Meta ограничивает расходы на токены для сотрудников

Meta ввела внутренние лимиты на использование ИИ из-за резкого роста расходов: в 2026 году только на сотрудников заложены миллиарды долларов, а общий бюджет на ИИ-инфраструктуру оценивается в 130–145 млрд. Вывод простой: даже у Big Tech ИИ перестал быть бесплатной игрушкой и требует жёсткого контроля затрат.

➡️ Читайте на сайте: https://aff.top/blog/meta-ogranichivaet-raskhody-na-tokeny-dlia-sotrudnikov

🧠 Ещё больше инсайтов → в канале AFF.top

Myth: First-price auctions made everyone overpay

The folk story is that when exchanges flipped from second-price to first-price around 2017-2019, buyers got fleeced — you now pay your full bid instead of the runner-up's price plus a cent.

Let's trace the actual mechanism.

1. In a true second-price auction (winner pays the second-highest bid + 0.01), the dominant strategy is to bid your true valuation. There's no penalty for honesty.

2. The catch: most "second-price" exchanges were never clean. Soft floors, fee-laden reserve prices, and unobservable publisher-side rules meant you often paid close to your bid anyway. The audit was impossible because you couldn't see the clearing logic.

3. First-price (you pay exactly what you bid) forces a behavioral change called bid shading — the DSP lowering your submitted bid toward the expected clearing price, estimated from won/lost log data.

4. Net effect: a well-calibrated shading model in first-price often clears lower than the murky second-price world it replaced, because now the clearing price is observable and the model optimizes against it.

The transition didn't raise prices structurally. It moved price discovery from the seller's hidden ledger into the buyer's model, where you can actually measure it.

Why it matters: blaming first-price for high CPMs usually means your shading is mistuned, not that the auction format is robbing you.

Cross-checking sellers.json removed paths carrying 6% invalid traffic

sellers.json (a public file where each SSP lists the seller accounts authorized to sell its inventory). One buyer used it to audit supply quality, not just spend.

1. They mapped every winning path's seller ID against the SSP's sellers.json and the publisher's ads.txt.
2. They flagged paths where the seller was listed as 'intermediary' but the chain length in the SupplyChain object exceeded two hops.
— 11% of spend ran through chains of three or more intermediaries.
3. They overlaid invalid-traffic rates from their measurement vendor onto those paths.

Evidence: the long-chain paths carried a 6.1% invalid-traffic rate versus 1.2% on direct and single-hop paths. Blocklisting them cut measured invalid traffic on the line nearly in half and shifted that budget to cleaner supply at a comparable CPM.

Why it matters: chain length in the SupplyChain object correlates with fraud exposure. Each undeclared intermediary is a place provenance breaks. Auditing seller IDs against sellers.json turns an opaque path into a traceable one, and untraceable paths are where invalid traffic hides.

Claude Cowork, Claude Design объединили в один Claude

➡️ Читайте на сайте: https://aff.top/blog/claude-cowork-claude-design-obedinili-v-odin-claude

🧠 Ещё больше инсайтов → в канале AFF.top

Microsoft планирует вставлять рекламу в игры

➡️ Читайте на сайте: https://aff.top/blog/microsoft-planiruet-vstavliat-reklamu-v-igry

🧠 Ещё больше инсайтов → в канале AFF.top

Myth: Pour in budget and the algorithm learns faster

The advice everyone repeats to new buyers: front-load spend so the DSP's optimization model "exits learning" quickly. More impressions, more signal, faster convergence. It treats learning as a function of volume alone.

The mechanism is narrower than that.

1. Most DSP bid models optimize toward a conversion or post-click event. The model learns from positive events (conversions), not raw impressions.

2. If your funnel produces 30 conversions a week, doubling impression volume doesn't double the training signal — it floods the model with negatives and dilutes the rare positives the model actually needs.

3. Worse: aggressive spend pushes you up the bid landscape into inventory you'd normally lose, changing the distribution of what you win. The model is now learning from a population it won't see at steady-state pacing.

4. Statistical convergence is governed by event count and event stability, not dollars. A campaign with 50 conversions/week on steady pacing trains more reliably than one with 50 conversions/week achieved through erratic budget dumps.

The fix is to stabilize the conversion stream and the won-inventory distribution, then let the model see a consistent landscape long enough to estimate it.

Why it matters: "spend to learn" wastes budget buying noise. Learning is bottlenecked by positive-event density and distribution stability, not by how fast you empty the wallet.

A cross-SSP frequency cap stopped 28% of impressions hitting fatigued users

Frequency cap (a limit on how many times one user sees a campaign). A buyer capped at 3 per day per SSP but tracked users separately on each path, so the real frequency ran far higher.

1. They unified user IDs across SSPs in log-level data and recomputed true daily frequency per user.
— 28% of impressions went to users already at 7-plus exposures that day, well past the intended 3.
2. The per-SSP caps each held, but four SSPs each delivered 3 to the same user, summing to 12.
3. They moved frequency capping to the DSP layer, enforced once across all paths.

Evidence: those 28% of wasted impressions were reallocated to fresh users. At flat spend, unique reach rose 19% and conversion rate per impression climbed 11%, since budget stopped pounding fatigued users with no marginal lift.

Why it matters: a frequency cap enforced per SSP is not a frequency cap. The same user arrives through every path. Capping must happen at the DSP, where all paths converge, or the real exposure is the sum of every per-path limit.

Myth: A higher win rate means a healthier campaign

Buyers obsess over win rate — won impressions divided by bids submitted — and treat climbing it as progress. The reasoning feels intuitive: winning more of what you bid on means you're competitive.

Walk the mechanism and it inverts.

1. Win rate is a ratio you control trivially. Raise your bid, win more, watch the number climb. It measures bid aggression as much as inventory quality.

2. There's a structural ceiling worth naming: the winner's curse — in any auction, conditional on winning, you tend to have over-estimated the impression's value, because you only win when you bid above everyone else, including bidders who knew something you didn't.

3. So a very high win rate often signals you're systematically the highest, most over-confident bidder in the room. You're not winning good inventory; you're winning contested inventory at a premium.

4. The healthier diagnostic is win rate at a fixed bid across time, or win rate segmented by supply path. A rising win rate at constant bid means real competitive improvement; a rising win rate from rising bids means you're funding the winner's curse.

Why it matters: win rate optimized in isolation rewards over-bidding. Always pair it with effective CPM and post-bid outcomes, or you'll celebrate a metric that's quietly draining margin.

Cookie matching vs ID solutions vs contextual: three addressability tools, three failure modes

As third-party cookies erode, buyers reach for replacements without noticing they have different mechanics and different blind spots.

1. Cookie matching (the legacy path): your DSP and the SSP sync user IDs via a match table. Works in Chrome-less environments poorly and not at all in Safari/Firefox, where the cookie is already blocked.

2. Universal ID solutions (UID2, ID5, and similar): a hashed-email-based identifier shared across the chain. Works without third-party cookies, but only when the user authenticated somewhere to seed the email — coverage is uneven by inventory type.

3. Contextual: no user identity at all; you target the content of the page. Immune to identity decay, but you lose frequency capping and cross-site retargeting.

4. The tradeoff axis is coverage versus precision. Cookies are precise where they exist and absent elsewhere. Universal IDs extend coverage but only on authenticated supply. Contextual covers everything but targets the page, not the person.

Why it matters: a single addressability strategy leaves a coverage hole. Measure match-rate by browser and supply path in log-level data. On Safari/in-app inventory where cookie match is near zero, a universal ID or contextual layer is not optional — without it that inventory is effectively un-targeted and you are paying personalized prices for anonymous impressions.