Anthropic обвинила DeepSeek, Moonlight и Alibaba в дистиляции

Статья о том, как китайские компании массово выкачивают ответы западных нейросеток, чтобы дообучать свои модели. Anthropic говорит о 24 000 фейковых аккаунтов и 16 млн запросов к Claude. Вывод простой: рынок ИИ входит в фазу жёсткого копирования, а китайские модели Qwen, Kimi и DeepSeek могут быстро догонять лидеров дешевле.

➡️ Читайте на сайте: https://aff.top/blog/anthropic-obvinila-deepseek-moonlight-i-alibaba-v-distiliacii

🧠 Ещё больше инсайтов → в канале AFF.top

A checklist for ruling out creative as the cause of lost auctions

Before blaming bids or floors for low delivery, eliminate the creative layer — it silently disqualifies you in ways the auction report won't show.

1. Confirm creative approval status per exchange. A creative pending or rejected on one SSP makes every bid there ineligible regardless of price.

2. Check declared vs actual ad markup. If your creative's declared size, format, or secure (HTTPS) status mismatches the impression, the SSP filters the bid pre-auction.

3. Audit creative weight and load time. Many publishers enforce a max payload; a heavy creative is dropped after winning, costing you a win that never renders — and damaging your seat's reputation score.

4. Verify all third-party calls (trackers, verification tags) are on the SSP's allowlist. One blocked vendor tag can disqualify the whole creative.

5. Check category and advertiser-domain declarations match the SSP's records — a domain mismatch trips brand-safety filters on the sell side.

6. Only after the creative is verified clean on every active exchange should you investigate price.

Why it matters: Creative disqualification looks identical to losing on price from the buyer's dashboard, yet no bid increase will ever fix it. A systematic creative-eligibility pass per exchange rules out an entire class of invisible losses before you waste a week tuning bids that were never even entered into the auction.

How to A/B test two bid strategies without contaminating the result

Comparing bid strategies on the same inventory is harder than it looks, because the two arms compete with each other. Here's a clean protocol.

1. Split on a stable hash of the user ID, not on time or geo. Time splits absorb daypart effects; geo splits absorb price differences. A user-hash split keeps both arms facing the same market.

2. Critically, ensure the two arms never bid in the same auction. If both strategies can enter one auction, they inflate each other's clearing price and you measure interference, not strategy.

3. Fix budget per arm proportional to its user share, and pace identically. A faster-pacing arm wins cheaper morning inventory and corrupts the cost comparison.

4. Run long enough to clear the conversion lag of your KPI — a 2-day test on a 7-day conversion window measures noise.

5. Compare on outcome per dollar (cost per action, viewable cost per mille), not win rate. A strategy that wins more can still cost more.

6. Check that win-price distributions didn't drift between arms; if they did, the split leaked.

Why it matters: Most in-platform bid experiments are invalid because the two arms cannibalize each other's auctions and budgets, producing a confident but meaningless winner. A user-hash split with auction isolation is the only design that lets the market treat both strategies fairly — without it, you're comparing two strategies that sabotaged each other.

Тем временем подстилка коричневых Кардиналов и лично Кустова главред Максим Огненный завёл собственный канал, где, наверное, опять будет писать стихи и анонсировать вьюхи с ме**дроновыми наркоманами.🤡🤡

Чем вообще известен этот персонаж? Собсна, только порцией отборного кринжа, например, не так давно он выебывался на Иванова в "Письмах кардинала", но недожал и тема осталась нераскрытой. ЕЮ заслужил даже высокоинтеллектуальные выпады, которые тот не понял в силу врождённого аутизма:

Высокохудожественная лексика героя, демонстрируемая им не только на своем канале, но и на любой публичной площадке, настолько пленит любого слушателя, что мало кто может стать собеседником жертвы во второй раз.

Нихуя не разбираясь в аффилке, он на серьёзных щах брал интервью у Мефмедова и пытался построить серьёзный диалог с объёбанной ракетой. Из остальных достижений только нытье в чатиках: персонаж настолько невзрачный, что даже хуесосить его не так интересно.¯\_(ツ)_/¯

А теперь Огненный завёл собственный канал, где собрался выебать всех, но, походу, ебать будет лишь подписоту своей графоманией. Похожей хуйнёй занимался и оунер ФБ Киллы, который так любит рассуждать про экономику и политику. Оно и понятно: что у коричневых, что у киллы, что у партнёркина — один инвестор и одна крыша, поэтому подопечные синхронно занимаются бестолковой хуйнёй. Кому не похуй, подписывайтесь, будет интересно (нет): https://t.me/+dr240TeLtPc2Nzdk

В арбитраже деньги есть, но только у тех, кто с LuckyCards 💵

How to read a win-price distribution to set a smarter bid

A single bid number is a blunt instrument. Reading the full distribution of prices at which you win tells you exactly where to set it.

1. For one segment, pull every winning clearing price and plot the distribution — the histogram of what you actually paid to win.

2. Identify its shape. A tight cluster means a stable, competitive market; a long right tail means a few auctions cost far more than the rest, dragging your average.

3. Compute the marginal cost of the tail: how much win-rate you gain by bidding to the 90th percentile versus the 75th, and what each incremental win costs. Often the top decile of price buys a trivial slice of wins at outrageous cost.

4. Set your bid at the percentile where marginal cost per win crosses your value-per-win threshold — not at the mean, which over-rewards the expensive tail.

5. Re-plot weekly; the distribution shifts as competitors enter and floors move.

6. For first-price exchanges, pair this with shading so you bid the percentile, not above it.

Why it matters: Buyers anchor on average win price and quietly overpay for a handful of expensive auctions in the tail that contribute almost nothing to scale. Reading the distribution exposes that tail and lets you set the bid at the point where each additional win is still worth its cost — turning a guessed number into a defensible one.

Decomposing win rate into three independent failures

'Our win rate is 12%' is a useless number until you decompose it. Win rate is the product of three conditional stages, and each fails for different reasons.

1. Bid rate — of auctions you were eligible for, how many did you actually bid on? Low here means targeting, budget pacing, or frequency caps are filtering you out before the auction.
2. Bid-to-win rate — of auctions you bid on, how many did you win? Low here is a price problem: your bids or your shading are below clearing.
3. Render rate — of impressions you won, how many actually rendered and were billable? Low here is technical: timeouts, viewability filters, or the publisher's ad server bumping you.

Multiply them and you get effective win rate. The diagnostic value is that each stage points to a different team and a different fix. A 12% effective win rate from 40% bid rate times 35% bid-to-win times 86% render is a completely different problem than 95% times 14% times 90%.

Why it matters: teams burn weeks raising bids when their bid rate is the bottleneck and price was never the issue. Always pull the three rates separately from log-level data before touching a single CPM.

Antropic раскрыла ферму с нейронными дейтинг-моделями

Anthropic раскрыла кейс китайской студии, которая запустила 20 дейтинг-приложений с нейропрофилями для удержания пользователей. Вместо стандартного слива дейтинг-трафика на чужие смартлинки, вебмастерам стоит присмотреться к разработке собственных LLM-сервисов. Затраты на токены и подключение платежных решений окупаются за счет прямого контроля над монетизацией и забора всей маржи рекламодателя.

➡️ Читайте на сайте: https://aff.top/blog/antropic-raskryla-fermu-s-neironnymi-deiting-modeliami

🧠 Ещё больше инсайтов → в канале AFF.top

Разработчик игры слил $220 в Google ads на установки ботами

Статья описывает кейс разработчика, потерявшего бюджет в Google Ads из-за фрода и специфики атрибуции. Платформа засчитала конверсии, которые не подтвердились в Google Play Console, так как алгоритм учитывает установки без прямой связи с кликом. Главный вывод: Google Ads может наливать ботов не меньше пуш-сетей. При работе за инсталы важно жестко контролировать качество трафика и сверять аналитику с бэкендом, в то время как Facebook на данный мо…

➡️ Читайте на сайте: https://aff.top/blog/razrabotchik-igry-slil-220-v-google-ads-na-ustanovki-botami

🧠 Ещё больше инсайтов → в канале AFF.top

Google расширил Data Manager

Google интегрировал инструмент Data Manager в GA и DV360 для удобной передачи first-party данных, что дает рост ROAS до 26%. Для арбитража трафика прямого применения у офлайн-данных нет, однако инструмент можно использовать для манипуляции алгоритмами: отправка синтетических конверсий через API поможет скорректировать оптимизацию и направить автостратегии на поиск нужной аудитории.

➡️ Читайте на сайте: https://aff.top/blog/google-rasshiril-data-manager

🧠 Ещё больше инсайтов → в канале AFF.top

A checklist for prepping a clean incrementality test in programmatic

Incrementality (the conversions your ads actually caused, versus those that would have happened anyway) is the only honest performance question. Set the test up correctly or it lies.

1. Build the control group as a true holdout — users eligible to be served who are deliberately withheld via a ghost bid or PSA, not users who simply weren't reached. Unreached users are a biased comparison.

2. Randomize at the user level before the auction, so exposure status is assigned independently of bid outcome. If control is defined post-hoc as 'people we lost the auction on,' you've selected for cheaper, different users.

3. Size the holdout for your baseline conversion rate. Low base rates need large holdouts to detect a real lift above noise.

4. Lock the measurement window to your conversion lag and freeze targeting and budget for the duration.

5. Suppress cross-DSP contamination — if another buy reaches your control group, lift collapses.

6. Pre-register the metric and window before launch so you can't cherry-pick afterward.

Why it matters: Almost every 'lift' number in programmatic is selection bias dressed as causation, because the control group was defined by auction losses rather than random assignment. A pre-randomized, pre-registered ghost-bid holdout is the difference between knowing your ads work and merely hoping they do.

The match-rate tax that silently caps your win rate

Before you can bid on a user meaningfully, your DSP must recognize them — match the exchange's user identifier to your own ID via cookie syncing or an identity graph. Match rate is the percentage of bid requests where that recognition succeeds, and it quietly caps everything.

1. On an unmatched request you either don't bid (lost reach) or bid blind without your audience data (wasted spend on the wrong user).
2. Match rates vary hugely by exchange and environment — high on the web with cookies, far lower on browsers that block third-party cookies, and structurally different in-app.
3. A 50% match rate means half your premium audience is invisible to you on that path no matter how high you bid.

This interacts with SPO: two paths to the same publisher can have very different match rates, so the 'best' path on take rate might be worse on addressability.

Why it matters: when an audience underdelivers, check match rate before bids. You may be winning auctions for the right inventory but the wrong people, or skipping winnable users because you never recognized them. As third-party cookies erode, match rate is becoming the binding constraint on programmatic — measure it per exchange and weight your supply paths accordingly.