The Evolution of Holding Companies

Over twenty years the large marketing groups bought the parts of the loop between a brand's goal and its next decision: digital skill, enterprise systems, identity and customer data and commerce operations. The original edition's records, not re-opened here, add creators, production and AI tools. This edition reads that record through dated deal, impairment and results documents. The groups are adding data, software and coordination to buying scale and specialist services. Those layers can improve client outcomes, reinforce principal-media economics or raise switching costs, and the record proves the bid, not the client's gain. Their value depends on the task: whether they improve creative quality, decisions, delivery or incremental results after the full cost of the arrangement. Buying scale persists beside the new tools. Brands keep accountability for goals and evidence while choosing the degree of operational control and portability they can justify. What a brand keeps is decided by the rights it can test, stop and take away: set the goal, act, stop the spend, challenge the proof, keep the learning. Part I gives the audited record for six groups and the alternatives. Parts II and III set the tests and three conditional paths.

A holding company is a parent firm that owns or controls several marketing businesses. This paper calls it a group. For twenty years the large groups have bought new kinds of work. They bought digital skill, enterprise systems, identity and customer data and commerce operations; the original edition's records, not re-opened here, add creators, production and AI tools. The deal records are real and dated. Publicis agreed to buy Digitas in December 2006. It closed Sapient in February 2015 and Epsilon in July 2019 [1][2][3]. Omnicom closed Flywheel on 2 January 2024 and IPG on 26 November 2025 [4][5]. Stagwell and MDC combined on 2 August 2021 [6].[1][2][3][4][5][6]fact A008 A009 A025 A036

This paper reads those purchases as parts of one loop: the loop that runs from a brand's goal to its next decision. The record proves the bid. It does not prove the client's gain. Filings show assets bought, charges taken and group margins earned. They almost never show a matched client result: profit after fees, internal cost and long-term brand effects. The one client-reported result in this review is a recognition rate, not a sales figure. It comes from a vendor release with no control group [7]. That is the gap this paper keeps open on purpose.[7]inference A028 A029 A048

Buying scale has not gone away. Principal media is media an agency buys on its own account and resells to the client. In March 2026 the ANA reported that 58% of 114 qualified client-side members had used principal media in the past year, up from 47% in its prior wave. Fieldwork ran from October 2025 to January 2026 [8]. That is a member sample, not a share of spend. It is still the only comparable pair of readings this review holds. It points up, not down.[8]attributed A061 A062 A063

The thesis is therefore narrow and testable. The groups are adding data, software and coordination to buying scale and specialist services. Those layers can do three different things: improve client outcomes, reinforce principal-media economics, or raise switching costs. The record proves the bid, not the client's gain. The value of each layer depends on the task: whether it improves creative quality, decisions, delivery or incremental business results after the full cost of the arrangement. Two effects sit behind that. Shared tools can make a group more useful to a client. The same tools can make the group's own media and data products easier to sell. The new way of doing the work can prop up the old way of earning money. Neither effect proves the other. A client must test both. What a brand keeps is decided by the rights it can test, stop and take away. Those rights explain who captures value; on their own they do not explain what creates it. A brand keeps accountability for its goals and its evidence while choosing the degree of operational control and portability it can justify.analysis A001

Control is the useful word. It has five parts. Who sets the goal? Who may act? Who can stop the spend? Who can challenge the proof? Who keeps the learning when the deal ends? No system design grants those rights by name. A contract, a log and a tested exit do. Everything that follows is built around that claim.analysis

Three paths can coexist. A brand may run the core itself. An agency may run it for the brand. A platform may run much of the work inside its own system. The right route can differ by market and by task. No route wins by name. An agency can earn a premium, a closed platform can create value, and a brand can delegate operation without giving up accountability; each is a case to test, not a rule. Part III sets out the three paths as conditions, not forecasts. It names the one decision a CMO can take in each.analysis

One more thing stays in view throughout. Cheaper production does not automate judgment: what to say, what to stop, what is distinct and what the brand should mean. Cheaper versions are not a stronger idea. That judgment is the part of the loop a brand must keep, or knowingly buy with a right to overrule.analysis A047

This is a public-source research edition. It is written for CMOs and senior leaders in media, advertising and technology. The research window runs from 1 January 2006 to 5 October 2026. It covers Publicis, WPP and WPP Media, Omnicom and the IPG lineage, Dentsu, Havas and Stagwell. It also looks at selected smaller firms, independent agencies and media platforms. It is a set of cases that matter. It is not a census of every subsidiary, affiliate or local brand.fact A004 A005

The evidence is strongest for large global groups and for US and European brand buyers. The surveys quoted are member samples: 114 ANA client-side members in 2026, 162 ANA respondents in 2023, 45 companies in the WFA and Observatory study, and 69 multinationals in the WFA 2026 study [8][9][10][11]. None is a market share or a global rate. This paper gives no full legal and financial comparison for China, India, Africa, Latin America or most Middle East markets. Conclusions about rights, proof and exit travel across markets. Prices, fee forms and legal buying roles do not.[8][9][10][11]analysis A063 A065 A068 A069

Coverage, with denominators

The audit behind this edition is dated 5 October 2026. It read named passages in 49 sources. It judged 72 selected atomic claims from the original paper and 20 more about Horizon, 92 in all. It recomputed the 32 inherited calculations; 31 reproduce, and one, a statutory-margin check, cannot be recomputed from the stored records and is not printed. The calculation ledger of this edition holds 104 numerical and comparability checks: those 32 recomputations, 48 numbers checked from the revised edition against their stored inputs, 15 task-model edge cases, 7 peer-growth observations and 2 precision rules. Only the 32 are recomputations in the strict sense. It listed all 362 paragraphs of the original edition. The original register of 232 claim records and 329 financial observations was not all audited again. Those records stay in a marked private archive. The counts the build computes for this page are 210 claim records, 51 source records and 255 correction entries.fact

In this edition's own review no source was re-opened. Network policy blocked every source host. So each of the 49 sources was rechecked only at search level: did its URL, title and date still appear in results dated on or before the cutoff? A search result is a discovery aid, not support. The stored audit passages were carried forward unread, and the source register labels them as such. Two protocol repositories were cloned through git and pinned to a commit. An external critical review dated 6 October 2026 reports that it opened 13 of these sources and reproduced all 47 arithmetic rows. Where this update relies on one of its checks, the text names the review. Those checks are attributed to it; they are not this paper’s own verification, and the re-opening of the 49 sources remains open.fact

The review used three passes: evidence, challenge and editing. One assistant made all three. They were not three separate reviews. There were no client interviews, no private fee audits, no contract reads and no live buying tests.fact

The evidence rule

Facts in this edition rest on one of two tiers. Tier 1 is a claim the audit found supported or partly supported, within its stated limits; the block names the claim id and the source. Tier 2 is a dated statement from one of the 49 register sources, written as what the source said on its date. Anything found only by search, or only in the original edition's archive, is not used as fact here. Where such a record would have mattered, the text says what this review did not show. It names the check that would settle it.fact

How to read the labels. A fact cites a source. A reported statement is what a company or body said on a date. A supported reported claim means the speaker made the claim, not that the promised result followed. Analysis is this paper's reasoning. An assumption or scenario states a condition, not a forecast. An illustrative number exists for arithmetic only. Missing public proof means not shown here. It does not mean the skill is absent. Where this paper says a brand must, should or never does something, that is a design choice of the paper, carried in a block labelled analysis, scenario, assumption or illustrative; it is not a finding about prevailing practice.analysis A045

Most confusion about groups comes from the unit of analysis. A parent owns businesses. An operating group sits inside a parent. An agency brand is a route to clients and talent; it need not be a legal company. A product is software or a service with an owner and a licence. A non-equity partnership is a contract. An equity partnership is a shared stake. A joint venture is a jointly owned business whose parents keep their own. Each has its own dates, rights and economics. The table fixes the working terms used here.analysis

Three rules follow. First, a rename is not an acquisition. When WPP Media replaced the GroupM name in May 2025, Mindshare, Wavemaker and EssenceMediacom stayed as client brands. No business changed hands [40]. Second, a retired brand is not proof that a function, a fee or a margin has ended. Omnicom's 2019 move of specialist staff into OMD, PHD and Hearts & Science kept central skills in place [35]. Third, a joint venture is not a merger. Horizon Global, announced on 29 September 2025, is a commercial joint venture of Havas and Horizon Media. The parents did not combine [33].[40][35][33]analysis A006 A022 A023 A032

Accounting labels need the same care. Gross or net revenue follows an accounting judgment about who controls the service before it reaches the client. It is not a clean proxy for the media bought for resale. The gap between the two is not a disclosed principal margin. Publicis reports €17.399 billion of revenue and €14.547 billion of net revenue for 2025. The €2.852 billion difference is pass-through cost, money billed on to clients for media and other third-party costs. It is not a profit line [43].[43]analysis A017 A018

The history reads as layers. From 2006 the groups bought digital breadth and built specialist buying teams. From the middle of the 2010s they bid for customer and enterprise work: consulting, engineering, CRM (customer relationship management systems) and identity data. From the late 2010s they merged brands and put a client officer across disciplines. The 2020s added commerce on the audited record; creators, production tools and AI rest on records not re-opened here. Each layer sat on the last. None replaced it. The dates below are anchors, not stages that every group passed on the same day.analysis

Two rules govern the table. Only events that a Tier 1 or Tier 2 record supports appear. And the status column keeps announced, closed, pending, renamed and impaired events apart. A plan, a purchase and a write-down are different facts. The original edition's timeline held about sixty events. The rows not checked again here stay in the private archive and are listed in section 31. They are not smoothed into a series.analysis

Read the table for what it does not show. It holds no price for most deals, because the audited records state none. It holds no abandoned deal. The original edition had two: the Publicis–Omnicom merger plan, ended in 2014, and the Dentsu international sale, ended in February 2026. Both are carried forward unread and listed in section 31. And it holds no client result at all. Twenty years of buying left a long record of what the groups wanted and a short record of what clients gained.analysis

The trading desk did not vanish

The trading desk is the clearest lesson in units. A trading desk is a team or system that buys media through automated auctions, which the industry calls programmatic buying. Xaxis launched in June 2011. It joined data and technology with GroupM's trading scale [42]. By 2019 Omnicom's media group had moved its specialist programmatic, search and social staff into OMD, PHD and Hearts & Science. It kept central product, engineering and data duties [35]. A name disappeared. The staff moved closer to the client teams. Nothing in either record says that a fee, a risk position or a principal margin ended with the name.[42][35]fact A006

The original edition traced seven desk lineages across the groups: Xaxis, Accuen, Cadreon and Matterkind, Amnet and Amplifi, Affiperf and Varick. That lineage is carried forward unread, listed in section 31, and not repeated here as fact. The conclusion it supports stands on the two records above: trace the function and the contract, not the name.analysis

Selection matters too. The largest surviving groups get most of the attention. Their histories alone cannot prove that scale is the best model. A fair test must also keep sold assets, impaired assets, failed plans and viable smaller firms in view. The group chapters that follow keep those counterweights beside each chain. Section 11 adds a digital-first challenger whose net revenue fell.analysis

This section carries no factual burden. It is the method the group chapters apply. The original edition set it out. The revised edition dropped it, though nothing in it was contradicted. It returns here because a dated deal record is useless without a way to read it.analysis

Six recurring motives

Deals in this field show six recurring motives. Buy a scarce skill. Buy a route to a client or a market. Buy a repeatable tool. Buy access to permitted data. Buy influence over a high-value decision. Or buy scale that lowers cost or improves terms. Most large deals mix several. A technology label can hide a large services business. A services label can hide valuable software, data processes and customer access. The first job is to name which motives a deal served, in the buyer's own words and then in this paper's.analysis

Five assessment steps

1. Start with the buyer's missing link. What could the group not do, or not do at scale, before the deal?
2. State what the target really supplied: people, clients, software, data, rights or scale. Separate the label from the contents.
3. State which rights changed. Owning a company is not owning its partners' data, its clients' permissions or its vendors' models.
4. State how the asset was joined to other work, and which result can be seen. Separate management's stated reason from this paper's inference.
5. Compare a build or partner route where one is plausible. Do not claim that the unobserved alternative would have happened.

Three value pools that can diverge

Shareholders may gain from cost removal. The acquired team may gain distribution. Clients may gain a better outcome, a lower total cost or less coordination work. None of these shows the others. A client win may support the combined offer while showing little about margin or measured effect. A goodwill write-down, a charge that cuts the book value of a past purchase, may challenge the deal thesis without proving that every acquired skill is useless. So the group chapters keep group economics and client effect in separate rows. They leave a row blank rather than fill it by inference.analysis A050

The chain template

Two rules close the method. One plausible deal logic is to enter more of the client's work: understand demand, know the customer, make the work, reach the customer, support the sale, then learn. That is a logic, not a ranking. No test in this review ranks one group's deal strategy above another, and this paper does not. The firmest cross-case inference is modest: groups want more recurring access to the client's decisions. A service contract can be renewed each year. A data process, a commerce system or an embedded workflow is harder to replace. That can create real value and switching cost at the same time. The right buyer question is whether the benefit of staying justifies the cost of leaving.analysis V3-116

Prior gap. Publicis set out to buy digital and direct-marketing skill at scale. Then it bought consulting and engineering. Then identity and customer data. Then commerce. Client coordination was a stated goal long before generative AI. The December 2015 plan put Chief Client Officers across disciplines and countries, effective January 2016 [39]. Media buying remains a named service. The sequence reads as a bid to own more of the links between a goal, customer data, content, execution and feedback. That is a reading of the deal order. It is not a claim that the links work as one system.[39]inference A007

Asset. The Digitas agreement of 20 December 2006 proposed a $1.3 billion deal. It named digital expertise and distribution as motives. The date is an announcement, not a closing [1]. Sapient closed on 6 February 2015; the closing release put the price at about $3.7 billion [2]. Epsilon's closing was announced on 2 July 2019, with the business consolidated from 1 July [3]. The half-year 2026 financial report, in note 3.2, dates the full purchase of Lotame to April 2025 [38]. Lotame and LiveRamp are different deals. The first is closed on the company's note. The second is not.[1][2][3][38]fact A008 A009 A020

Intended link. The July 2019 Epsilon release announced completion and consolidation [3]. The idea is simple. One customer identity lets work from several teams reuse the same view of the customer. That could help from one campaign to the next. It also raises the questions this paper keeps asking. Who can license the identity? Who can export matched audiences? Who can check where the data came from? Who can change the measurement vendor? Who keeps the learning after exit?[3]attributed

Use on record. In June 2023 Publicis described retail media products in which bids are triggered by stock and price signals, with a content product that recommends page changes [19]. That is a named input, decision and action. It is stronger than a logo chart. It is still a seller's description of an available skill, not an outside result. On the five gates it sits at the second: documented implementation.[19]attributed

Client effect. Not shown in any source reviewed. No matched client comparison with full cost was found.analysis A048

Group economics. The 2025 Universal Registration Document shows €14.547 billion of net revenue and €17.399 billion of revenue. Net revenue rose from €13.965 billion in 2024; that is reported growth of 4.2%, recomputed from the two figures [43]. The half-year 2026 report shows net revenue of €7.229 billion and an operating-margin subtotal of €1.261 billion. That is 17.4% after €7 million of LiveRamp transaction costs and 17.5% before them; the company headline uses the 17.5% [38][46]. The subtotal is a company measure, not statutory operating income.[43][38][46]fact A017 A018 A019 A020 A021

Countercase. The 2016 results booked a €1.44 billion non-cash charge. Of that, €1.392 billion fell on Publicis.Sapient. The release spoke of legacy digital assets and delays in joining them up [37]. The charge is not a write-off of the Sapient purchase alone; the unit held more than that deal.[37]fact A010 A011 A012

Rival reading. The 2016 charge is evidence against an easy platform story. Buying more can raise the value of joining the parts while making the joining harder. Results consistent with a strategy do not isolate its cause. Simpler cross-selling to the same clients through one client officer fits the same record. So does retention: a shared identity asset raises the cost of leaving whether or not it raises the client's return.analysis

LiveRamp: a pending link, by date

The LiveRamp record is five events, and the paper keeps them apart. Publicis announced the agreement on 17 May 2026, a Sunday. The release stated an equity value of $2.546 billion, net cash of $379 million and an enterprise value of $2.167 billion. It said closing was subject to conditions [17]. The announcement also states an offer price of $38.50 per share; the external review of 6 October 2026 confirmed the price, the equity value and the enterprise value in that announcement, which this edition did not re-open. A LiveRamp shareholder vote on 17 August 2026 appears in the search record of LiveRamp's filings. On 16 September 2026 a Publicis bond release still said the purchase was subject to the usual approvals and expected to close before the end of 2026 [18]. That is the latest dated issuer status checked. No completion was found by the cutoff of 5 October 2026.[17][18]fact A013 A014 V3-009

The proposed asset is not like an agency. A data-collaboration platform joins data held by many firms. Its value may depend on rivals of its owner still using it. Publicis has promised neutral, open access [17]. That is a promise to test. It is not outside proof of neutrality, and it is not evidence of misuse. The test is this: equal terms of access, walls around sensitive data, service quality, and whether rivals keep connecting. A scarce shared utility can lose value if its owner is not trusted to be neutral. Nothing in this paper draws an owned link.[17]analysis A015 A016

Prior gap. WPP's gaps came in this order: search and ad technology, audience buying at scale, research scale, and then one data and workflow layer shared across its agencies. Each gap was met by a different kind of move. A purchase. A launch built from inside. A staged sale. A rename.analysis

Asset. Xaxis launched in June 2011. It joined the group's own data tools with GroupM's trading scale; it was built from existing teams, not bought [42]. WPP announced the InfoSum purchase on 3 April 2025 as an investment in its data offer. The release did not disclose the terms [26].[42][26]fact

Intended link. On 28 May 2025 WPP Media replaced the GroupM name. Mindshare, Wavemaker and EssenceMediacom stayed as client-facing brands [40]. That is an operating change. It does not show one contract, one team or one data right for every client. It is the May 2025 launch structure, not a current legal entity map.[40]fact A022 A023

Use on record. The InfoSum purchase and the WPP Media launch are the recorded steps this review holds. The original edition described more: a three-layer stack, counts of staff using the platform, a production site, a self-serve route and an agent hub. Each is a sign of use reported by the company, and none is a client result.analysis

Client effect. Not shown in any source reviewed.analysis A048

Group economics. The 2026 interim results, published on 6 August 2026, show revenue of £6,373 million and revenue less pass-through costs of £4,745 million. The net measure fell 4.7% like for like, that is, on a constant scope and currency as the company defines them. Headline operating profit was £398 million, which is 8.4% of the net measure. Reported operating profit was £261 million, which is 4.1% of revenue [47]. Those two percentages differ in both numerator and denominator. They cannot be read as one profit test.[47]fact

Countercase. The Kantar sale is the recorded counterexample to endless asset building. On 5 December 2019 WPP said the sale of 60% of Kantar had closed for about 90% of the business. The remaining steps were expected in 2020, and WPP kept 40% [41]. The 40% is the December 2019 position, not a checked current stake. The original edition also recorded a 2020 combination of AKQA and Grey that was separated in 2025, with £641 million of goodwill impairments in the 2025 annual report. If that record holds, it is the largest reversal of an integration for any group in this paper, and it is the first WPP item in section 31.[41]fact A024

Rival reading. A new name and a shared platform had not yet produced group-wide growth by the cutoff. A cost reset after weak results fits the same record as well as a strategy of selective control does. WPP attributed WPP Media's H1 2026 decline partly to prior-year client losses [47]; the external review of 6 October 2026 confirmed that attribution in the interim release, which this edition did not re-open. It is management's explanation of its own division, not an independently established cause. Shared tools can lower WPP's own cost while leaving client cost, quality and control unchanged. Both readings can be true at once. A new structure must earn its keep through delivery, not through less overhead.[47]analysis V3-117

What a buyer can test now. Ask to see one brief pass through the real teams. Which steps are shared? Which still need manual repair? Who may reject a group product? Can the client keep an outside proof team? The claim weakens if lower group costs leave client cost, quality and control unchanged.analysis

Prior gap. Omnicom filled its gaps in this order: shared analytics, then consulting access to client systems, then commerce data and operations, then scale. IPG began with first-party data work, the data a company collects from its own customers. It moved to applications and activation, and then into Omnicom. The two histories stay apart until 26 November 2025. After that date IPG is a lineage inside Omnicom. It is not a peer to add to a total.analysis

Asset. Omnicom closed Flywheel on 2 January 2024. The closing release proposed a link between commerce data and Omni, the group's shared platform [4]. Omnicom completed its purchase of Interpublic on 26 November 2025 [5]. Both are closing records, not plans. One boundary from that record is worth a caution. Acxiom and LiveRamp are different companies on different paths. Nothing in the audited records joins them.[4][5]fact A025

Intended link. The Flywheel closing release stated the design: link commerce data with Omni [4]. The original edition reported that Omni launched in 2018 around partners the group did not own and that a January 2026 release named Acxiom and Flywheel inside the platform. That shift, from partner-neutral to owned, is the chapter's central point, and it is kept here as analysis of the original edition's account.[4]analysis

Use on record. On 28 August 2026 Omnicom launched Hearts United by combining Hearts & Science with Mediahub. The release listed UM, PHD, OMD and Initiative as siblings. UM was not the merger partner [36]. This is a joining of two brands, not a purchase. In July 2026 an Acxiom release in Omnicom's newsroom quoted a Guitar Center executive: recognised website visitors rose from 23% to 35% after an Acxiom link into its Salesforce stack [7]. That is 12 percentage points, or roughly half again in relative terms. It is a vendor-published before-and-after with no disclosed control group. It supports a named use. It does not show added sales caused by the wider group.[36][7]fact A026 A027 A028 A029

Client effect. One reported recognition rate, and nothing on added sales or profit.analysis A029

Group economics. The second-quarter 2026 release, dated 28 July 2026, reports Core Operations revenue of $5,995 million and core adjusted EBITA of $1,068.8 million. EBITA is earnings before interest, tax and amortisation. The core base leaves out sold and held-for-sale businesses. Its 2025 comparator combines Omnicom and IPG [45]. It is a chosen perimeter, not a statutory series. The revised edition printed a bridge from total revenue to core revenue. In it, the excluded amount was just the gap between the other two numbers. This edition does not repeat it.[45]fact

Countercase. In 2019 Omnicom's media group moved specialist staff out of a central desk and into OMD, PHD and Hearts & Science [35]. That is integration running the other way. The original edition held more countercases: IPG's sale of Huge and R/GA before the close, a 2024 goodwill impairment on IPG's digital specialists, a merger with Publicis abandoned in 2014, and US regulatory conditions on the merger. They would make the chapter's countercase larger, not smaller.[35]analysis A006

Rival reading. Once a skill becomes common work, placing it near the client can make more sense than selling access to a distant team. A margin on a chosen perimeter cannot say how much came from breadth, how much from scale and how much from cost removal. Cutting duplicate cost and gaining bargaining power can lift a margin with no client gain at all. Selling weak units makes the survivors look better while the sold units take the losses. A bigger group makes a shared platform easier to sell to the same clients.analysis

What a buyer can test now. Test breadth at the seams. Can commerce stock data change a media plan? Can a creative team challenge the targeting logic? Can a proof team disagree with a sales team? The group thesis gains strength only when those tests work at full cost across repeated client scopes.analysis

Prior gap. Dentsu's gaps were reach outside Japan, digital skill in those markets, and then customer data and CRM. Aegis answered the first two. Merkle answered the third. Both are closed records. The question the record leaves open is whether the group could run those assets at a global cost base.analysis

Asset. The Aegis agreement was announced in July 2012 and the deal completed on 26 March 2013, when Dentsu Aegis Network was established [29]. On 15 April 2020 Dentsu confirmed 100% ownership of Merkle, with payment for the remaining stake still due in the third quarter of 2021 [30]. Ownership, payment and reporting dates differ, which is why the entity map needs separate fields for each.[29][30]fact

Intended link. The case was to link demand with customer experience. Aegis brought media reach. Merkle brought customer systems and services. The original edition added more: a 2016 Merkle rationale, a 2023 one-management-team structure and a 2026 audience product.analysis

Use on record. No Tier 1 or Tier 2 record in this review documents a named link between Merkle's work and the group's media work. That is a gap in the review, not a finding of absence.analysis

Client effect. Not shown in any source reviewed.analysis A048

Group economics and countercase. The FY2025 financial report records goodwill impairment of ¥396,074 million. Of that, ¥299,656 million relates to the Americas and ¥96,418 million to EMEA; the two regional figures sum exactly to the total [31]. The charge is revised value against book amounts, not that year's cash. From regional totals it cannot be assigned wholly to Merkle or to any single deal.[31]fact A030 A031 A050

The original edition leaned on a regional reading: Japan grew in the first half of 2026 while the Americas, EMEA and APAC outside Japan fell. It also reported a February 2026 decision not to sell the international business and a change of chief executive in March 2026. If those records hold, they show a group choosing to run the acquired assets rather than sell them, which sharpens the cost-base question without settling it.analysis

Rival reading. The charge is real contrary evidence to a simple deal-success story. It is not proof that all the acquired skills failed. Strong acquired skills run at an unaffordable global cost base fit the record as well as a failed strategy does. Japan's model may fund a focused recovery. Both can be true in different markets.analysis

What a buyer can test now. Test the proposed local team, its continuity and its full cost. Ask which team owns the brief, the customer data model, the creative work, the buying accounts and the proof of impact in the market concerned. The group's size is a weak substitute for that work.analysis

Prior gap. Havas wanted depth in customer experience and digital work, a health specialism, creative work that stands out, and scale without owning every layer of technology. Its route mixes a few purchases with partnerships. The test is whether a mixed system can deliver better work with clear client rights. The client must also know which firm stands behind each part.analysis

Asset and structure. Havas listed on Euronext Amsterdam on 16 December 2024, after its spin-off from Vivendi [32]. A separate listing does not by itself show who controls the company. The original edition reported family-linked holdings as of September 2026, and purchase records for FullSIX, Havas CX, Uncommon, Channel Bakers and the 2026 deals. Those records are carried forward unread; section 31 lists them with their effect on this chapter.[32]fact

Intended link. On 29 September 2025 Havas and Horizon Media announced Horizon Global, a commercial joint venture [33]. Horizon's own December 2025 release describes it as a 50/50 partnership with Havas Media Network [34]. The parents did not merge. The combined billings quoted in coverage are not Havas revenue.[33][34]fact A032 H018

Use on record. No Tier 1 or Tier 2 record in this review documents a named client use of the Havas operating system or of the joint venture. The original edition named released tools.analysis

Client effect. Not shown in any source reviewed.analysis A048

Group economics. For 2025 Havas reported net revenue of €2,783 million, adjusted EBIT of €358 million and organic growth of 3.1%; organic growth is growth excluding deals and currency, as the company defines it. The margin is 12.9%, and the ratio recomputes to 12.9% at one decimal [48]. That is a full-year, company-defined adjusted measure. It cannot be ranked against a peer's half-year or statutory margin.[48]fact A033 A034 A035

Countercase. The original edition reported two things from the listing prospectus: a large budget loss at one health client, and a €400 million programme for 2024 to 2027 of which about three-fifths was operating expense. Both matter to a claim of breadth, and both would count against it. No impairment record for Havas was audited.analysis

Rival reading. A joint venture is a route to scale without ownership. It also leaves two firms behind one service. Partnerships can keep the brand dependent on outside systems and leave the joining work with people. Open does not mean low cost or easy exit. A workable shared process need not own all the data or tools below it. But the contract must name the party that fixes a broken handoff.analysis

What a buyer can test now. Test the joint service under load. Name the firm that repairs a failed handoff. Demand a clear route to move the work if one partner leaves. Ask which data can move to another partner and which results can be tested by an outside party.analysis

Prior gap. Stagwell is a test of whether a smaller group can connect specialist firms, software and media access without erasing the agencies that brought it clients. Its gaps were scarce creative talent, research, software products, geographic reach and preferred supply. The group did not begin with one finished system.analysis

Asset. The Stagwell and MDC combination was announced as complete on 2 August 2021. Creative, research and digital firms appear in the roster [6]. The original edition also recorded 72andSunny, Anomaly, the Harris assets, Brand New Galaxy and ADK GLOBAL.[6]fact A036

Intended link. The original edition described three 2026 products. In January, a workflow layer across existing tools. In June, a media operating system with more than twenty agents and human decision gates. In July, a marketplace of preferred supply. Together they are the clearest statement by any group that workflow and preferred access are meant to feed each other. Orchestration means joining tasks, data, tools and rights across one process. The limit this paper places on the whole orchestration thesis is here: a shared workflow can steer buying toward the supply the group prefers. Those three releases are carried forward unread and listed in section 31; the argument stands here without them.analysis

Use on record. The second-quarter 2026 earnings presentation, filed on 30 July 2026, labels $16 million as committed enterprise-software revenue for the first half [27]. That is a commercial signal. The slide alone does not show recognition in the period. It does not separate subscriptions from services. It does not show that agent systems are the group's main profit source.[27]attributed A037

Client effect. The original edition held one cost claim from a named client; the audited record holds no client result.analysis

Group economics. The same presentation shows first-half net revenue of $1,216.2 million against $1,162.3 million a year earlier. The organic bridge is $38.9 million, which is 3.3% of the prior-year base [27]. The company headline rounds that to 3%; the original edition printed it with an extra decimal it never had. Stagwell changed its organic-growth method from the third quarter of 2025. Acquired businesses now enter organic growth after twelve months, and the deal adjustment uses current-period acquired revenue rather than prior-period revenue [28]. The 3.3% is on the new method. It must not be spliced onto any earlier series without a constant-method bridge. No such bridge exists in the audited records.[27][28]fact A038 A039 A040 A072

Rival reading. A set of specialist firms sharing tools is a services business with a software label. Software sales may be a way to sell services. A marketplace of preferred supply may make supply economics stronger, not replace them. The gap between gross and net revenue is not a disclosed principal margin.analysis

The smaller-group test: S4 Capital

S4 Capital's audited 2025 results, presented in March 2026, show net revenue of £673.0 million, down 10.8% as reported and 8.4% like for like on the company's definition. Operational EBITDA (earnings before interest, tax, depreciation and amortisation) was £81.2 million and year-end net debt £86.9 million [44]. The decline does not isolate an AI effect. The revised edition said the margin rose and net debt fell. Those directions rest on prior-year values not in the audited passage, so this edition states the 2025 figures only.[44]fact A041 A042 A043 A044

What a buyer can test now. S4 is the counterweight to any claim that one digital brand protects demand. A digital model can improve cost and cash while demand weakens. Consolidation and technology coexist with client caution. So judge the actual team and the joins, not the size of the parent. Ask whether the default supply choices inside a buying system improve the client's result, and whether an outside party can challenge them. Keep three things apart: delivery the group owns, work done with affiliates, and access to software. A smaller group can offer all three under one name.analysis

The rival to a large group is not one kind of firm. The set includes consulting firms, listed digital networks, groups backed by private equity, founder-led firms, teams embedded in the client, and the media platforms themselves. Each enters the client through a different need: media buying, enterprise systems, creative production, commerce or a skill the brand wants inside. Each can then grow into the next task. This paper compares where each one gets its grip on the client, not their revenues. A consulting firm's total revenue, an agency's fee revenue, media billings and software commitments do not measure the same thing. This review holds no audited revenue for any consulting firm's marketing arm.analysis

Not all independents are open, and not all groups are closed. Independence from a listed group can mean private-equity ownership with its own return targets, founder control with minority investors, or an agency that sells embedded teams and owned software. None of those makes advice more impartial or exit easier by itself. The test is the same as for a group: rights the client can exercise, a proof route the seller does not control, and a working export.analysis

The in-house boundary

In-house work moves the boundary. It does not end the need for an agency in any sample held here. The ANA's 2023 member survey had 162 respondents, with fieldwork in February and March 2023. It reported 82% with an in-house agency and 92% still using outside agencies [9]. The WFA and Observatory International study of 45 companies, published on 20 December 2023, reported 66% with an in-house agency and all respondents still working with outside agencies [10]. These are incidence rates in member samples, not shares of spend. The two populations must not be pooled.[9][10]attributed A064 A065 A066 A067 A068

The results support coexistence. They do not show that agency work or spend stayed level. A brand can bring one task inside and buy more of another outside. A small internal team can own a key choice while a partner does most of the work. Dependence falls only when the brand keeps the brief, the account ownership, the test plan and the budget authority, and can run an exit. Fixed cost rises with headcount whoever owns the decision.analysis

Compare tasks, not labels. Who owns the brief, the account, the test plan and the final budget? Which skills are needed every week, and which twice a year? Count the full cost of staff, tools, training, cover, rework and the links between teams. Map dependence by task and by exit time, the time it would take to move each task to another supplier. Do not map it by the presence of an internal team.analysis

Horizon, as an independent

Horizon Media is the independent agency this paper follows. Its releases describe a way of working, not a deal. Horizon Media Holdings is a private parent [34]. On 18 December 2025 it launched HorizonOS. The release described human specialists, client pilots and a partner marketplace. It also disclosed the 50/50 Horizon Global partnership with Havas Media Network [34]. On 18 June 2026 it announced an agentic layer for media decisions, with APIs, MCP connections and agentic buying. An API is a software interface; MCP is a standard for connecting AI tools; 'agentic' means software agents act within the system. The one named client, SharkNinja, was using audience APIs with retail partners. That is audience collaboration, not autonomous buying [13]. On 1 October 2026 it named a third cohort of 11 partners, chosen from more than 600 submissions. More than 50 partners had been through the programme. The next step was to scope pilots with clients [14].[34][13][14]attributed H001 H002 H003 H004 H005 H006 H008 H009 H010 H011 H017 H018

Three limits apply, and section 22 works through them. A marketplace is not portability. The promise of an ecosystem with no vendor lock-in is a sales claim until a contract and a migration drill back it. A count of partners is not a count of deployments. An announcement is not proven autonomy, and the SharkNinja example does not prove autonomous buying. And no fee mix, principal margin or client economics for Horizon appears anywhere in this paper, because no release reviewed states them. Horizon is a proposed way of working. It sits closest to the agency-led path. It is not proof that independents have won.analysis H012 H020 H006

Platform substitutes

Platforms automate the work inside their own inventory. Google's help page describes Performance Max as a campaign type that spans Google Ads inventory and bids toward goals the advertiser sets. The page is a living document with no version history [15]. Amazon's page dated 21 March 2025 had a title that put Brand+ in open beta. Its body, when the audit read it, said both Performance+ and Brand+ were available globally. By the cutoff the title in search results had changed again. The page cannot date the later stage on its own, and this paper does not claim the exact stage from it [16]. On the five gates, both sit at the second, documented implementation, as described by the seller.[15][16]attributed A057

The substitution is real and bounded. A platform removes manual buying work inside its own silo. It gives no outside view across platforms. It keeps its own model state, the learned settings behind its bidding, under its own terms. An agency can compete with that automation. It can also run it for the client. So platform use is an input to an agency as well as a rival to it. Neither a platform's report nor an agency's claim settles whether the sales were new, moved from another channel, or bought with discounts.analysis A052

For the CMO the evidence supports a choice of operating model, not a choice between old and new logos. Use a group where shared talent, local access and cross-discipline work solve a real coordination problem. Use a consultancy where implementation and enterprise change dominate. Use embedded teams where closeness to the brief and repeat production matter. Use software where the brand can own the process and govern the outputs. These are decision rules, not rankings. Part III turns them into three paths.analysis

Numbers in this field mislead in four ways. Gross revenue is compared with net revenue. A quarter is compared with a half-year. An organic rate on one definition is compared with a like-for-like rate on another. And a rounded headline gains a decimal it never had. This section prints each observation on its own row with its own definition. It says which can be recomputed from disclosed figures and which can only be quoted.analysis

Growth observations, one row each

Five rows from the original peer chart are missing on purpose. They are Publicis first-half 2026 organic growth, Omnicom second-quarter 2026 core organic growth, Dentsu first-half 2026 organic growth, Havas first-half 2026 organic growth and S4 Capital first-half 2026 like-for-like change. Each rests on a document this update did not re-open; section 31 lists them with the primary document that would settle each. Even when they return, they must not share an axis. A quarter is not a half. A chosen core perimeter is not a statutory group. A changed organic method is not an unchanged one.analysis A049

Margins, one decimal, with their definitions

No two of these six ratios share both the top line and the bottom line. The inputs are rounded to the nearest million. A fourth decimal would be false precision. One decimal is all the inputs allow. No figure in this paper is converted to another currency or scaled to a full year.analysis A021

LiveRamp: which number is which

Press reports of the proposed LiveRamp deal quote $2.5 billion and about $2.2 billion for the same transaction. Both are right. They measure different things. The Publicis release of 17 May 2026 states an equity value of $2.546 billion, net cash of $379 million and an enterprise value of $2.167 billion [17]. Equity value is what shareholders receive: the per-share price times the fully diluted share count. Enterprise value is what the buyer pays net of the cash it acquires: $2.546 billion less $0.379 billion equals $2.167 billion, which rounds to $2.2 billion. Publicis’s 17 May 2026 announcement states an offer price of $38.50 per share, an equity value of $2.546 billion and an enterprise value of $2.167 billion; completion remained conditional in its 16 September update [18]. The price was confirmed in the announcement by an external review on 6 October 2026; this edition did not re-open the source, and its stored passage carries the two values but not the price. The net-cash figure is as of a date stated in the release and will differ at closing. Neither number is a completed purchase price. The two must never be added or swapped.[17][18]fact A013 A014 V3-009

Survey denominators

The ANA reported in March 2026 that 58% of respondents had used principal media in the past year, against 47% in its prior wave. The sample was 114 qualified client-side members, with fieldwork from October 2025 to January 2026 [8]. The gap is 11 percentage points between two member samples that were not matched. The prior wave's sample size and the base for the 58% are not in the audited passage. No weighting is stated. The figure is a rate of use, not a share of spend. A relative change should not be computed from it. The 2023 in-house figures are rates of use too: 82% and 92% of 162 ANA respondents, and 66% of 45 companies in the WFA and Observatory study. They must not be pooled [9][10]. The WFA's 2026 study covered 69 multinationals, 71% of whose respondents work in global marketing procurement. It reported that 63% expected more performance-based fees, and that such fees are still a small share in most disciplines [11].[8][9][10][11]attributed A061 A062 A063 A064 A065 A066 A067 A068 A069 A070 A071

What growth cannot say

Growth can come from price, wins, scope, currency or demand. Profit can rise through staff cuts. Cash can rise through client payment timing. None of these, on its own, shows that AI or one deal caused a better client result. The records in this section do not isolate a causal effect of AI or of any one deal. A strong test needs the same scope, quality and cost boundary on both sides. No such test has been published for any group in this review.analysis A048

Value created for a client and value captured by a seller are different quantities. Value is captured in fees, spreads, subscriptions, data services, switching costs or platform spend. A group can gain margin without a client gaining profit. A client can gain from a closed system. An open system can cost more to run. This section does not rank business models by virtue. It names each mechanism, who gains from it as volume rises, whether the client can gain too, and what evidence would show that the client did.analysis

Three readings follow from the table. First, spreads, subscriptions, data services and switching costs may allow a seller's revenue to grow faster than its delivery labour; in this review's register they are also the mechanisms with no client-side test on record. This review does not compare their scalability or client returns. That is not an accusation. It is a measure of where the proof is missing. Second, the one mechanism with task-level evidence, labour saving, comes from customer support, not advertising. The study's authors note that the gain varied with skill [12]. It cannot be applied to agency cost or profit. Third, every row has a client-side test that a contract can require. None has been published for a named group account.[12]analysis A060

The reading rule for the rest of the paper is this. For each mechanism, state who funds spend before proof arrives, who funds failed tests and who captures savings. A spending cap is not a liability ceiling. Spend awaiting proof is not a loss. A group that gains margin has not thereby cost a client anything. A client that gains from a closed system has not thereby lost control. The question for each mechanism is whether the brand can see it, test it, stop it and leave it. Decision rights explain who captures value; the quality of the work, which section 19 takes up, is what creates it. Part II puts eleven claims about these mechanisms under pressure. Part III turns the tests into three paths.analysis

Part I set out the record. This part tests the claims people draw from it. The eleven below are the ones a CMO is likely to hear in pitches, trade press and board papers; claim (b) is graded in three parts, because a revenue model, a delivery model and client value are three different claims. For each, the table gives the strongest support in this edition's register, the strongest contrary evidence in it, a rival explanation, the condition under which the claim would hold, and the test that could change this paper's reading. Every row uses only the sources in this edition's register. A blank is a gap in what this review inspected, not a finding of absence.analysis

Where the support is thinnest

Claim (b) is graded in three parts. The revenue-model claim (b1) has the thinnest support in this register, because its evidence is a category error. Launches, colleague counts and committed software revenue show intent and skill. A software revenue model shows up in audited accounts as a recurring fee line with retention. No group in this edition's register shows that line. Publicis' FY2025 accounts split revenue from net revenue by pass-through cost, not by software [43]. Stagwell's $16 million is committed and mixed [27]. This paper treats (b1) as unestablished for every group. The delivery claim (b2) is different in kind. An agency can run an orchestration service on retainers, project fees or managed-service fees, so a service-fee model can pass the delivery test without passing the revenue test. The register holds described designs, not a logged cross-service workflow on a named account, so (b2) is described but unshown. The client-value claim (b3) is unshown for every group. A subscription product does not pass it by having recurring revenue, and a managed service does not pass it by having a wide scope. The rival readings, a cost reset and cross-selling, fit the same facts at least as well.[43][27]inference A037 A017 A018

Claim (f) is weak because the documents that exist are the wrong kind. A draft for public comment [20], an SDK description [21] and a repository at a pinned commit [22] are proposals and specifications. An SDK is a software kit for building on a protocol. Compatibility is a property of two named implementations at a named version after a test. Portability is a property of state, rights and licences. Neutrality is a property of incentives. The AdCP FAQ itself says shared transport does not settle shared meaning [23]. This review ran no compatibility test. It found no public cross-vendor result. Until one exists, (f) is a hope with a schema attached.[20][21][22][23]inference A053 A054 A055 A056

Claims (g), (h) and (j) fail together, and for one reason. Each assumes a right the brand has not been shown to hold. Autonomy improves outcomes where a fast signal is a true signal; the 2019 experiment comparison says it often is not [51]. A seller's assessment of its own media can be trusted where someone outside the sales line is free to publish an adverse result; no reviewed contract shows that freedom. Access to data becomes usable learning when the export, licence and transfer rights the task needs are in the contract. The one case in this edition that promises no lock-in has not been verified [14]. The common fix is in Parts II and III: a tested stop, a separate proof function (the team that tests results and reports to the brand, not to the spender) and an exit drill (a rehearsed export and rebuild with another supplier) before scale.[51][14]inference A051 H012 H005

'Agentic' is sold as one thing. It is six different grants. A tool that drafts a plan and a tool that moves budget across platforms without fresh approval carry different risks. They need different logs and fail in different ways. The ladder below names the grant at each level. The same product can sit at level 1 for copy, level 3 for pacing (keeping spend on schedule) and level 5 for a hand-off between two systems. The level belongs to the task, not to the vendor. The ladder is a teaching aid. It orders grants by the discretion they give an agent, and it should not be read as saying that more coordination means more autonomy; the task profile below records the dimensions the ladder mixes as separate facts.analysis