A hidden 'soft floor' was turning second-price into first-price on 30% of wins

A buyer on a nominally second-price exchange noticed they often paid close to their full bid. Log-level data exposed why.

Second-price (winner pays one cent above the runner-up bid). Soft floor (a publisher minimum that, when no bid clears it cleanly, makes the winner pay their own bid instead).

1. They computed, per win, the ratio of price paid to bid submitted.
— On 30% of wins the ratio exceeded 0.95, meaning they paid nearly their full bid.
2. Those wins shared one trait: no competing bid cleared the publisher's undisclosed soft floor.
3. With no valid runner-up, the auction fell back to charging the winner's bid, effectively first-price.

Evidence: lowering bids by 8% on those specific placements cut paid CPM 11% while win rate held at 26%, because the bids still cleared the soft floor with room to spare.

Why it matters: 'second-price' describes the rule, not the outcome. When thin competition leaves no clearing runner-up, soft floors convert the auction to first-price and you pay your bid. Measure price-to-bid ratio to find where that is happening.

ChatGPT 5.6 Luna и Terra подешевели

OpenAI резко снизила цены на GPT-5.6 Luna и Terra спустя три недели после релиза, сделав их заметно дешевле для массового использования. Sol осталась премиальной по прежнему прайсу, но получила Fast mode с ускорением до 2,5 раза. Вывод: компания давит ценой и скоростью, чтобы быстрее нарастить спрос и долю рынка.

➡️ Читайте на сайте: https://aff.top/blog/chatgpt-5-6-luna-i-terra-podesheveli

🧠 Ещё больше инсайтов → в канале AFF.top

Доменная зона .web делегирован в корневую зону DNS

Verisign вывела .web в корневую зону DNS: теперь это полноценный TLD, но открытая регистрация ещё не стартовала. Сначала доступ получат владельцы совпадающих доменов в .com через LRP. Вывод для рынка: хорошие EMD в .com могут стать входным билетом в .web, если успеть занять брендовые имена раньше общего запуска.

➡️ Читайте на сайте: https://aff.top/blog/domennaia-zona-web-delegirovan-v-kornevuiu-zonu-dns

🧠 Ещё больше инсайтов → в канале AFF.top

В Facebook Ads появился раздел «Conversations»

Facebook Ads добавил Conversations с автоответом на комментарии: по ключевым словам можно сразу отправлять сообщение в личку. Для арбитража это новый способ прогрева и передачи ссылки без клоаки: в креативе можно просить оставить комментарий, а заинтересованных уводить с вайта на блэк уже в ДМ. Идея спорная, но её стоит тестировать.

➡️ Читайте на сайте: https://aff.top/blog/v-facebook-ads-poiavilsia-razdel-conversations

🧠 Ещё больше инсайтов → в канале AFF.top

A fee audit on one SSP recovered 14 cents of every dollar to working media

Supply path economics are not only about CPM; the take rate (the share of spend an SSP keeps as fee) varies sharply by path. One buyer measured it directly.

1. Under a DSP that exposed gross-versus-net spend per SSP, they computed the realized take rate per path.
— Path A (direct SSP): 11% take rate.
— Path B (reseller-fed SSP): 25% take rate, for the same publisher and audience.
2. They confirmed both delivered comparable viewability and invalid-traffic rates.
3. They shifted 70% of the publisher's budget from Path B to Path A.

Evidence: working media (spend that actually reaches the publisher and buys impressions) rose by roughly 14 cents per dollar on the reallocated budget, increasing won impressions 11% at flat total spend.

Why it matters: two paths to the same impression can have wildly different take rates. CPM tells you the price; the fee tells you how much of your dollar becomes inventory. Measuring net-to-gross per path is how you find the leak.

A daypart bid multiplier rebuilt from log data lifted conversions 18%

A performance buyer used flat bids around the clock. Log-level analysis showed clearing prices and conversion rates moved on opposite schedules.

1. They bucketed every auction by hour and computed two curves: median clearing price and post-click conversion rate.
— Conversion rate peaked 22:00 to 01:00; clearing prices peaked 18:00 to 21:00 (prime-time competition).
2. The two peaks were offset by several hours, so flat bidding overpaid during expensive low-converting hours and underbid during cheap high-converting ones.
3. They built hourly bid multipliers: -20% in the 18:00 to 21:00 window, +25% in 22:00 to 01:00.

Evidence: at flat total spend, conversions rose 18% and cost per acquisition fell 15%, because budget moved into hours where the impression was both cheaper to win and more likely to convert.

Why it matters: clearing price and conversion value rarely peak at the same hour. A single flat bid averages over that mismatch. Daypart multipliers, built from your own log-level curves, buy the hours where the two align in your favor.

Removing 7 of 18 header-bidding partners raised yield 9%

More demand partners is assumed to mean more competition. A publisher tested that assumption and found the opposite past a point.

1. They logged each of 18 header-bidding partners' bid rate, win rate, and average bid over 30 days.
— 7 partners had win rates under 0.5% and bid below the floor 90% of the time.
2. Those partners added latency to every auction (each bidder call extends the timeout window) without ever setting the price.
3. They removed the 7, narrowing the auction to 11 active bidders.

Evidence: average auction latency fell 280ms, letting them tighten the timeout and capture more complete bids from the 11 real competitors. Yield per impression rose 9%, and timeout-related bid loss on the strong partners dropped by a third. The dead-weight bidders had been inflating latency, which forced a wider timeout that hurt everyone.

Why it matters: in header bidding, each bidder is a latency cost paid on every auction. A partner that never wins and never sets the price adds delay without competition. Pruning non-competitive demand can raise yield by letting real bidders respond in time.