How identity match-rate silently rewrites your auction economics

Match-rate (the share of users a DSP can resolve to a known identifier on a given supply path) is treated as a targeting input. It's actually a price-setting force.

The chain of effects:
— On a path with high match-rate, you can identify valuable users and bid up for them specifically.
— So can every other buyer with the same identity graph. Competition concentrates on identifiable users, raising their clearing prices.
— On a low-match path, the same users arrive anonymous. Fewer buyers recognize their value, so competition is thinner and clearing prices are lower.

The non-obvious implication: the highest-value users are often cheapest to win where they are least identifiable, because the auction can't price what it can't see. Buyers who can resolve identity through first-party signals or a stronger graph than competitors can find under-priced reach on low-match supply.

How to exploit it:
— Measure match-rate and median clearing price per path together.
— Look for paths where your private match-rate exceeds the market's typical match-rate — there you identify value others miss and pay thin-competition prices for it.
— Treat identity coverage as a sourcing advantage, not just a targeting filter.

Why it matters: an auction prices visible value. Where identity is scarce, value hides, and the buyer who sees it without revealing it through aggressive bids captures the arbitrage.

Z.ai анонсировала новую GLM-5.5

Z.ai готовит релиз флагманской GLM-5.5: модель обещают показать в августе 2026 года.

Главная интрига — рост до 1 трлн параметров при том же контекстном окне в 1 млн токенов. Новинка снова будет заточена под код и агентные задачи.

Почему версия сразу 5.5, без 5.3 и 5.4, и что это может означать для рынка — в блоге.

➡️ Читайте на сайте: https://aff.top/blog/z-ai-anonsirovala-novuiu-glm-5-5

🧠 Ещё больше инсайтов → в канале AFF.top

Telegram запустил собственный сервер для ботов

Telegram запустил собственный сервер для ботов и мани-приложений: теперь backend можно размещать прямо внутри инфраструктуры мессенджера.

Сервер работает на JavaScript/TypeScript, через вебхуки, и позволяет подключать SQL-базу для сбора контактов без посредников.

Пока неясны цена и ограничения — что именно уже можно тестировать, а где скрыт подв…

➡️ Читайте на сайте: https://aff.top/blog/telegram-zapustil-sobstvennyi-server-dlia-botov

🧠 Ещё больше инсайтов → в канале AFF.top

Pacing is a feedback controller fighting your bid price

Budget pacing (spreading spend evenly across a flight instead of front-loading it) seems like simple budget math. Mechanically it's a controller that quietly distorts every bid you make.

How it interferes:
— The pacing system sets a spend target per time slice and watches actual spend against it.
— If you're under-pacing, it loosens — raising bids or bid-rates to win more, regardless of whether the inventory is well-priced right now.
— If you're over-pacing, it throttles — lowering bids or skipping auctions, even excellent ones.
— This multiplier sits on top of your value-based and shading-based bid.

The distortion: in low-supply hours, an under-pacing line item bids up to catch up, paying inflated CPMs precisely when good inventory is scarce. In high-supply hours it may throttle and skip cheap, valuable impressions to avoid over-pacing.

The diagnostic:
— Plot hourly CPM against hourly pacing multiplier. If your most expensive impressions cluster in your most aggressive pacing windows, pacing is overriding price discipline.
— Consider asymmetric pacing: allow front-loading when clearing prices are demonstrably low, and tolerate under-delivery rather than chasing spend into expensive hours.

Why it matters: pacing optimizes for even spend, not for surplus. Left naive, it forces you to overpay when supply is thin and walk away from value when supply is cheap — the exact inverse of good buying.

Google картинки станут конкурентом Pinterest

Google Картинки начали превращать в полноценную платформу с персональной лентой по прошлым запросам — по сути, в аналог Pinterest.

Во вкладке For you уже тестируют подборки, а ещё обещают коллекции и генерацию изображений во встроенной Nano Banana.

Как это будет работать и когда новинка дойдёт до других стран — в блоге.

➡️ Читайте на сайте: https://aff.top/blog/google-kartinki-stanut-konkurentom-pinterest

🧠 Ещё больше инсайтов → в канале AFF.top

sellers.json and SupplyChain: the trust graph buyers underuse

sellers.json (an SSP's public list of who it pays and is authorized to represent) and the SupplyChain object (the per-bid record of every intermediary that touched an impression) together form a verifiable graph of the supply chain. Most buyers check them for fraud only. They also encode cost.

Reading them as cost structure:
— Each node in the SupplyChain object is an intermediary taking a margin. A four-node chain to reach a publisher is, by construction, more expensive than a one-node direct path to the same publisher.
— sellers.json marks each seller as PUBLISHER, INTERMEDIARY, or BOTH. Long runs of INTERMEDIARY nodes signal resold-of-resold inventory with stacked fees.
— Comparing the same publisher's impressions across paths, the chain depth correlates with the gap between your price paid and the publisher's actual received revenue.

The action, step by step:
— Parse the SupplyChain object on incoming bid requests and tag each by node count and authorization status.
— Cross-check declared sellers against the publisher's own ads.txt and sellers.json to flag unauthorized or undeclared resale.
— Prefer the shortest authorized path to each publisher; throttle deep chains unless they uniquely deliver inventory you can't get direct.

Why it matters: the SupplyChain object isn't just a fraud signal — it's an itemized receipt of the intermediaries between your money and the publisher. Buyers who read it as a cost graph route around fee stacking that buyers who read it only for fraud never see.

Россияне не смогут покупать стейблкоины

Россиянам могут закрыть доступ к покупке стейблкоинов: в новой версии закона их приравняли к иностранным активам.

Купить такие токены смогут только квалифицированные инвесторы — например, с активами от 24 млн рублей или доходом от 12 млн в год.

Что это значит для обычных пользователей и когда правило заработает — в блоге.

➡️ Читайте на сайте: https://aff.top/blog/rossiiane-ne-smogut-pokupat-steiblkoiny

🧠 Ещё больше инсайтов → в канале AFF.top

Your loss data is censored — and that biases every model you build

Every buyer's bid log has a structural blind spot: censored loss data (you see that you lost, but rarely the price you lost to). Ignoring this censoring quietly biases your price models.

The problem, precisely:
— On wins, you observe everything: your bid and the price you paid.
— On losses, you typically observe only the fact of losing. The winning price is hidden.
— So your sample of clearing prices is truncated: it over-represents auctions that cleared low enough for you to win.

If you estimate a segment's typical clearing price from your wins alone, you systematically under-estimate it, because every auction that cleared above your bid is invisible to you. Your model then shades too aggressively, loses more, and the bias compounds.

The statistical fix:
— Treat this as a survival-analysis problem. Each lost auction is a censored observation: the true clearing price is somewhere above your bid.
— Estimate the clearing-price distribution accounting for that censoring (the same math used for time-to-event data), rather than averaging only observed wins.
— Where SSPs provide minimum-to-win on losses, use it — it converts a censored point into a known one and sharpens the estimate dramatically.

Why it matters: naive win-only analysis is not just noisy, it's biased in a known direction — downward. Correcting for censoring is the difference between a shading model that reflects the market and one that systematically misjudges it.

Reading rec

If this channel's your speed, @CreativeRadar runs a sharp feed on Ad creatives / spy tools. Different angle, same depth — worth a follow.

QPS throttling secretly biases which auctions you ever see

QPS (queries per second — the rate of bid requests a DSP accepts from an SSP) is an infrastructure setting. But the way DSPs shed load under QPS limits biases your entire view of the market.

The mechanism:
— SSPs send far more bid requests than any DSP can evaluate. The DSP caps intake at a QPS ceiling per SSP.
— When requests exceed the ceiling, the bidder drops the overflow. The question is which requests get dropped.
— Naive load-shedding drops whatever arrives when the queue is full — which correlates with traffic peaks, certain geographies, and high-volume publishers. Those are dropped disproportionately.

The consequence: your bid log is not a random sample of available inventory. It's skewed toward the requests that happened to arrive when you had spare capacity. Any segment-level price or performance model you build inherits that sampling bias.

The diagnostic and fix:
— Compare your seen-request distribution per SSP against the SSP's reported available supply. Large gaps on specific publishers or hours reveal where QPS shedding blinds you.
— Move to value-aware shedding: prioritize intake of requests likely to be high-value (known segments, target geos) and drop low-value overflow first, so your sample skews toward inventory you'd actually buy.

Why it matters: you can only buy what your bidder sees, and QPS shedding decides what it sees. An uncontrolled drop policy means your market view — and every model built on it — is shaped by your queue, not by the market.