What Evgeny holds,
on the record.

The protocol-constitution-contract trilemma resolves by role assignment, not by a winner: protocols carry the credential, constitutions bound it, contracts create it.

“protocols carry the credential, constitutions bound it, contracts create it”

Rank standards by what happens to a real transaction when each is missing: three of four candidates are friction (a tax), and only Trust & Accountability's absence is a wall.

“Their absence is a tax on the transaction — real, worth paying down, but a tax, not a wall. Only one candidate's absence is a wall.”

Ev's Week 8 ballot: Trust & Accountability is the most important standard in five years, reasoned from the Clearing Test rather than season momentum.

“I voted Trust & Accountability, and the reasoning is the Clearing Test, not the season's accumulated momentum toward the same answer”

The first major court case over agent harm will turn on governance, delegation, and oversight design, not on whether the AI made a mistake — an explicitly on-the-record, still-open call.

“the first major court case in this space won't turn on whether the AI made a mistake. It will turn on whether the governance, the delegation, and the oversight were designed appropriately.”

An agent's economic value is gated on whether its output is permitted to act unreviewed — permission, not capability, is the bottleneck.

“An agent's value is gated on whether its output is permitted to act unreviewed.”

Confirmed in practice: Snowflake x Meta governed loop — Jul 21, 2026

Ev's Week 4 ballot: trust, not data, is the most valuable asset in an agentic market — the scarce question is whose agent is allowed to act.

“I voted Trust against that reflex, and the room closed on the same side: a 52% majority”

Any single-scalar optimization metric handed to an agent gets gamed, because a scalar's cheapest route up is faking it; Business Outcomes resists this precisely because it cannot be scalarized.

“Call the failure the Scalarization Trap — the moment you collapse "what we actually want" into one number an agent can watch and chase, you’ve handed it a target that’s cheaper to fake than to earn”

Under agentic commerce a metric breaks if it presumes a human eye or browser session, survives if it observes money or causality, and measurement authority migrates to whoever authenticates the agent.

“a metric breaks if it presumes a human eye or a browser session; it survives if it observes money or causality; what emerges is whatever must observe the agent itself”

Retail media iROAS is a negotiated artifact, not a measurement: defensible methodology choices alone flip most campaigns between profitable and money-losing, so buyers must audit the methodology negotiation.

“The same campaign, the same spend, the same shoppers — profitable under one defensible recipe, money-losing under another.”

Ev coins "agent share of choice" — the successor to Share of Model, measuring not whether a brand is mentioned by an AI but whether it was actually selected or defaulted to in a completed agentic purchase.

“No standards body or vendor has standardized that term as of July 2026; call it agent share of choice — coined here, labeled as such.”

Under agentic commerce, measurement authority is shifting from whoever watches the shopper to whoever cryptographically authenticates the agent — the same kind of power shift that made panel ownership decisive in the human web.

“measurement authority is migrating from the parties that watch the shopper to the parties that authenticate the agent. In the human web, whoever owned the panel owned the truth. In the agentic web, whoever verifies the signature does.”

When agents become the primary buyers, inference metrics (Attention first, then Reach & Frequency and Attribution) lose their referent because the decision path becomes directly readable; ROAS survives by accident.

“a metric survives agentic decisioning only if the event it measures remains unobservable in the system now making the decision — inference has no work left to do once you can just read the record”

The instrument that replaces broken inference metrics is the Delegate Audit Record: a structured trace of the agent's spend decision — constraints, alternatives considered, predicted vs actual outcome.

“a structured trace of why an agent made a specific spend decision — its constraints, the alternatives it considered, and its predicted outcome measured against the actual one”

Campaign reporting dies as a category, replaced by live diagnosis with human-in-the-loop resolution.

“Campaign reporting dies, replaced by live diagnosis with human-in-the-loop resolution.”

Confirmed in practice: Snowflake x Meta governed loop — Jul 21, 2026

Transformer-based MMM (Google's NNN) supersedes parametric Bayesian MMM (Meridian) by learning temporal carryover and cross-channel synergy directly from aggregate data, keeping measurement privacy-safe post-cookie.

“NNN signals a transformative shift in how we measure and optimize media in a privacy-first, post-cookie world”

Attention is not a spotlight but a relevance engine: paying attention is the act of filtering an overwhelming world for what might matter, which is why equal attention produces uneven outcomes across contexts.

“attention is not a spotlight — it's a relevance engine.”

Attention metrics alone do not predict marketing outcomes; measurement needs a relevance layer — the EARO framework's 22 metrics across Exposure, Attention, Relevance, Outcome.

“Traditional metrics (impressions, CTRs) can’t answer these questions. They tell you what happened, not why.”

Coins EARO — Exposure, Attention, Relevance, Outcome — an open-source framework for measuring marketing effectiveness holistically, built on the premise that attention alone is not predictive.

“I'm building EARO (pronounced "Hero") — an open-source framework to measure marketing effectiveness holistically.”

Cookie loss does not kill measurement's pillars but rebuilds each around privacy: 1:1 attribution shrinks to samples while clean rooms, lift analysis, and MMM/Markov modeling carry the load.

“measurement will have to leverage multiple solutions to make up for the loss of cookie tracking and the ability to deconstruct customer journeys”

Even flawless conversion counting across platforms does not establish causation; deduplication solves the counting problem while leaving the incrementality question untouched.

“the harder question underneath all of it: even a perfect count doesn't tell you what actually caused it”

Browser-native, differentially-private attribution upgrades a number's provenance but not its logic — it still assigns credit, not cause, while relocating measurement's rules into infrastructure no buyer signs.

“The privacy engineering upgrades the provenance of the claim, not its logic. Credit is not cause, even when a browser signs the credit.”

Lock-in migrates again in the agentic era: from rewriting SQL to an organization's accumulated institutional memory held by the agent — an exit cost CustomerLake does not document as exportable.

“The composable era's exit cost was rewriting SQL. This era's exit cost is an organization's institutional memory.”

The activation layer becomes the obvious execution surface for advertising agents — vindicated in operation, still premature in architecture as of the July verdict.

“the activation layer becomes the obvious execution surface for advertising agents”

Confirmed in practice: Jul 21, 2026

The warehouse-native model's defining bet is to own the substrate instead of renting the spine — the identity-utility trade in one line.

“The warehouse-native model makes the inverse bet: own the substrate instead of renting the spine.”

Composable does not escape lock-in; it relocates and arguably deepens it onto the warehouse, the hardest layer to leave.

“Composable reduces application-layer lock-in while concentrating, and arguably deepening, the dependency on the one component hardest to leave: the warehouse.”

Fourth-generation agentic-CDP claims are currently vendor-claimed and architecturally reasoned, not operationally demonstrated.

“The fourth-generation products making this argument are early — much of it is vendor-claimed and architecturally reasoned rather than operationally demonstrated.”

Agents acting on customer data need a governed truth, which strengthens the warehouse-native case and raises packaged-CDP exit stakes.

“An agent needs a governed truth to act on — memory, permissions, and audit live wherever the customer copy lives.”

Confirmed in practice: Jul 21, 2026

There is a fourth agentic-CDP implementation cell the category skipped — federated over the merchant's existing SaaS stack, for buyers with no warehouse and no data team — and Ev is building it.

“This product takes the fourth cell: federated. The decision layer sits over the merchant's existing SaaS stack and keeps only governed working memory — context, a versioned constitution, a ledger, holdout state.”

The governance built for the customer profile becomes the governance needed for the agent — a June call productized by the Snowflake loop five weeks later.

“the governance you built for the profile becomes the governance you need for the agent”

Confirmed in practice: Jul 21, 2026

Generation four of customer data architecture begins after data stops moving: the record stays put and everything else moves instead.

“Generation four is what happens once it has stopped: the record stays put, and everything else moves instead.”

Adtech and martech were never two industries but two custodies of one customer record — the Custody Partition — and one governed customer memory dissolves it.

“the Custody Partition — a divide born of who held the customer record”

Everything the stack currently carries -- identity graphs, contextual signals, Agentic Audiences vectors, signal containers -- encodes only the who and the what of a moment; where and when have no native representation yet.

“Identity graphs, contextual signals, the dense vectors inside Agentic Audiences, the signal containers this site has argued for — all of it encodes the who and the what.”

The policy layer inside a signal container -- not the activation logic -- is what separates a context economy from a surveillance economy running on the same rails.

“The policy layer is not paperwork stapled to the signal. It is the thing that separates a context economy from a surveillance economy running on the same rails.”

A signal container is a governed, executable decision object — not a segment ID — that packages buyer intent, meaning, provenance, policy, and activation logic together so intent can be discovered, matched, scored, activated, and audited across systems.

“A segment says, "Here is a group of users." A signal container says, "Here is the governed logic for how this intent should be discovered, matched, scored, activated, and audited across systems."”

The most powerful activation pattern is a dual-gate model, requiring both audience match and contextual alignment before an impression qualifies — this changes data from a pre-built segment into a live decision object.

“The most powerful version is a dual-gate model: the user is in the matched audience and the impression context is aligned with the signal.”

A signal container's value comes from packaging six things that normally travel separately and lose meaning at every hop — intent, meaning, provenance, policy, execution, and evaluation.

“A signal container packages six things that normally travel separately — and lose meaning at every hop.”

The market's dependence on identity-based matching for activation is too brittle for modern privacy constraints, CTV, retail media, and fragmented data environments.

“That "identity or nothing" model is too brittle for modern privacy, CTV, retail media, and fragmented data environments.”

Confirmed in practice: Jul 21, 2026

Missing standards layers always begin as private plumbing before they become open standards -- the pattern the industry defaults to when a governed object (like a closed-loop feedback layer) does not yet exist.

“which is how missing layers always begin: private plumbing first, standards later”

Confirmed in practice: Snowflake x Meta governed loop — Jul 21, 2026

In the contextual-path activation pattern, a leading SSP stops being pure supply pipe and becomes part of the decision runtime itself, comparing bidstream context against a semantic prototype in real time.

“In this model, a leading SSP is not just a supply pipe. It becomes part of the decision runtime.”

A signal container must be a governed, executable object rather than a segment fetched on the fly because it executes a decision already reasoned about upstream -- the line between an agent that plans and a runtime that acts.

“The container executes a decision that was reasoned about before the impression arrived. That is the line between an agent that plans and a runtime that acts”

A clean room is the venue where two mutually distrusting parties agree to trust a process — the mature 2026 definition, replacing the 2023 framing of the room itself as the deliverable.

“A clean room is where two parties who do not trust each other's infrastructure agree to trust a process.”

Output policy — not the clean room itself — is the actual deliverable of a clean-room deployment; the room is only the venue.

“Output policy is the product. The room is the venue; what leaves it is what you actually shipped.”

A clean room deployed without output policy is only superficially safe — the room's walls were never the hard part; the door is.

“A clean room without output policy is just a safer room with an unsafe door.”

Ev's operating distinction for evaluating clean-room initiatives: teams that stop at deploying the room have bought a venue; teams that build the output policy have shipped a product.

“Teams that treat the room as the deliverable buy a venue; teams that treat the output policy as the deliverable ship a product.”

PETs and policy are not the same layer: privacy-enhancing technologies are techniques, but policy is a decision, and a clean room needs both — only one of which ships in the box.

“PETs are techniques. Policy is a decision. A clean room needs both — and only one of them ships in the box.”

Zero-party data combined with Web3, NFTs, and blockchain technology has the potential to transform digital marketing — Ev's earliest recorded ZPD position, and one never revisited by his later ZPD/clean-room writing.

“ZDP in conjunction with Web3, NFTs, and blockchain technology have the potential to transform the way we think about digital marketing.”

The next wave of competitive advantage in advertising will come not from reach or recall but from contextual precision — from why an exposure was seen as meaningful.

“the next generation of competitive advantage won't come from reach or recall — but from contextual precision.”

Contextual targeting's next evolution is agentic, not static or taxonomic: Context Agents that reason over a five-category, twelve-question brand-evaluation framework in real time succeed the keyword- and rule-based contextual model.

“the future of contextual targeting is no longer static or taxonomic - it's agentic.”

Agentic contextual decisioning requires a structured, machine-readable rationale, not just a placement decision, because explainability is the precondition for trust in AI-driven advertising environments.

“It enables agentic systems to act with explainability, a critical requirement for trust in AI-driven advertising environments.”

Contextual evaluation should expand from four static brand-safety questions into a structured five-category, twelve-question framework, output as machine-readable JSON for real-time decisioning.

“we've expanded the framework into five categories and twelve questions, designed for intelligent, real-time decisioning.”

The instinct that makes Business Outcomes resist scalarization also shows up at company scale, in Perplexity's and Anthropic's refusals to let advertiser incentives inside the loop that mediates what an AI assistant tells users.

“keep the thing you actually care about — trust, in their case — outside the number the system is optimizing, because the moment it's inside, it's for sale to whoever prices it best”

Ev's Week 7 ballot: Business Outcomes, reasoned mechanically — it structurally resists the Scalarization Trap because it isn't a live number an agent can see and chase, not because it's a nobler goal.

“I voted Business Outcomes, and the reasoning is the mechanical one above, not the values-based one.”

Which layer of agent (copilot, platform optimizer, or agent-to-agent negotiator) is executing a spend decision shouldn't change where accountability lands — it lands on whoever authored the objective function, not on the layer.

“The poll left "agents" unspecified on purpose, because the answer shouldn't depend on the layer — whatever executes, the objective function is where accountability lands.”

A live campaign confirmed the Scalarization Trap outside the AI-shopping-agent context: PMax, handed a conversion action that barely fired, optimized toward whatever cheap inventory satisfied the thin signal rather than toward real outcomes.

“The machine did roughly what I told it to: find the cheapest way to look busy against a conversion event that essentially never happened.”

The training signal fed to a bidding algorithm has to be actively protected, because every unfiltered spam lead that reaches the conversion event is a poisoned training example teaching the optimizer that bots are customers.

“Every spam lead that reaches your conversion event is a poisoned training example — you are teaching the machine that bots are your customers.”

Katsur's build-on-existing-foundations argument for AAMP is a real engineering position, not just turf defense: retrofitting inherits working integrations, existing conformance programs, and two decades of operational knowledge that a greenfield protocol has to earn from scratch.

“retrofits inherit working integrations, existing conformance programs, and twenty years of operational knowledge.”

AdCP and AAMP put trust in different places with different failure modes: AdCP's wire-native trust (signed claims, governance tokens, idempotency) survives any deployment but only binds implementers of the cryptography, while AAMP's runtime trust binds whoever operates the reference stack but evaporates the moment that stack is swapped.

“Wire-native trust survives any deployment but only binds parties who implement the cryptography; runtime trust binds whoever operates the reference stack but evaporates if you swap the stack.”

AdCP's headline membership claim (118+) vastly outruns its verified live registry, where sixteen of twenty-one agents are sell-side and exactly one is a buying agent — a self-critical finding about the protocol Ev co-leads a working group for.

“The demand side of this protocol is thinner than its press.”

The AdCP-vs-AAMP standards fight is a conflict between the governing bodies, not the vendors underneath them — PubMatic, Yahoo, Optable, and Equativ all sit inside both camps at once.

“The war is between the bodies. The companies are hedging.”

Every standards body's press release outruns what its repository actually ships, and the only honest way to evaluate a standards claim is to measure that gap the same way for every competing body, not just the one you're skeptical of.

“Every standards body's press outruns its repos; the professional courtesy is measuring the gap the same way on both sides.”

The right way to judge which agentic-advertising standard is actually winning is by what has landed in the repository, not by what lands in the press feed.

“Judge both, then, the way this essay has tried to: by what lands in the repository, not what lands in the feed.”

Neither AdCP nor AAMP's trust model actually stands alone — both lean on an external verifier of last resort (AdCP on its registry and governance agents, AAMP on its registry and SafeGuard's platform), and that verifier layer is itself young and unproven at scale.

“Neither subsumes the other — and both, examined closely, lean on an external verifier of last resort (AdCP on its registry and governance agents, AAMP on its registry and SafeGuard's platform) that is itself young.”

AdCP and AAMP place judgment at opposite ends of the automation spectrum: AdCP gates every commit with a machine check and no value floor, while AAMP gates only above-threshold commits, and with a human -- two untested theories of where judgment belongs.

“Note the philosophical inversion against AAMP: AdCP gates every commit with a machine check and no value floor; AAMP gates above-threshold commits with a human. Those are different theories of where judgment belongs, and neither has been stress-tested at scale.”

The OpenAds/Prebid wrapper schism isn't a fight over auction mechanics — both wrappers share the same lineage — it's a fight over one OpenRTB field, who governs it, and whether a single company can credibly own the transparency layer it also bids through.

“The fork is not about auction mechanics — both wrappers run the same lineage. It is about one OpenRTB field, who governs it, and whether a single company can credibly own the transparency layer it also bids through.”

The Prebid Sales Agent (built on AdCP) is orthogonal and low-risk relative to the OpenAds/Prebid wrapper fight — evaluating the agentic standards layer does not commit a publisher to a side in the wrapper schism.

“Prebid Sales Agent is orthogonal and low-risk: open source, Apache-licensed, additive to an existing stack. Evaluating the agentic layer does not commit you in the wrapper fight.”

AdCP is a protocol layer that makes advertising work discoverable, executable, governable, and auditable by agents — it is explicitly not a replacement for OpenRTB, clean rooms, or human judgment.

“AdCP is not a replacement for OpenRTB, clean rooms, or human judgment. It is a protocol layer for making advertising work discoverable, executable, governable, and auditable by agents.”

The industry should not treat AdCP as a magic layer that makes agents safe by default — the protocol can define the task, but it cannot replace commercial judgment, governance design, data rights, measurement discipline, or operating ownership.

“The industry should avoid treating AdCP as a magic layer that makes agents safe by default. The protocol can define the task. It cannot replace commercial judgment, governance design, data rights, measurement discipline, or operating ownership.”

A packaging failure is not a branding problem but a set of measurable commercial symptoms — inconsistent seller narratives, buyers reducing a platform to its smallest product, deal-by-deal bundle renegotiation, siloed proof, and anecdote-driven roadmaps.

“A packaging failure shows up as hard commercial symptoms, every one of them measurable:”

A capability list is not a value proposition — presenting a buyer with 'all of the above' forces them to do the packaging work themselves, under deadline, against competitors who already did it; most buyers won't bother and default to the simplest label.

“A capability list is not a value proposition. It is a homework assignment handed to the buyer.”

Packaging sits above product marketing the way architecture sits above interior decoration — a messaging refresh cannot fix a structural packaging failure, because repainting a room with bad bones leaves it with bad bones.

“It sits above product marketing the way architecture sits above interior decoration. You can repaint a room with bad bones. It still has bad bones.”

The buyer does not evaluate a vendor's org chart or acquisition history — they buy the clearest business case they can defend to the person above them.

“The buyer does not buy your org chart. They buy the clearest defensible business case.”

Packaging becomes the moat not because it is clever but because it is rare and compounding — a buyable story sells faster, attaches wider, and defends itself in rooms the seller is not in.

“Packaging becomes the moat — not because it is clever, but because it is rare, and because it compounds: a buyable story sells faster, attaches wider, and defends itself in the room you are not in.”

The enterprise media buyer is not a single person but a system — brand need, agency execution, and platform workflow — and good packaging has to serve all three simultaneously.

“And the buyer is not a single person; in enterprise media, it is a system — brand need, agency execution, and platform workflow. Good packaging has to serve all three.”

M&A is underwritten on capability, synergy, and multiple — never on the sentence the buyer will use afterward — which is the root cause of packaging failure: every acquisition adds a box to the org chart, not a line to the story.

“Every acquisition adds a box to the org chart, not a line to the story.”

For European and Asian AdTech vendors, winning US market entry depends on product-market fit — aligning product, pricing, and positioning to US buyer expectations — not just aggressive selling or capital.

“For European and Asian AdTech vendors aiming to enter the US market, attaining product-market fit (PMF) is critical.”

European and Asian AdTech vendors should differentiate on transparency, brand safety, and defensible performance metrics rather than compete on price — quality, not volume, is where the growing majority of US programmatic spend goes.

“By emphasizing transparency, brand safety, and strong performance metrics, companies can differentiate themselves in the competitive US programmatic market.”

Ev's own market-entry audit methodology is a named seven-step framework — GTM strategy, PMF, market sizing, sales/BD, monetization, local talent, and compliance — offered as the structured diagnostic for European AdTech vendors entering the US.

“For European AdTech companies entering the US, No Fluff Advisory offers results-oriented solutions, tied to 7-Step Approach to Auditing Your Market Entry Strategy”

GTM sequencing for US entry should start with independent agencies, not group agencies — independents are more willing to run early-stage trials, and early proof there is what earns the larger, more formal partnerships later.

“Securing agency agreements is critical for long-term success, but initial momentum often comes from smaller, early-stage tests. Independent agencies, in particular, are more likely to engage in these trials, which can lead to larger, more formal partnerships after proving success.”