The raw research behind the report

Executive summary

Every major player that controls ad-delivery infrastructure — Google Ad Manager and DV360, The Trade Desk, Amazon Ads, Magnite, PubMatic, Microsoft Monetize, Criteo, OpenX, Yahoo DSP, Meta, TikTok — shipped the same kind of standardized AI layer onto its existing platform in 2025-2026, available identically to every customer: copilots that answer natural-language questions about performance (#1, #27), generative tools that auto-build campaigns from a media plan or a prompt (#25, #26, #70), and ML models that auto-optimize bidding, pacing, and creative (#3, #24, #42, #46). Adoption of these features is now near-universal within single platforms — The Trade Desk's Kokai went from "nearly all" clients trying it to "almost 100%" running through it in two quarters (#15, #16); Meta's Advantage+ runs at a $75B annual revenue run rate (#46). That is the textbook signature of Process Automation at Scale: the same tool, sold identically to every competitor, automating a job (bidding, reporting, campaign setup) that already existed.

But this domain also produced an unusually rich cluster of Institutional Capability candidates relative to what a market this saturated with standardized copilots would predict — concentrated specifically in proprietary decisioning/auction engines built over years, not bought off a shelf. AppLovin's AXON engine, refined since 2023, re-rated the company from a mobile-gaming discovery tool into a $5.48B-revenue single-segment ad-tech business (#58, #59, #61). Unity rebuilt its ad-targeting engine from scratch as Vector after a 2022 model-corruption failure cost it $110M in guidance, and Vector now runs above a $1B annualized revenue rate (#65, #66, #67). Comcast's VideoAI, built from years of internal Comcast/NBCUniversal/Sky deployments, is now compounding into externally-sold products like FreeWheel's 2026 Context Engine (#4, #5). The Trade Desk, PubMatic, Magnite, and Index Exchange are each racing to own the newest layer — agent-to-agent transacting infrastructure (#9, #35, #37, #39) — though as of this research pass that race is still splitting between genuine multi-partner platform capability and single-company bespoke pilots (see below). None of this displaces the base-rate finding: most of the dollars and most of the announcements are still automation, not invention.

The whole domain also sits under a structural cloud that has nothing to do with AI capability: Google's ad-server and ad-exchange businesses were found to be an illegal monopoly in April 2025 (#12, #71), with a US remedies ruling still pending as of this writing and a separate €2.95B EU fine under appeal (#73, #75). Whoever ends up controlling ad-serving infrastructure at all could be reshaped by these rulings independent of who has the best model.

A follow-up research pass (2026-08-22, after this report's thesis was grilled — see `01-brainstorm/grill-me-log.md`) sharpened and partly corrected the picture above. The orchestration/transacting layer is standardizing fast — but around two rival open protocols, not one: IAB's ARTF (finalized March 2026, #87) and the independent AdCP/AgenticAdvertising.org (127+ members, #89), and the industry's first real agent-to-agent transaction ran on AdCP, not ARTF (#89). What these protocols actually do, per their own technical documentation, is a linear discovery-then-approve exchange with no negotiation step (#98) — the "negotiation, not just bidding" framing applied earlier to the PubMatic/Optable and Magnite/Scope3 deals doesn't survive closer reading of the protocol spec or the deals' own mechanics (#96, #97, #99). And The Trade Desk's Koa/Kokai — cited below as this domain's clearest open-web Institutional Capability evidence alongside AppLovin and Unity — was directly contradicted by Morningstar's March 2026 downgrade of TTD's economic moat to "none" (#101), with TTD's stock and growth continuing to deteriorate through August 2026 (#102, #103). TTD's own CEO, in his own words, draws exactly the line this report's thesis draws — open access layers versus proprietary decisioning and "exclusive" data (#106, #107) — but Morningstar's reasoning is that TTD's actual data asset, UID2, functions as an open, multi-tenant identity standard other DSPs can also use, not a closed exclusive network the way LiveRamp's governed graph or AppLovin's and Unity's owned SDK telemetry are (#106). The distinction between genuinely exclusive data and an open standard merely branded as proprietary turns out to be load-bearing.

How AI is actually being used

Publisher-side ad servers & yield management. Google Ad Manager's core yield-management pitch is explicitly "grow revenue... without the need for an in-house data scientist" via ML-driven Dynamic Allocation, Exchange Bidding, and Programmatic Guaranteed (#3) — a standardized substitute for institutional data-science capability, available to any publisher, stated almost verbatim by the vendor. On top of that sits "Ask Ad Manager," a Gemini-powered conversational agent for diagnosing delivery problems and generating custom reports, launched free and unlimited in June 2026 (#1), plus a November 2025 wave of brand-safety, generative reporting, CTV Live-biddable, and "Buyer Direct" tools (#2). Magnite's SpringServe added ML "smart podding" that predicts the minimum bid-request volume needed to hit fill targets (#7); on its Q1 2026 call, management described AI as still an early-stage contributor to inventory valuation and execution, attributing part of that quarter's margin gains to "early AI-related productivity gains" alongside a separate cloud-to-on-prem infrastructure shift (#36) — a rare instance of a company explicitly characterizing its own AI impact as modest and just beginning, rather than transformative. Microsoft Monetize runs standardized LLM-based "Ad Quality Models" (Clickbait, Image Sensitivity & Aesthetics) across its whole publisher network (#10).

Agentic and agent-to-agent transacting. This is the newest and most contested layer. FreeWheel built AI-agent infrastructure on its own MCP server, piloted so far with a single agency partner, PMG (#6). Magnite embedded an AdCP seller agent into SpringServe and, per CEO Michael Barrett, executed "the industry's first agent-to-agent campaign" with Scope3 as buyer agent for MiQ across LG and Warner Bros. Discovery inventory in Q4 2025 (#8, #96) — a single named demonstration, described by Magnite itself only in general terms ("interpret," "match," "transact") with no disclosed negotiation mechanics or deal terms (#97). By April 2026, Magnite generalized this into platform-wide "Intelligent Assistance" and "Agentic Execution," with named activations across Kepler, MiQ, Disney Advertising, Publicis Media Exchange, and Spectrum Reach (#9), then launched "Magnite Orchestration" in June 2026 as a coordination layer for third-party buyer agents, in beta with Dentsu and DIRECTV (#37). PubMatic's AgenticOS, live since January 2026, had run over 80 agentic campaigns across all five global holding companies by mid-2026, with CEO Rajeev Goel projecting ~25% of PubMatic's ecosystem trading agentically by 2028 (#31); its partnership with Optable lets a sell-side agent package first-party audience data for a buyer agent, but the deal itself executes through "PubMatic Activate, a direct-to-supply media bidder" (#99) — conventional real-time bidding under the hood, not a distinct negotiated-deal mechanism, despite marketing language describing it as agent-to-agent negotiation (#35, #100). The underlying protocol both companies build on, AdCP, confirms this structurally in its own technical documentation: a linear get_products → syncCreatives → createMediaBuy sequence with no counter-offer or negotiate task — "not truly bilateral negotiation," per AdCP's own docs (#98). Underpinning all of this, IAB Tech Lab released the Agentic RTB Framework (ARTF) v1.0 for public comment in November 2025 and finalized it by March 2026 (#40, #87) — an industry-wide, co-located-auction standard backed by Index Exchange, Netflix, Paramount, The Trade Desk, and Yahoo, meant to cut bid latency by up to 80%. But ARTF is not the only open standard racing to become that layer: a rival protocol, AdCP, governed by the independent AgenticAdvertising.org (127+ member companies, including ARTF backers Yahoo and PubMatic), ran the industry's first real agent-to-agent transaction — LG Ads DOOH inventory, October 2025 — before ARTF was even finalized (#89). By August 2026, IAB Tech Lab's own COO was publicly arguing the two standards are complementary rather than redundant, and the one cited real-world production example, VIOOH, hedges by registering under both rather than adopting either exclusively (#90).

DSP-side AI buying platforms. The Trade Desk's Kokai is the most extensively documented case in this research pass: adoption went from "nearly all" clients trying it and 85% using it as default (Q3 2025, #15) to "almost 100%" running through it (Q4 2025, #16), with company-reported average gains of 26% better CPA, 58% better cost per unique reach, and 94% better CTR versus its prior platform (#17), plus a string of named-advertiser case studies (SpecSavers, Danone, Ikea, Best Western, Cheerios — #20, #21). Deal Desk, its AI deal-forecasting product, reports deals performing ~35% better than the legacy approach after historically ~90% of deal IDs never scaled (#19). Amazon Ads' Performance+ automates DSP campaign setup and optimization via first-party signals (#24); its November 2025 "Ads Agent" builds campaigns from natural language or an uploaded media plan, auto-generates Amazon Marketing Cloud SQL queries, and by Q2 2026 users saw 8% lower cost-per-impression and 6% lower CPA than non-users, per Amazon's own comparison (#25, #50). Google positioned Gemini as DV360's "core operating layer" at its March 2026 NewFronts, auto-translating an uploaded media plan into a full campaign (#26). Yahoo DSP's agentic layer, built on "Yahoo Blueprint" (evolved from the earlier AdLearn engine), includes a live Troubleshooting Agent for pacing/delivery issues (#27). Criteo expanded its self-serve "GO" platform with an AI Onboarding Agent that can launch a cross-format campaign in five clicks (#70), aimed explicitly at SMBs who can't build equivalent bidding/targeting in-house. Basis Technologies added AI-driven unified reporting and automated paid-search activation for agencies (#29). Microsoft, meanwhile, is shutting its Xandr/Invest DSP by early 2026 to replace it with a not-yet-built Copilot-powered buying product, while its sell-side Monetize business continues unaffected (#11, #28).

Mobile, in-app, and retail-media ad networks. AppLovin's AXON engine is the strongest single-company AI story in the domain: CEO Adam Foroughi credited AXON 2.0 for a record $406M software-platform revenue quarter in 2023 (#58); the company's FY2024 10-K names AXON's algorithmic efficacy as a primary driver of results (#59); by FY2025 AppLovin reported $5.481B revenue (+70% YoY) and $3.334B net income after divesting its entire mobile-gaming business to become a pure-play ad-tech company built around AXON and its separately-sold mediation product, MAX (#60, #61); in June 2026 AppLovin opened AXON to full self-serve sign-up after 14 years as a closed platform (#62). Unity's Vector, a full rebuild of its prior ad-targeting engine after a 2022 corrupted-data failure cost ~$110M in guidance and ~37% of its stock in a day (#65), reached general availability in May 2025 (#66) and drove Grow Solutions revenue to $389M (+35% YoY) with Vector itself surpassing a $1B annualized run rate by Q2 2026 (#67). Kevel's "Kai" suite adds ML forecasting (Kevel Forecast) plus a bring-your-own-model feature that lets retail media networks plug their own proprietary models directly into Kevel's decisioning loop (#13). Criteo's Commerce AI predicts conversion probability per user-product pair, trained (per Criteo's own marketing claim) on over $1 trillion in cumulative commerce data (#68); its Retail Media auction-based on-site display format grew 65% and reached ~21% of on-site media spend across 49 live retailers in Q4 2025 (#69).

Platform-native optimization (walled gardens). Meta's Advantage+ end-to-end tools reached a $75B annual revenue run rate in Q2 2026 (#46), with over 9 million small businesses using at least one AI creative tool (#47) and named case studies like Indian apparel retailer Underneat (13% incremental purchase lift, #48); separately, Meta's own internal ad-ranking R&D — upgraded user-understanding and GEM sequence-learning models — drove an 8.3% increase in Facebook ad clicks and 15.7% uplift in conversions (#49), a genuinely different (institutional, Meta-controlled) layer from the commoditized Advantage+ tools advertisers buy on top of it. TikTok claims Smart+ delivers a 53% average ROAS improvement, explicitly positioned to close the gap with Meta (#52) — a vendor claim, not an independently verified figure. On Google Search, AI Max/Performance Max adoption reached just over 30% of customers as of Q1 2026 (#44), with Hilton EMEA cited as a showcase advertiser (#45); Microsoft's Performance Max now auto-imports campaigns directly from advertisers' Google Ads accounts, with ADAC Car Insurance cited as a ~600% ROAS case study (#51).

CTV-specific tooling. Roku is pairing geographic targeting with "AI-driven creative pipelines" for localized ad variants in 2026, predicting (its own forecast, not a verified outcome) a 30% effectiveness boost versus linear TV for local advertisers (#14). FreeWheel's Context Engine, launched April 2026 on Comcast's VideoAI, generates contextual video signals (emotions, objects, keywords) for cookie-free targeting across FreeWheel's Streaming Hub (#4). Index Exchange integrated xpln.ai's ML-based "Attention KPI" — trained on eye-tracking data across 20-25 exposure signals — directly into pre-bid SSP decisioning in February 2026, shifting attention measurement from post-campaign reporting into a real-time, actionable signal (#39).

Identity infrastructure. With Chrome's third-party cookies staying live by explicit 2025 reversal (#76) and Google retiring more than ten Privacy Sandbox APIs for low adoption in the same announcement (#77), the field has gone to deterministic and proprietary identity graphs rather than a browser-native standard: The Trade Desk's Unified ID 2.0 is adopted by Warner Bros. Discovery, Roku, and LG Ad Solutions (#78) and now sits inside Kokai's full decisioning stack via "Identity Alliance," which CEO Jeff Green called "the very essence of what it means to be a DSP" (#79). LiveRamp launched "agentic orchestration" in October 2025, giving AI agents governed access to its identity resolution and clean-room tools across roughly 900 partners (#80); Publicis agreed to acquire LiveRamp outright for ~$2.17B in May 2026, explicitly framed around owning "data co-creation... in the age of artificial intelligence" (#81) rather than renting it.

Quality, attention, and fraud tooling. OpenX's "Results" suite claims (vendor figures, unaudited) up to 23x higher CTR and 11x improved CPA for one campaign (#41); its June 2026 "OpenX IQ" suite bundles inventory-quality scoring, audience/attention modeling, and selling agents into a standardized decisioning product (#42). DoubleVerify's November 2025 "DV AI Verification" analyzes ~2B AI-agent interactions per month to distinguish declared legitimate agents (ChatGPT, Claude, Perplexity) from evasive scraper bots (#84). Smaller vendors (CHEQ, Pixalate, Adloox) market comparable ML-based fraud detection, with no independent verification found of their specific performance claims (#85).

Internal engineering productivity. The one clear Personal Productivity data point in this domain: The Trade Desk's CEO disclosed that "every engineer at TTD is using AI tools to write and/or test code" (#22) — individual-level coding-copilot adoption, distinct from the institutional Kokai/Koa platform investment sitting one layer up.

Where the investment is concentrated

The bulk of activity described above is Process Automation at Scale: a standardized tool (copilot, auto-bidder, auto-campaign-builder) sold identically to every customer of a platform, automating a job — bidding, reporting, campaign setup, moderation — that already existed. See `investment-landscape.md` for the full initiative-by-initiative quadrant mapping.

Unusually for a market this dominated by standardized tooling, several genuine Institutional Capability candidates emerged, and they cluster in one specific place: proprietary decisioning/auction engines built over multiple years, not shipped as a feature. AppLovin's AXON (#58-61, #104) and Unity's Vector (#65-67) are the domain's strongest evidence — both run on fully owned, closed-loop data (each company's own mobile SDK telemetry) that no rival can access, and both correlate with externally verifiable, sustained outsized financial results. Comcast/FreeWheel's VideoAI (#4-5) fits the same pattern. The Trade Desk's Koa/Kokai architecture with Identity Alliance (#18, #79) was included in this category in the original research pass, but a follow-up pass found this doesn't hold up as cleanly: Morningstar formally downgraded TTD's economic moat from "narrow" to "none" in March 2026 (#101), reasoning that a pure-play DSP's AI decisioning isn't sufficient without owned supply, auction mechanics, and behavioral data — and that TTD's identity asset, UID2, functions as an open, multi-tenant standard other DSPs can also use, not an exclusive network (#106). TTD's stock and growth continued deteriorating through August 2026 (#102, #103), roughly three years after Kokai's 2023 launch, with none of AXON's or Vector's inflection. The comparison is now this domain's clearest evidence that a proprietary decision engine alone is not enough — it needs genuinely exclusive data underneath it, not just AI branded as proprietary. PubMatic's Activate/Optable flywheel ("better outcomes attract more advertisers, more campaigns and more data, creating a compounding advantage" — #34, #35) and LiveRamp's identity graph, now being acquired outright by Publicis rather than merely licensed (#80, #81), fit the exclusive-data shape better than TTD's UID2 does. Meta's internal ad-ranking models (#49) are institutional capability for Meta as the platform operator, even though what advertisers buy on top of them (Advantage+) is commoditized for the buyer.

The newest layer — agent-to-agent transacting infrastructure — looked at first like a split between the two invention quadrants, but a follow-up pass complicates that reading. FreeWheel's MCP pilot with a single agency partner (#6) and Magnite's first AdCP demonstration transaction (#8) still look like Individual Invention: real, but bespoke and unrepeated at the time of disclosure, and — per management's own later description — not a documented negotiation, just an early proof of concept (#96, #97). Magnite's subsequent platform-wide rollout (#9, #37) and PubMatic's AgenticOS-plus-Optable integration (#31, #35) generalized into multi-partner infrastructure, which looks closer to Institutional Capability on the surface — but what's actually being standardized is a thin discovery-and-approval protocol with no negotiation step (#98), executed through each side's existing RTB machinery (#99). That protocol layer is not where the moat is forming; it's exactly the kind of interoperability plumbing that standardizes the way Prebid.js standardized header bidding (#91-95), and it is already racing between two rival open standards (ARTF and AdCP, #87-90) rather than settling into one company's proprietary version. Whether any single company's decisioning intelligence sitting behind that commoditizing protocol becomes a durable moat — the way AppLovin's and Unity's did, and TTD's arguably has not — is this report's refined thesis for the next phase (see `01-brainstorm/grill-me-log.md`).

Blindspots & challenges

Antitrust and structural risk sit above the whole domain. A US federal court found in April 2025 that Google illegally monopolized the open-web publisher ad server (DFP) and ad exchange (AdX) markets and unlawfully tied them together (#12, #71); as of this research date the remedies ruling is still pending, with DOJ seeking divestiture of AdX (and possibly DFP) and Google proposing behavioral remedies instead (#72). The European Commission separately fined Google €2.95B in September 2025 for the same underlying conduct — finding Google informed AdX in advance of rival bids it needed to beat, since at least 2014 (#73) — with a Commission official saying only a structural remedy (a forced sale) may be sufficient (#74); Google is appealing that decision on 17 grounds with no suspensory effect (#75). Whoever ends up owning ad-serving infrastructure at all could be determined by these rulings, independent of AI capability.

Advertiser "black box" complaints are concrete and growing, not fringe. Digiday documented advertisers actively cutting Performance Max spend — one agency client cut Google spend 50%, another halved Performance Max spend after pacing swung "almost nothing" to "double or triple the budget" with no diagnosable cause (#53). An independent Haus study of 640 Meta geo-lift tests found Advantage+ beat manual campaigns in only 42% of tests, and even when it won, delivered 12% lower incremental ROAS at 18% lower spend than manual (#54) — directly undercutting the platforms' own self-reported case studies. A 1,306-respondent global PPC practitioner survey found 62% cite "opaque, black-box platforms" as their top challenge and 53% say the job has gotten harder in two years (#55). Google's own May 2025 response — adding channel-level, search-term, and asset-level reporting to Performance Max — was reported as a direct concession to this pressure (#57), though it makes the black box visible after the fact rather than transparent in real time.

AI-generated ad-creative adoption is outrunning consumer trust. IAB's tracking survey found 83% of ad executives report deploying AI in the creative process (up from 60% in 2024), but the gap between executives' belief that consumers feel positive about AI-generated ads (82%) and consumers' actual sentiment (45%) widened to 37 points (#56) — a trust deficit sitting underneath the same Process-Automation-at-Scale tools described above.

Fraud and invalid traffic remain an adaptive arms race, not a solved problem. The ANA's 2023 audit found ~21% of audited impressions were made-for-advertising inventory and only $0.36 of every advertiser dollar reached the consumer (#82) — still the standard reference baseline, with no comparably rigorous update found. HUMAN Security's own 2026 telemetry (a vendor claim about its own platform, not independently audited) found bot traffic growing 23.5% YoY versus 3.1% for human traffic, and traffic from AI agents/agentic browsers specifically growing 7,851% YoY (#83) — the offense scaling faster than the defense by the defender's own measurement. DoubleVerify's response was to start classifying legitimate declared AI agents apart from evasive scrapers rather than claim to have solved detection (#84).

Consolidation is real but not uniformly AI-driven. Equativ's 2024 acquisition of Sharethrough was explicitly attributed by Adweek to competing "among undifferentiated SSPs" — evidence for, not against, the domain's commoditization thesis (#43). By contrast, MediaMath's June 2023 bankruptcy was checked specifically for an AI-capability angle and found none — contemporaneous coverage attributes it to failed M&A and rising interest rates, a reminder that not every industry failure in this space is an AI story (#23). Microsoft's decision to shut its Xandr/Invest DSP was justified in the company's own words by an AI-strategy rationale ("not achievable with the industry's current DSP model") even as it consolidates around sell-side products instead (#11, #28).

Data-practice and model-fragility incidents. Short sellers alleged in March 2025 that AppLovin's AXON relied on prohibited cross-platform "fingerprinting"; AppLovin denied it, and the SEC's resulting investigation concluded in mid-2026 with no enforcement action (#63, #64). Unity's 2022 Audience Pinpointer failure — corrupted training data from one large customer degrading the whole targeting model — is a concrete illustration of how a single automated system's failure can cascade across an entire ad business ($110M in lost guidance, ~37% stock drop in a day) before being rebuilt as Vector (#65).

A proprietary AI decision engine, on its own, is not a proven moat outside mobile's closed ecosystem. The Trade Desk's Kokai rollout coincided with the company's first revenue miss in 33 quarters, a >30% single-day stock crash, and securities-fraud litigation alleging TTD overstated Kokai's readiness (#102); by Q2 2026, growth had decelerated to 3% YoY with a 7-year stock low and an HSBC "reduce" rating (#103). Morningstar formally downgraded TTD's economic moat from "narrow" to "none" in March 2026 (#101). PubMatic shows the same pattern: its loudest "AI is transformational" marketing coincides with its weakest financial results (Q1 2026 revenue down 2% YoY, EBITDA margin collapsing from 13% to 4%, management calling agentic campaigns an "immaterial percentage of the business"), while Magnite — which explicitly frames AI as an efficiency tool rather than a moat — posts steadier if unspectacular growth (#105). No open-web/CTV company examined shows AppLovin- or Unity-level convergence between AI claims and hard-to-replicate financial outcomes.

Flagship reports now in this research folder

Unlike domains anchored by big analyst-firm outlooks, this domain's evidence base leans heavily on primary company disclosure — earnings-call transcripts (Alphabet, Meta, Amazon, Microsoft, The Trade Desk, Magnite, PubMatic, AppLovin, Unity, Criteo) and SEC filings (AppLovin's FY2024 10-K and FY2025 8-K, #59, #61) — because most of the players making AI claims are public companies disclosing directly rather than being profiled by a third-party analyst firm. The genuine industry-report/primary-document layer drawn on:

• IAB Tech Lab, Agentic RTB Framework (ARTF) v1.0 (Nov 2025) — the industry's proposed co-located auction standard for an agentic bidding era (#40)
• IAB, "The AI Ad Gap Widens" (Jan 2026) — Insights Engine/Attest survey tracking AI ad-creative adoption vs. consumer trust (#56)
• ANA, 2023 Programmatic Media Supply Chain Transparency Study (Jun 2023) — the still-standard MFA/waste baseline (#82)
• State of PPC 2026 Global Report (Jan 2026, PPCsurvey.com/TrueClicks, 1,306 respondents) — practitioner-level "black box" quantification (#55)
• HUMAN Security, 2026 State of AI Traffic & Cyberthreat Benchmark Report (Apr 2026) — vendor-published but widely-cited bot-traffic telemetry (#83)
• US v. Google (E.D. Va.) liability ruling (Apr 2025) and European Commission Case AT.40670 decision (Sep 2025) — the two primary legal documents anchoring the antitrust section (#12, #71, #73)

What's still unverified

Company-selected, single-advertiser case studies dominate the performance-claim evidence base — Meta's Underneat (#48), TTD's SpecSavers/Danone/Ikea/Best Western/Cheerios (#20, #21), Amazon's Ads Agent CPA figures (#50), Microsoft's ADAC ROAS figure (#51), OpenX's 23x/11x claim (#41), TikTok's 53% Smart+ claim (#52), Criteo's $1 trillion training-data claim (#68) — none independently audited. The one independent (non-platform-reported) incrementality check found, the Haus/Meta study (#54), points in the opposite direction of the platforms' own narrative but was flagged by its own author as predating a subsequent Meta ranking-model upgrade and comes from a vendor with a commercial interest in the finding — suggestive, not dispositive. Several primary sources (PubMatic and Magnite investor-relations pages, Magnite's 10-K) blocked direct fetch in this research pass and had to be corroborated via secondary trade-press coverage instead (see notes on #30, #32, #33, #35). Revenue and margin growth cited throughout this report (e.g., Magnite's CTV contribution ex-TAC growing 36% YoY in Q2 2026, #38) often reflects broader CTV ad-spend growth as much as any specific AI initiative — the claims above should not be read as proof that AI itself is the primary growth driver just because a company mentions both in the same earnings call. Both the US and EU antitrust remedies decisions remain undecided as of this writing — any characterization of a forced Google breakup as "ordered" would be premature. Whether Magnite Orchestration or PubMatic's AgenticOS becomes a durable proprietary moat, or gets commoditized by the open IAB ARTF standard, is an open question with no evidence yet either way.

Status & next steps

This closes the `00-research` phase for Ad Servers (this file plus `sources.md`, `investment-landscape.md`, `industry-challenges.md`, and `unsolved-problems.md`) and reflects one full grill-me session plus a follow-up research pass resolving its open questions (`01-brainstorm/grill-me-log.md`). The thesis that survived grilling, refined by that follow-up pass:

The durable AI moat in ad-serving does not sit in the orchestration/transacting layer — that layer is standardizing fast, the way Prebid.js standardized header bidding under neutral non-profit governance while the SSPs plugging into it stayed commoditized (#91-95). If anything, agentic orchestration is standardizing even faster and more contested than header bidding did: within nine months of one open standard's launch, a rival open standard had already run the industry's first real transaction, and the standards body itself had to publicly argue the two aren't redundant (#87-90). And what these protocols actually do isn't negotiation — it's a discovery-then-approve handshake running through the same RTB machinery underneath (#96-100). The moat sits where The Trade Desk's own CEO puts it in his own words: not in open ecosystem access, but in proprietary decisioning plus genuinely exclusive data (#106, #107). But "exclusive" is the operative word, and it isn't automatic — Morningstar's formal moat downgrade of The Trade Desk itself (#101), specifically because its AI decisioning sits on an open identity standard rather than closed proprietary data, is the domain's clearest evidence that decisioning intelligence alone is not enough. AppLovin's AXON and Unity's Vector remain the sharpest moat evidence precisely because they run on fully owned, closed-loop data no rival can access (#58-61, #65-67, #104) — TTD is the cautionary case proving what's missing without it.

Researched deliberately without treating AI as the lens — before testing any constraint against AI, per the framework's "hunt the impossible" step. The point is to name the industry's real structural problems on their own terms, not to reverse-engineer problems that happen to be shaped like whatever AI tools are for sale. Cited claims reference `sources.md` by number; several points below are general industry-structure background knowledge rather than claims on the sourced list — those are marked [background, uncited] and should not be treated as sourced facts.

The finding: one root structural conflict, three direct consequences

Ad-serving infrastructure has one problem sitting underneath almost everything else: the largest player in the market operates on both sides of the transaction it referees — running the dominant publisher ad server and the dominant ad exchange, deciding in real time whose bid wins. Three of the industry's other recurring complaints (competitive thinning, opacity/trust erosion, and the fraud arms race) are best read as downstream consequences of a market built around that same basic shape — concentrated, self-dealing infrastructure — rather than independent root causes.

1. Vertical integration and conflict of interest at the infrastructure layer — root cause

The core structural fact: the entity operating the auction has historically also been a bidder in it, or has favored its own exchange when running someone else's auction.

• A US federal court found Google illegally monopolized both the open-web publisher ad server market (DoubleClick for Publishers) and the open-web ad exchange market (AdX), and unlawfully tied the two — preventing publishers from getting real-time AdX bids unless they also ran DFP (#12, #71). This is not an allegation; it is an adjudicated liability finding as of April 2025.
• The European Commission separately fined Google €2.95 billion for the same underlying conduct: favoring AdX in DFP auctions, including informing AdX in advance of rival bids it needed to beat, since at least 2014 (#73).
• Both regulators are still deciding what to do about it as of this research date: the US remedies ruling is pending, with DOJ seeking AdX divestiture and Google countering with narrower behavioral remedies (#72); the EU has floated a structural remedy — "selling some part of its Adtech business" — while Google appeals the finding entirely (#74, #75). Neither case is resolved. This is the single largest open variable in the entire domain: any AI-driven decisioning advantage a vendor builds today is being built on top of infrastructure whose ownership structure may be forcibly restructured within the next one to two years.

Why this is root, not a symptom: it's not a temporary bad-actor problem — it's the incentive structure the market was built on. An ad server that both runs the auction and has a stake in who wins it will structurally prefer itself, absent either competitive pressure or regulatory constraint, regardless of how sophisticated its AI gets.

2. Thinning, undifferentiated competition among independent (non-walled-garden) vendors — consequence of #1 and of commoditization

Outside the two dominant platforms under antitrust scrutiny, the independent ad-tech middle has been consolidating and, in at least one documented case, failing outright.

• MediaMath, once a major independent DSP, filed for Chapter 11 bankruptcy in 2023, owing over $100M to creditors (mostly SSPs and publishers) after acquisition talks collapsed — contemporaneous coverage attributes this to financing/credit conditions and failed M&A, not an AI-capability gap (#23).
• Microsoft is exiting the DSP market entirely (Xandr/Invest shutting down by early 2026), citing that "the industry's current DSP model" isn't compatible with its future strategy, and redirecting investment to sell-side products instead (#11, #28).
• Equativ's acquisition of Sharethrough was characterized by trade press explicitly as consolidation to "compete among undifferentiated SSPs" (#43) — a direct statement that the mid-market SSP layer has stopped being able to differentiate on its own.

Why this reads as consequence, not an independent root cause: each of these is a company responding to a market where the largest, structurally-integrated player captures disproportionate value (#1) and where the remaining layer competes on an increasingly commoditized feature set (see `investment-landscape.md`'s Process-Automation-at-Scale concentration) — not enough differentiation to sustain a crowded field of independent vendors.

3. Publisher/advertiser trust erosion from black-box decisioning — consequence, but with its own now-substantial evidence base

Even independent of any specific AI failure, advertisers and publishers increasingly report they cannot see or verify how their money is actually being allocated inside the platforms they depend on.

• 62% of PPC professionals surveyed across 50+ countries cite "increasingly opaque, black-box platforms" as their top challenge; 53% say managing PPC is harder than it was two years ago (#55).
• Named agencies have acted on this: one advertiser cut Google spend 50% (mostly Performance Max) and shifted it to the open web; another cut Performance Max spend in half for a client after erratic, unexplainable pacing (#53). An independent study of 640 Meta incrementality tests found Advantage+ beat manual campaigns in only 42% of cases, with caveats about the study's currency and the vendor's commercial interest in the finding (#54).
• Google itself conceded ground by adding channel-level reporting and search-terms visibility to Performance Max in 2025, explicitly in response to "persistent criticism... about Performance Max's 'black box' nature" (#57) — evidence the platforms know this is a real commercial problem, not just noise.

Why this reads as consequence: opacity is a natural output of concentrated, self-interested infrastructure (#1) combined with a commoditized-automation business model that has every incentive to abstract away the mechanics it's selling (see `investment-landscape.md`). It would exist even without any AI — it is the decades-old "black box of programmatic" complaint — but AI-driven auto-bidding has intensified rather than resolved it.

An independent structural challenge: the ad-fraud and invalid-traffic arms race

Unlike the three points above, this does not trace cleanly back to vertical integration — it is a persistent, adversarial problem inherent to any large-scale, automated, real-time bidding marketplace, and it is getting worse for a new reason (AI agents), not a familiar one.

• A 2023 industry-association audit of $123M in spend across 21 major advertisers found made-for-advertising inventory at ~21% of audited impressions (~$13B/year industry-wide) and confirmed invalid traffic on top of that; only 36 cents of every dollar entering a DSP reached the consumer (#82). This figure is now three years old and has not been refreshed, so should be read as a dated baseline, not a current-state fact.
• A newer vector is emerging on top of the old one: one verification vendor's telemetry shows AI-agent and agentic-browser traffic growing 7,851% year-over-year against a broader bot-traffic growth rate of 23.5% (vs. 3.1% for human traffic) — roughly 8x faster than legitimate traffic growth (#83, vendor-reported, flagged accordingly). Verification vendors have begun launching products specifically to distinguish "declared" AI agents (ChatGPT, Claude, Perplexity acting as a proxy for a real consumer) from malicious undeclared scrapers (#84) — a genuinely new classification problem that didn't exist in this form a few years ago.

Why this is listed separately: it isn't downstream of the DFP/AdX conflict-of-interest story — it would exist in a perfectly competitive, non-monopolized ad-serving market too. It is a standing tax on the entire marketplace that both walled gardens and the open web pay, and the AI-agent-traffic surge is a new and currently unresolved twist on it.

Two structural forces considered and set aside as background context, not top-tier findings

• Walled-garden dominance of the open web. [background, uncited as a single figure, but consistent with sourced revenue scale] Google, Meta, and Amazon's advertising products are large enough that Meta's Advantage+ alone runs at a >$75B annual revenue rate (#46) and Amazon's Ads Agent has already expanded across 11+ additional countries within six months of an initial US launch (#50) — scale that dwarfs any single independent ad server or SSP in this evidence set. This is real and shapes everything else in the domain (it is the backdrop against which #1's antitrust fight is being fought), but it reads as the setting the other four problems play out in, rather than a distinct, independently diagnosable problem with its own binding constraint.
• Post-cookie identity fragmentation. Contrary to the "cookies are dead" narrative that shaped ad-tech strategy for years, Google reversed course and is not deprecating third-party cookies in Chrome, retiring most of its own Privacy Sandbox replacement APIs instead (#76, #77). The practical result is a fragmented identity landscape — cookies persist by default in Chrome but are blocked by default in Safari/Firefox/Brave — with industry-run deterministic networks like UID2 (#78) and LiveRamp's RampID (#80) filling the gap browser-by-browser rather than a single standard emerging. This is a real and current structural headache for anyone building cross-platform AI decisioning, but it is more accurately described as unresolved plumbing than a business-model-level challenge on par with the four above.

What this means for the domain analysis

This AI-agnostic pass should anchor, not be replaced by, whatever binding-constraint analysis follows in `unsolved-problems.md`. The throughline worth carrying forward: nearly every AI capability catalogued in `investment-landscape.md` — better bidding, better yield, better fraud detection, better creative — is being built and sold on top of an infrastructure layer whose basic ownership structure is legally contested (#1) and whose competitive middle is thinning (#2). An AI feature that makes a conflicted auction run faster or a black box harder to see through does not resolve either problem; it can even make problem #3 (trust erosion) worse by adding another layer of automated opacity on top of an already-opaque incentive structure. Whether any AI investment in this domain constitutes a genuine new capability or just faster automation within a structurally compromised system is the question worth pressing hardest in the next research phase.

Problems long considered intractable. For each: what was the binding constraint, and does AI remove it — or just speed up work within it?

Problem 1: Simultaneous auction transparency for both buy- and sell-side

• Why it was considered unsolvable: Whoever operates the ad-serving/exchange infrastructure that clears an auction has both the technical ability and the commercial incentive to see more of the auction than the parties transacting on it — and historically there was no neutral third party clearing the trade. This isn't a modeling problem; it's a conflict of interest baked into who owns the pipes.
• Binding constraint: Structural information asymmetry held by the infrastructure owner, which that owner has no commercial incentive to give up voluntarily.
• Does AI remove the constraint? Mostly not yet, and the strongest evidence is regulatory, not commercial. The European Commission's €2.95bn decision found Google's DFP-AdX integration let Google inform AdX in advance of rival bids it needed to beat, since at least 2014 — the asymmetry was engineered directly into the pre-AI auction plumbing (#73), and the same tying conduct was the core finding in the April 2025 US liability ruling (#12, #71). The AI-era "transparency" features that have shipped since — Google's channel/search-term/asset-level reporting expansion in Performance Max (#57), Ask Ad Manager's natural-language reporting (#1) — add visibility into what already happened after the auction cleared. That is automation of reporting, not removal of the underlying asymmetry in how the auction itself is run. Buy-side evidence points the same way: 62% of practitioners still cite "opaque, black-box platforms" as their top challenge (#55), and advertisers are responding by cutting spend on automated products they can't audit rather than trusting new AI dashboards to fix it (#53) — a sign the constraint is being worked around, not removed, years into heavy ML investment in the same platforms.
• Evidence anyone is attempting real removal: IAB Tech Lab's Agentic RTB Framework (ARTF), backed by Index Exchange, Netflix, Paramount, The Trade Desk, and Yahoo, reframes the transaction as a co-located, explicit exchange rather than an opaque internal waterfall (#40, #87). Magnite's and PubMatic's agent-to-agent deals (#8, #9, #35) were originally read as going further structurally — a named buyer agent and a named seller agent seeing and agreeing to specific terms, mechanically different from bidding blind into a black-box auction. Correction (2026-08-22 follow-up pass): this reading doesn't survive closer evidence. AdCP's own technical documentation defines a linear get_products → syncCreatives → createMediaBuy sequence with no counter-offer or negotiate task — "not truly bilateral negotiation," per the spec itself (#98). Magnite's own CEO described its flagship deal only in general terms with no disclosed negotiation mechanics (#96, #97), and PubMatic/Optable's deal executes through "PubMatic Activate, a direct-to-supply media bidder" (#99) — conventional RTB underneath. The agent sees a richer, structured product description before bidding; it still doesn't negotiate, and the auction clearing mechanism itself is unchanged. That's automation of discovery, the same category as the reporting-layer automation above, not a removal of the structural information asymmetry.
• Institutional-capability formulation: Transparency would need to live in the transaction protocol itself — an explicit, inspectable, negotiable agreement between named buy- and sell-side agents, ideally on an open standard no single infrastructure owner controls — rather than in a reporting dashboard bolted onto an opaque auction after the fact, or a richer discovery query bolted onto an unchanged auction. No protocol examined in this research (ARTF, AdCP, or any proprietary implementation) has shipped the negotiation step this would actually require.
• Verdict: ☑ Automation in disguise, for everything shipped as "AI transparency" to date — both the reporting layers (unchanged auction, more visibility after the fact) and the agent-to-agent "negotiation" claims (unchanged auction, richer discovery before the fact, per the protocols' own documentation). No evidence found anywhere in this domain of the underlying information asymmetry actually being removed rather than worked around or reported on more legibly.

Problem 2: Cross-platform identity resolution without cookies or device IDs

• Why it was considered unsolvable: Recognizing the same user across publishers and platforms at fine grain requires either a shared ground truth no single company controls, or an invasive tracking mechanism that regulators and browsers are actively closing off. No amount of modeling skill substitutes for parties who won't share the underlying data.
• Binding constraint: A coordination and consent problem across many independent controllers of the user relationship (browsers, publishers, advertisers) — not a technical resolution problem.
• Does AI remove the constraint? No — and the six-year browser-native attempt to solve it architecturally, without relying on any party's proprietary data, just failed. Google reversed its own cookie-deprecation plan and confirmed Chrome will keep third-party cookies live by default (#76), and in the same October 2025 update retired more than ten Privacy Sandbox APIs — including the interest-based Topics API — for low adoption (#77). That failure wasn't a modeling failure; publishers and advertisers simply didn't build on top of the privacy-preserving alternative. What scaled instead is deterministic, consent-based identity: The Trade Desk's Unified ID 2.0, adopted by Warner Bros. Discovery, Roku, and LG Ad Solutions (#78), and LiveRamp's identity graph, which Publicis is acquiring outright for ~$2.17B specifically to own "data co-creation... in the age of artificial intelligence" rather than license it (#80, #81). In both cases, AI (Kokai's "Identity Alliance," #79; LiveRamp's "agentic orchestration," #80) sits on top of an already-solved consent/matching layer as an access and activation interface — it is not the mechanism that resolves identity across non-cooperating parties.
• Evidence anyone is attempting it: Every example found (#78, #79, #80, #81) is AI layered on top of deterministic, consented identity infrastructure that someone already owns outright — none is AI inferring identity across parties who have not agreed to share data.
• Institutional-capability formulation: The moat here was never going to be a smarter matching model; it's outright ownership of a consented identity graph plus enough network scale (LiveRamp's ~900 partners) that being outside it is a competitive disadvantage — which is exactly why Publicis chose to buy the asset rather than license the API. Confirmed by a follow-up finding (2026-08-22): The Trade Desk's UID2, cited above alongside LiveRamp, turns out to be the negative case proving this distinction — Morningstar's March 2026 downgrade of TTD's economic moat to "none" reasoned specifically that UID2 functions as an open, multi-tenant identity standard other DSPs can also use, not a closed, exclusive network the way LiveRamp's graph is (#101, #106). Open/consented identity infrastructure and exclusive identity infrastructure are not the same asset, even when both are described as "proprietary" in company marketing.
• Verdict: ☑ Automation in disguise. This is the clearest test-script-trap example in the domain: AI is the interface layer on an already-solved (by consent and acquisition, not by AI) identity problem, not the thing that removed the original binding constraint. The binding constraint — no shared ground truth across non-cooperating parties — is unchanged; it has been routed around via deterministic opt-in identity and institutional ownership, not solved by AI.

Problem 3: Ad fraud and invalid traffic that adapts faster than detection can generalize

• Why it was considered unsolvable: This is an adversarial arms race, not a static classification problem — any detection signal, once learned, gets reverse-engineered and evaded, so a fraud-detection model decays the moment it succeeds.
• Binding constraint: Adversarial adaptation speed. Historically the defense had a tooling advantage over individually-operated fraud rings; that advantage disappears once both sides have access to similarly capable AI/automation tooling.
• Does AI remove the constraint? The evidence found says no — AI is currently accelerating the offense faster than the defense. HUMAN Security's own 2026 telemetry (a vendor's claim about its own platform, not independently audited, but the only recent measurement found) reports bot/automated traffic growing 23.5% year-over-year versus 3.1% for human traffic — roughly 8x faster — and traffic specifically from AI agents and agentic browsers growing 7,851% year-over-year (#83). DoubleVerify's November 2025 response is telling in what it didn't claim: rather than announce improved detection accuracy, it launched a system to classify declared, legitimate AI agents (ChatGPT, Claude, Perplexity) apart from evasive scrapers (#84) — triage and labeling, not elimination. The ANA's 2023 baseline — ~21% of audited impressions made-for-advertising, only $0.36 of every advertiser dollar reaching the consumer (#82) — remains the standard reference figure years later; no comparably rigorous update was found showing that ratio has improved despite years of AI-based verification tooling (DoubleVerify, HUMAN, CHEQ, Pixalate, Adloox) being sold and deployed at scale in the interim. Smaller vendors' AI/ML fraud-detection claims (CHEQ's "military-grade NLP," #85) are undated marketing copy with no independent performance verification found.
• Evidence anyone is attempting it: DoubleVerify (#84), HUMAN (#83), CHEQ/Pixalate/Adloox (#85) — the entire MRC-accredited verification industry has adopted AI/ML detection, and it is sold identically to every advertiser as a standardized layer.
• Institutional-capability formulation: No evidence was found of any player restructuring the underlying incentive (fraud is profitable because impressions are paid for regardless of whether a human saw them) rather than racing to detect fraud faster after the fact. A genuine constraint-removal approach would change what gets paid for, not just get better at spotting what shouldn't have been.
• Verdict: ☑ Automation in disguise — and by the defense's own published numbers, currently losing the acceleration race, not closing it. This is the domain's clearest example of AI scaling both sides of an unsolved problem at once rather than resolving it.

Problem 4: Measuring true incrementality versus what the ad server already optimized toward

• Why it was considered unsolvable: Whoever runs the auction and decisioning also tends to grade its own campaign's success, using signals shaped by the same model that made the buying decisions — a self-grading conflict of interest. Historically, isolating true causal lift required slow, expensive randomized geo-holdout experiments that the platform running the campaign had little commercial incentive to run rigorously against itself.
• Binding constraint: The party operating the ad-decisioning engine and the party measuring its success are typically the same party (or use the same underlying signals) — a conflict-of-interest problem, not a lack of measurement sophistication.
• Does AI remove the constraint? The one independent (non-platform-reported) check found in this research pass says no, and points the opposite direction from the platforms' own narrative: Haus's study of 640 Meta geo-lift tests found Advantage+ — Meta's most AI-automated campaign type — beat manual campaigns in only 42% of tests, and even when it won, delivered 12% lower incremental ROAS at 18% lower daily spend than manual (#54). That is precisely what the self-grading constraint would predict: more automated does not mean more incremental, because the same engine that buys the media also generates the metrics used to judge it. This single independent finding sits against a wall of platform-reported, company-selected case studies making the opposite claim — Meta's own Underneat lift figures (#48), The Trade Desk's SpecSavers/Danone/Ikea/Best Western/Cheerios figures (#17, #20, #21), Amazon's Ads Agent CPA comparison (#50), Microsoft's ADAC ROAS figure (#51) — every one of which is the ad server reporting on its own campaign, self-selected, none independently audited.
• Evidence anyone is attempting real removal: None found. No source in this research pass showed a platform submitting its AI-optimized campaigns to independent, adversarial incrementality testing as standard practice rather than as an occasional vendor-commissioned study. Haus itself is flagged in its own reporting as a measurement vendor with a commercial interest in finding exactly this result, and its spokesperson noted the study predates a subsequent Meta ranking-model upgrade (#54) — so the finding is suggestive, not dispositive, but it is the only independent data point against dozens of self-reported ones.
• Institutional-capability formulation: Genuine removal would require incrementality measurement to sit outside the ad server's own control — an independent, standardized holdout methodology the industry trusts the way accounting relies on external audit — rather than a platform-selected case study or a single commercially-motivated third-party report. No evidence of that infrastructure existing yet was found anywhere in this research pass.
• Verdict: ☑ Automation in disguise, provisionally. No evidence found of AI removing the self-grading structural conflict; the one independent check available suggests the opposite of what every platform's own reporting claims. This is the sharpest test-script-trap pattern in the whole domain: a large volume of company-reported percentage gains on one side, and exactly one independently-run, commercially-caveated study on the other.

Goal: map where money and attention are going in this domain, and classify it into the 2x2. "Ad Servers" here means the companies that control the infrastructure for ad delivery: publisher ad servers (Google Ad Manager, FreeWheel, Magnite SpringServe), sell-side platforms/exchanges (Magnite, Index Exchange, OpenX, PubMatic, Equativ), demand-side platforms (The Trade Desk, Amazon DSP, DV360, Yahoo DSP, Microsoft), plus the identity and verification layers (LiveRamp, UID2, DoubleVerify/HUMAN/IAS) that plug into all of them.

Funding & spend snapshot

• Total AI investment in domain: No single reliable aggregate figure was found (no CB Insights/Gartner "AI-in-adtech market size" figure survived verification in this pass — unlike fashion, no market-sizing claim made this list at all). What's verifiable instead is company-level revenue tied to AI-branded products: AppLovin's AXON-powered software-platform business drove FY2025 revenue of $5.48B (#61); Unity's Vector-driven Grow Solutions revenue hit a $1B+ annualized run rate in Q2 2026 (#67); PubMatic's AgenticOS-led "Emerging" revenue line reached ~15% of total revenue, up ~100% YoY (#32); Meta's Advantage+ alone runs at a >$75B annual revenue rate (#46).
• Largest named initiatives: The Trade Desk's Kokai (near-100% of TTD's client base within a year of full rollout — #15, #16); Google's Performance Max/AI Max (>30% of Search customers — #44); Meta's Advantage+ (>$75B run rate, #46); Amazon's Ads Agent and Performance+ (expanding to 11+ new countries in H1 2026, #24, #25, #50); AppLovin's AXON (opened to full self-serve in June 2026 after 14 years closed, #62).
• Enterprise/infrastructure spend trend: The clearest domain-specific pattern is sell-side platforms racing to stand up agent-to-agent transacting infrastructure in 2026: Magnite (SpringServe AdCP agent, then platform-wide rollout, then "Orchestration" — #7–9, #36–37), PubMatic (AgenticOS, launched Jan 2026, already ~15% of revenue by Q2 — #31–35), FreeWheel (MCP Server, #6), and Index Exchange (pre-bid attention signals, #39) all shipped agentic or AI-decisioning infrastructure inside a roughly 12-month window (mid-2025 to mid-2026). This is a genuine infrastructure investment wave, not just a marketing relabeling. Update (2026-08-22 follow-up pass): the plumbing underneath it is standardizing around two rival open protocols racing each other — IAB Tech Lab's ARTF, finalized by March 2026 (#40, #87), and the independent AdCP/AgenticAdvertising.org (127+ members, including ARTF backers Yahoo and PubMatic), which ran the industry's first real agent-to-agent transaction before ARTF was even finalized (#89). Neither protocol's own technical spec includes a negotiation/counter-offer step (#98) — both are closer to a richer, schema-based version of the same propose-then-clear job RTB already does.

Initiative classification (the key table)

PP = Personal Productivity · PA = Process Automation at scale · II = Individual Invention · IC = Institutional Capability

Claims that don't fit the 2x2 — structural/regulatory context instead

These describe market structure, litigation, or infrastructure decisions rather than a specific AI capability's productive use, and are deliberately not forced into a quadrant:

• Antitrust and market structure: the US v. Google liability ruling and pending remedies (#12, #71, #72), the EU's €2.95B fine and Google's ongoing appeal (#73, #74, #75). These bear directly on how durable any single vendor's AI-driven ad-serving advantage can be — a structural breakup of DFP/AdX would reshape who can even build an institutional capability on top of that infrastructure.
• DSP market exits: Microsoft shutting down Xandr/Invest (#11, #28) and MediaMath's 2023 bankruptcy (#23) are both consolidation/financial-structure events; contemporaneous coverage explicitly does not attribute either to an AI-capability gap.
• SSP consolidation without an AI rationale: Equativ's acquisition of Sharethrough (#43) is framed by trade press as a response to SSP-market commoditization, not a capability play — arguably evidence for the framework's thesis (an undifferentiated middle of the market consolidates) rather than a specific AI investment to classify.
• Cookie/identity infrastructure: Google's reversal on deprecating third-party cookies and retirement of most Privacy Sandbox APIs (#76, #77) is a platform-policy fact that changes what identity signal AI decisioning systems have to work with, not itself an AI investment. UID2's adoption (#78) is the plumbing an AI system like Kokai's Identity Alliance runs on top of (#79), not an AI capability itself.
• Financial/market-sizing backdrop: Magnite's CTV contribution ex-TAC growth (#38) and the ANA's programmatic-waste baseline study (#82) are included as context for where growth and risk concentrate, not as evidence of a specific AI initiative's outcome.

Observed concentration

Partially confirms, partially complicates the framework's usual bottom-half pattern. By raw count, Process Automation at Scale still dominates: roughly 40+ of the ~67 quadrant-classifiable claims are PA — nearly every AI feature an advertiser or publisher can buy today (Kokai, Performance Max, Advantage+, Ads Agent, OpenX IQ, PubMatic's publisher platform, fraud verification) is a standardized tool available to any customer with a budget, exactly as the framework predicts. The "black box" backlash evidence (#53–57) is itself further confirmation: when a PA-quadrant tool's internal logic isn't visible to the buyer, advertisers experience it as a loss of control, not a capability gain — the commoditization critique in concrete, sourced form (a Crowd Louder Media client cutting Google spend 50%, #53; 62% of PPC professionals citing "opaque, black-box platforms" as their top challenge, #55).

The complication: this domain shows a denser, better-evidenced cluster of genuine Institutional Capability claims than is typical — concentrated almost entirely among the companies whose core business is the infrastructure itself rather than among the advertisers/publishers buying tools from them. AppLovin's FY2025 results (70% revenue growth, 111% net-income growth, #61) and Unity's Vector crossing a $1B run rate three years after a public AI-model failure (#65) are the sharpest evidence that owning the proprietary decisioning engine — not just reselling access to it — is where the durable value in this domain has actually landed so far.

Update (2026-08-22 follow-up pass) — the cluster is smaller and more specific than first mapped. Two of the six items originally counted in this cluster don't hold up on closer evidence. (1) The Trade Desk's Koa/Kokai architecture and Identity Alliance (#18/#79) — Morningstar formally downgraded TTD's economic moat to "none" in March 2026 (#101), reasoning that decisioning AI without owned supply/data isn't sufficient, and specifically naming TTD's identity asset (UID2) as an open, multi-tenant standard rather than an exclusive network (#106); TTD's stock and growth kept deteriorating through August 2026 (#102, #103). (2) Magnite's Orchestration layer and the generalized agent-to-agent rollout (#9, #37) — the underlying protocol (AdCP) has no negotiation step and executes through conventional RTB (#96–100); it's commoditizing plumbing, structurally the same job Prebid.js standardized in header bidding (#91–95), not a new transacting capability. What survives as clean IC evidence: AppLovin's AXON, Unity's Vector, Comcast's VideoAI, PubMatic's compounding data flywheel, and LiveRamp's governed agent-access network — all built on data the owning company controls exclusively, which the TTD case now suggests is the load-bearing ingredient, not the AI decisioning layer by itself.

Individual Invention is thin and short-lived by design: only two claims qualify (#6, #8), both single-partner pilots (FreeWheel×PMG, Magnite×Scope3/MiQ) explicitly framed by their own companies as precursors to generalized rollout — both graduated to Institutional Capability claims (#9, #37) within one to two quarters of the pilot. This is a useful domain-specific data point: in ad-serving infrastructure, Individual Invention seems to function less as an end state and more as a staging step before a capability either gets generalized (IC) or commoditized (PA), which is a faster and more visible transition than in slower-moving domains.

Personal Productivity is nearly absent from the evidence base: the only clean example found is TTD engineers using AI coding tools internally (#22) — this domain's AI investment is overwhelmingly aimed at the product (what gets sold) rather than internal staff tooling, likely because the product-market fit for "AI that runs the auction/campaign" is so much more commercially visible than "AI that helps an engineer write code."