audiences.

These two days are not regime-matched, and that bounds every comparison in the report. 08-02 sits wholly inside one regime, R5 at $500/day, which ran 08-01 04:10 to 08-03 04:20 on Meta's clock. 08-01 does not: its window opens on the tail of an earlier $500/day setting, runs at $875/day, and returns to $500/day at 11:10 UTC, with $435.60 of its spend falling before that change and $298.34 after. That is a consequence of the clock rather than of the buying. Regimes are read from Ads Manager's activity history and cannot be recovered from any export.

12Reproduce

Report day 2026-08-02 00:00 → 24:00 UTC, restitched from Meta account days 08-01 17:00–23:59 and 08-02 00:00–16:59, the account being fixed at UTC-7. A UTC pull therefore needs the previous Meta day's hourly export as well as the day's own. Payment records cut 2026-08-05 10:59 UTC, which gives 08-02's installs 61.4 to 82.9 hours and 08-01's 83.1 to 106.9 hours, so both 24-hour cohorts are complete. Country figures come from one account-wide Country × Day pull and are a Meta-day cut.

Meta rows are selected by Ad ID, not by ad-set or creative name: E2 carries byte-identical creative names to this campaign's, and its first ad set was itself named ww_broad_purchase_adv_cbo150 for ten hours on 08-02–08-03. A name filter over-attributes by ~70%.

cd C:/Projects/VideoGenAI_Related
CFG=_runs/2026-08-05_0804_settled/run_config.json          # tz: UTC
python -m ad_ops.measures      --config $CFG --day 2026-08-02 --prev 2026-08-01
python -m ad_ops.cohort_arpu   --config $CFG --days 2026-07-31,2026-08-01,2026-08-02
python -m ad_ops.report_tables --config $CFG --day 2026-08-02 --prev 2026-08-01
python -m ad_ops.report_charts --config $CFG --day 2026-08-02 --prev 2026-08-01 \
    --days 2026-07-31,2026-08-01,2026-08-02 \
    --out ad_ops/figures/2026-08-02 --relpath figures/2026-08-02

# META clock — the instrument gate, §4's per-creative tests, and the regime table
python -m ad_ops.measures --config _runs/2026-08-05_0804_settled/run_config_meta.json \
    --day 2026-08-02 --prev 2026-08-01
python -m ad_ops.day_compare --config $CFG --today 2026-08-02 --prev 2026-08-01 --h 24 \
    --campaign bailingxia_meituan_ww_cvr_260730 --ads ww_bikini_10s,ww_bodysuit_10s,ww_pool5
python -m ad_ops.asset_estimator hourly --horizon 12 \
    --meta ad_ops/_data/b1_meta_exports/B1-Ads-Jul-30-2026-Jul-30-2026.csv \
           ad_ops/_data/b1_meta_exports/B1-Ads-Jul-31-2026-Jul-31-2026.csv \
           _runs/2026-08-03_aug2_hourly/exports/ads_hourly_aug1_settled.csv \
           _runs/2026-08-03_aug2_hourly/exports/ads_hourly_aug2_settled.csv \
           _runs/2026-08-04_e2_start_and_0803_report/exports/ads_hourly_aug3_settled.csv \
    --lum _runs/2026-08-05_0804_settled/lum \
    --fx  _runs/2026-08-02_today_read/fx_rates.json \
    --ads ww_bikini,ww_bodysuit,ww_pool5

report_tables --prev emits §2's matched-hours pairing and §4's pooled funnel as well as §7's three standing tables. report_charts writes all nineteen charts and a per-section block of image links.

The measures run raises by name if the previous Meta day's hourly export is missing, rather than short-filling the first seven hours of the UTC day. The 08-03 hourly file is required by the estimator for the same reason in the other direction: the campaign delivered past midnight, and without it attach_cohorts carries every post-midnight install back into the last 08-02 cell.