No Fluff Advisory · Research · Holding companies Full paper

Plain-language edition

The Evolution of Holding Companies, in plain language

Who runs the work, who proves it worked, and who keeps what was learned.

By Evgeny Popov · No Fluff Advisory · October 2026

12 minto read, about 2700 words
87.6Flesch reading ease, measured on this page with textstat 0.7.13 (the full paper scores lower)
210atomic claims checked in the full paper

Current status · 7 October 2026

The findings and sources are unchanged since the update of 6 October 2026. Since then the 20 figures were rebuilt and checked twice against the live page, a 21st figure in section 20 sets the present-day loop beside a possible future, and the retailer example in section 26 now pauses new spending after a failed evidence gate. No source was re-opened; the research cutoff is still 5 October 2026. Every dated change is listed in the update history after the method section. Update history

Read this first

This is the short, plain version of a long research paper. The full paper is called The Evolution of Holding Companies: Control, client value and three paths for marketing work. It has the sources, the tables and the math. This page has the same findings in fewer and shorter words. The numbers are copied from the paper. Each part links to the place in the paper that backs it up. This page matches the paper's findings as updated on 6 October 2026, after an outside review. The paper was last updated on 7 October 2026.

The research runs from January 2006 to 5 October 2026. It looks at six big ad groups: Publicis, WPP, Omnicom with IPG, Dentsu, Havas and Stagwell. It also looks at some smaller firms and at the big ad platforms. Section 2 sets out the scope.

What the paper asks

Big ad groups used to buy media for brands. Over twenty years they bought much more. They now want to help set a brand's goal, run the work, hold the data and judge the result. The groups are adding data, software and coordination to what they already sell. Those layers can help the brand, prop up the group's own media business, or make the group harder to leave. Which one depends on the task, after all the costs are counted. So the paper asks one thing. When a group does all of that, what does the brand still control, and what does it still have to prove? Section 1 has the argument.

What a holding company is

A holding company is a parent that owns many marketing firms. The paper just calls it a group. Think of a family that owns ten shops on one street. Each shop has its own name and its own staff. But one family sets the rent, the rules and the prices.

Now think of the work between a brand's goal and its next move as a loop. Set the goal. Make a plan. Buy the ads. Measure. Give credit. Report. Adjust. Plan again. The groups spent twenty years buying the parts of that loop. Section 1 names them.

The dates are real. Publicis agreed to buy Digitas in December 2006. It closed Sapient in February 2015 and Epsilon in July 2019, and it bought Lotame in April 2025. Dentsu closed Aegis in March 2013 and took full control of Merkle in April 2020. Stagwell and MDC joined on 2 August 2021. Omnicom closed Flywheel on 2 January 2024 and IPG on 26 November 2025. Havas listed its shares in December 2024 and announced a joint venture with Horizon Media in September 2025. WPP built Xaxis in 2011, sold 60% of Kantar in 2019, and gave GroupM the new name WPP Media in May 2025. Section 4 has the full dated table.

Each group gets its own chapter: Publicis, WPP, Omnicom and IPG, Dentsu, Havas and Stagwell. Each chapter asks the same things. What gap did the deal fill? What was bought? How was it meant to link up? Did any client gain? What did the group earn?

What we found

1. The record proves the bid, not the client's gain

A receipt shows what you paid. It does not show that the meal was good. The deal records are like that. They show what the groups bought, what they wrote off and what margin they earned. They almost never show a matched client result: profit after fees, inside costs and the long-term effect on the brand. The one client result in the whole review is the share of one retailer's website visitors that a data firm could recognise. It rose from 23% to 35%. It is not a sales figure, and it comes from a vendor release with no control group. Section 1 and section 14 keep that gap open on purpose.

2. Buying scale has not gone away

Some agencies buy ad space for themselves and sell it on to clients. The trade calls this principal media. In March 2026 the ANA, a body whose members are advertisers, said that 58% of the 114 client-side members in its survey had used principal media in the past year. In its prior survey the figure was 47%. That is a count of members, not a share of spend. But it is the only like-for-like pair of readings the paper holds, and it points up. Section 13 explains the survey numbers.

3. The record holds real setbacks

Not every deal paid off. In 2016 Publicis took a charge of €1.44 billion, and its release cited delays in joining up its digital assets. Dentsu wrote down ¥396.1 billion of goodwill for its 2025 financial year. WPP sold 60% of Kantar, its research arm, in December 2019 and kept 40%. And S4 Capital, a digital-first challenger, saw its 2025 net revenue fall 10.8% as reported. None proves a strategy failed. Each says the simple success story is not the whole story. Sections 6 to 11 hold the cases.

4. The biggest new link is still pending

Publicis announced a deal to buy LiveRamp on 17 May 2026. LiveRamp is a data platform that lets many firms match their data with each other. You will see two prices in the news: $2.5 billion and $2.2 billion. Both are right. The first is what shareholders get, $2.546 billion. The second takes off LiveRamp's cash and comes to $2.167 billion. The offer was $38.50 a share. The paper did not open that announcement itself; an outside review on 6 October 2026 did, and found the price there, so the paper says who checked it. On 16 September 2026 a Publicis bond release still said the deal needed approvals and was expected to close before the end of 2026. The paper found no record that it closed by 5 October 2026. So the paper never calls LiveRamp owned. Section 6 and section 13 walk through it.

5. Value made and value kept are different things

A group can earn more without a client earning more. Spreads on media, subscriptions, data fees and the cost of leaving can let a seller earn more without adding staff. In the paper's sources they are also the ways with no proof of client gain on record. The paper does not rank them. The paper shows this with one small sum. Say a task costs an agency 60 in staff and 20 in other costs, and the client pays 100. New tools cut the staff cost by a quarter and add 5 in tool and review costs. The cost falls to 70. If the fee stays at 100, the agency keeps all the saving. At a fee of 90, the client gets all of it. The sum is made up. It proves nothing about any real agency. It shows that a saving has to go somewhere, and the contract decides where. Section 18 has the model and section 14 lists who gains from each route.

What "control" means

Control is the key word in the paper, and 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?

Think of a house you rent. It matters little who built it. It matters who holds the keys, who can change the locks, and what you can take with you when you leave. No system grants those rights by name. A contract, a log and a tested exit do. Section 1 sets this out. Section 20 draws a loose, typical-looking loop and shows what it may lack: an owner for the record of decisions, a proof team apart from the team that spends, a stated level of authority per task, a stop that has been tested, and an exit with a price. Many teams already have some of these. The paper does not claim to know how many.

Money also leaves before proof comes back. Fast signals arrive in hours; real proof takes weeks. One made-up case in the paper: $10,000 a day for seven days is $70,000 out the door before the first real read. That is not a loss, but someone carries the gap, and the deal should say who. Section 18 covers it.

The paper adds two tools. One is a ladder of six levels of authority, from "suggest a plan" to "move real budget with no fresh approval". A tool can sit at a low level for one task and a high level for another. Section 16 has the ladder. The other is a list of thirteen controls a buyer should test, such as whether a stop reaches every booked order. The paper proposes the tests. It did not run them. Section 17 has the list.

Three paths, and the baseline

The paper first draws a loose, typical-looking loop, not a measured one: eight tasks in seven stages, with a fast clock for daily tweaks and a slow clock for real proof. Attribution, the step that hands out credit, is not proof that the ads caused growth. Section 20 is the baseline.

The paper calls that loop PAM: plan, activate, measure. It is one wheel that turns once per campaign. A figure in section 20 sets it beside a possible future. The work stays the same, but it runs as three loops at three speeds. A fast loop acts within its limits in hours, and it never counts as proof. A middle loop runs over weeks: it tests what worked and decides what agents may do next. A slow loop revisits the goal over months. Only the two slower loops may change the rules, and only with an approval on the record. A stop can halt all three at any time. In the figure, dots run around each ring at their own speed. See the loops in section 20.

Then it sets out three paths. These are "could happen" designs, not forecasts. The paper gives no odds for any of them. A brand can use all three at once for different tasks.

Path A: the brand runs the core. The brand owns the goals, the permissions and the record of choices and tests. Agencies supply strategy, craft and skilled work. It can be costly to run, because the brand pays for the joins and the drills. The paper has no cost cases to rank the paths, so it compares them on one list of nine cost lines instead. Section 21 sets it out.

Path B: an agency runs the system. The brand keeps the goals, the stop rights, the proof and the big budget calls. Horizon Media's public design is the closest case, as a proposal, not a proven result. Section 22 sets it out.

Path C: the platforms run most of the buying inside their own walls. Local results can be good. But the view across platforms stays patchy, and the learned state stays behind. Section 23 sets it out.

Pick by task, not by supplier. The paper recommends that three tasks stay in-house for any brand: set the goal and its limits, hold the measurement mandate and its budget, and own the contract, the audit and the exit. Everything else is a cost-and-risk comparison. That split is the paper's advice, not a rule it found in the data. Section 24 has the table.

What a CMO can do in 90 days

The paper proposes one small pilot and one exit drill. In weeks 1 to 4, pick one decision to test. Add up all the money: media, fees, data, tools and inside staff. Write down where you stand now. Then put the goal, the caps and the stop rule on one signed page. In weeks 5 to 8, run the pilot, with the tool allowed only to act on preset rules or to adapt within set bounds. In week 4, before any money goes out, test one stop, one refused request, one retry and the spending cap. Nothing gets live authority until all four pass. In week 6, rehearse the stop while the pilot runs, and test one export. In weeks 9 to 10, have the proof lead or an outside reviewer challenge the result in writing. In weeks 10 to 12, rebuild a small piece of the work outside the current supplier and count what was lost. In week 13, apply the rule you signed in week 4. A pilot that fails but teaches you something is a success of the method. Section 28 has the full plan.

A small team holds the rights. The paper names the roles and who can stop what. One way a small brand can do it is with three people: one for goal and budget, one for data and proof, one for creative. The doing can be shared, part-time or bought, as long as each right has a named owner. Section 27 has the team. The creative lead matters. Cheaper versions of a weak idea are still a weak idea. Rights protect the work; a clear brief, a shared view of the customer, decisions on time and useful feedback are what make it better. Section 19 keeps both in view.

What the paper could not prove

The paper is open about four gaps. First, no source gives a matched client result for any deal. Second, LiveRamp was still pending at the cutoff. Third, no protocol, tool or group has passed a public test of compatibility, stop, revocation or exit, and the paper ran none. Two rule books for AI buying agents are at an early stage. AdCP had a published release, 3.2.1, dated 30 September 2026, checked from a copy pinned on 5 October 2026. The IAB Tech Lab's OpenProposal was a draft open for comment until 22 October 2026. A published rule book is not a passed test. Section 17 has the stages.

Fourth, no source was re-opened in this review. The network blocked every source site. So each of the 49 sources was checked only by search, to see whether its title, web address and date still showed up. The passages used for facts were carried forward from an audit dated 5 October 2026. That audit checked 72 claims from the original paper plus 20 about Horizon, 92 in all, and re-did 32 sums; the paper's number file holds 104 checks in all, and only those 32 are true recomputations. The review was three passes by the same assistant. That is three processes, not an outside review. An outside review dated 6 October 2026 did open 13 of the sources and checked the sums. The paper names it where it leans on one of its checks, and treats those as the reviewer's checks, not its own. The paper still has to re-open its sources. Section 2 and section 31 state these limits.

Where the full ledgers are

The full paper ends with four tools. A version map shows what changed across the editions, and a dated update history records every change, the latest on 7 October 2026. A claim ledger lists every checked claim with its verdict. A source register lists every source with its stored passage and its status at the cutoff. A corrections ledger lists every fix. The build counts 210 claim records, 51 source records and 255 corrections. The data files can be downloaded from the paper. Section 31 has the ledgers.

Who wrote this

The author, Evgeny Popov, has worked at Samba TV since May 2025. Samba TV measures TV viewing and sells ad data. Before that he worked at Hearts & Science, which Omnicom combined with Mediahub into Hearts United in August 2026, and at Lotame, which Publicis bought in 2025. Both are named in the paper and judged on the public record alone. He co-leads a working group at AgenticAdvertising.org, the body that publishes AdCP. Samba TV is listed among AdCP's members, and No Fluff Advisory, his own practice, is a member of the body. None of that earns AdCP any stage in the paper, and his own work on it is not cited. No company named in the paper funded, reviewed or saw it before publication. The full disclosure is at the end of the paper.

Read the full paper

The full paper has the sources, the data and each step of the math. Read The Evolution of Holding Companies in full.

Update history 6 dated notes
  • Update of 6 October 2026, after an external critical review dated the same day. This update changed: the edition label, which no longer says 'Final'; the thesis in section 1 and summary point 5; the LiveRamp offer price and the WPP client-loss attribution, now printed as statements the external review confirmed in the cited sources and this edition did not re-open; proposition (b) in section 15, split into three graded claims; section 29, now research questions and update triggers; one brand-side cost boundary across sections 20 to 24; the retailer's budget tranches, discount arithmetic, exclusion dates and aggregation rule in section 26; the baseline label and the pre-live control tests in sections 20 and 28; bounded wording in place of rankings and absolutes; a value-production loop in section 19; a task profile beside the authority ladder in section 16; the partnership definitions in section 3, the Digitas status in section 6 and the top evidence gate in section 2; and the claim ledger's wording versions and status field, with 'numerical and comparability checks' in place of 'recomputations'. Every record this edition carries forward unread is now listed once in section 31 with its effect on the conclusion. The re-opening of the 49 inherited sources remains open: no source was opened in this update either, and the external review's source checks are attributed to it, not adopted as this paper's own verification. The calculator reset (review finding F14) and the print and figure typography repairs (F15) were made at build and figure level later on 6 October 2026, after this manuscript edit: the reset now restores the five defaults (100, 60, 20, 25, 5) and recomputes the chart, table and sentence together; the figure text equivalents print once in an appendix and each scenario prints once; the figures were then rebuilt with larger labels, as the next note records. These are build and figure changes, not manuscript changes; the corrections ledger logs them (ER-F14, ER-F15).
  • Figure update, later on 6 October 2026, after the review update above. This is a rendering and interaction pass, not a new independent audit. No source was re-opened. No fact after the 5 October 2026 cutoff was added. No research fact, source or date changed. Every figure now sits in one shared shell: a number, a status badge, a one-line takeaway, controls, an Enlarge button that opens a viewer (fit width, fit all, 100%, zoom, pan, Escape to close), a caption, a text equivalent, a sources-and-assumptions note and an editable SVG download. On phones each figure shows a compact layout in place of a drawing that scrolls sideways. The fourteen existing figures were redrawn from data files that carry claim and source ids. The five control questions (section 1) are five cards, not a cycle. The timeline (section 4) runs on one linear axis with filters, a dated event list and a lens view. How to read a deal (section 5) is a numbered checklist with twelve worked cases. Value capture (section 14) and the seven fee routes (section 18) are one commercial-model explorer with two entry views. The task profile (section 16) shows stated conditions for two tasks at a time. The task-saving figure and the calculator (section 18) share one calculation and four presets; the equal split at a fee of 95 is new and illustrative. The baseline and paths A, B and C (sections 20 to 23) are one comparison component with compact chapter views and a six-question table. The Horizon loops (section 22) open on the client loop, with a step-through. Openness and control (section 24) is a described two-by-two with no vendor count. The 90-day plan (section 28) is four phase cards with visible gates. Six figures were added on 6 October 2026: from a claim to evidence (section 2, with a protocol view linked from section 17); who owns what, since when (section 3, with chapter views in sections 6 to 11); spending to proof (section 18); how judgment creates value (section 19); what leaves with the brand (section 25); one retailer, three agents (section 26). The paper now has 20 figures and 20 SVG downloads. Nine points wait for the author and are marked open in the corrections ledger: the $90,000 cap default and the commitments input in the spending figure; the returns wording and the weekly-budget dollar example in the retailer walkthrough; the row labels of the section 24 table; the two proposed measures in the judgment figure; whether to print the AdCP 3.2.2 patch date; the capability columns and the S4 Capital row in the ownership figure; the proposed roles in the control questions; the wording for records the deal figure leaves blank; and a compact protocol table for section 17. One assistant checked the figures in three passes, in a browser at seven widths, in light and dark themes, and in the PDF. A screen reader and a physical phone were not used.
  • Figure repair, later on 6 October 2026, after the author’s live validation of the figure update (a browser and representation review of the published page by one assistant, received the same day). The validation found the figures clearer but not ready for sign-off, and listed ten findings. All ten are addressed here, and the open points of the note above are settled. The enlarged view now moves the whole figure into the dialog, so every switch inside it works and stays in step with its list and detail panel; the zoom readout is measured from the drawing after every change, so it matches what the reader sees. The spending figure (section 18) now holds spend to the cap: each day spends at most what the cap leaves. A slow stop is a labelled assumption, off by default, and any spend past the cap is drawn as its own line, not hidden. The read schedule is stated: after the first read, each day’s spend is read a set number of days later. The $90,000 cap stays, marked as borrowed from section 26 and not a recommended limit. The retailer figure (section 26) now keeps three steps apart: a proposal can be eligible, then wait for the creative lead, then run only if a tranche has money left. With tranche 1 spent before its read, nothing runs. The returns assumption is stated in the text, the table and the figure, with an alternative offered as a figure option only; the weekly budget behind the 10% shift rule is defined. The ownership figure (sections 3 to 12) now keeps apart what an entity is, how it relates to the group, the deal’s status, and the dates. Digitas reads as owned on the record but not checked again here; Lotame and LiveRamp are businesses, LiveRamp still pending; the $16 million Stagwell line is a reported figure, not a product; Horizon is the independent case with its own private parent; SharkNinja is a named client. Each group has its own download, and the all-groups file is a labelled reading sheet. Exports wrap their text for the fallback font. The deal figure names the worked case it shows, and two labels replace one: not assessed in this review, and not shown in the reviewed record. The baseline and path figures (sections 20 to 23) no longer print a yes-or-no “brand decides” line; each stage names who runs it, whose sign-off it needs and what is delegated. The section 24 table uses the figure’s definition: open means providers can be connected and replaced. The AdCP row keeps 3.2.1 as the version analysed, with the v3.2.2 patch tag of 4 October 2026 recorded beside it, not substituted. Section 17 carries a short protocol table from the same data. The judgment figure keeps its two proposals as questions, not measures. The control questions say a title grants no legal authority. No research fact, source or date changed, and nothing was fetched. Still open: on the two tallest views, fitting the whole drawing to the screen makes labels about 14 pixels (fit to width keeps them at 18 or more); the Horizon single-loop views are scaled copies of the overview; the ownership print takes five half-empty sheets; and one author decision, whether execution pauses when the week 4 read fails while tranche 1 still has money. Verification record: one assistant ran four repair passes and one merge, not independent reviews. On the merged build, the browser check ran every control of every interactive figure both in the page and in the enlarged view (272 transitions, each in both places) and checked the readout against the drawing’s real scale. The author’s headline cases were reproduced and pass. A screen reader, a physical phone, real browser zoom and other browsers were not used. The test record is published with the data downloads as figure-test-record.md.
  • Update of 7 October 2026, after the author’s live check of the figure repair. The check confirmed the repair on the live page and left two items, both now done. First, the author decided what happens when an evidence gate fails. Section 26 now says: at a failed interim gate, pause new commitments and queued actions that can still be stopped; hold the next tranche and return the plan to the CMO and budget owner; unspent budget does not carry execution authority forward; resume only under a logged human decision consistent with the mandate; keep non-cancellable commitments visible. A delayed or unclear read is not a failure. Any continuation must be set in advance, with a time limit and a spend cap. In this example, after a delayed week 4 read, work may go on for up to 7 days from what is left of tranche 1, capped at $22,500; there is no continuation after week 8. Actions already taken stay logged. The retailer figure follows the same rule. Second, the ownership figure was too tall to read whole. Each row now shows the name, type, relationship and status; dates, checks and source passages sit in the panel when a row is selected. The capability mapping is a separate view. The viewer’s buttons now say Fit width (reading) and Fit all (overview), and on screens 900 to 1,199 pixels wide it opens with its side panel hidden. A correction: the note above said that fitting the tallest views to the screen gave labels of about 14 pixels. That was not true of the Horizon ownership view, which gave about 10. After this update, at 1366 by 900 pixels every ownership view keeps labels at about 18 pixels or more, whether fitted to width or whole. No research fact or source changed. One assistant made and checked these changes. A screen reader, a physical phone, real browser zoom and other browsers were not used. The measured sizes are in the test record. Later the same day, the figures were set to one reading width: on any desktop a drawing is at most 880 pixels wide, so its labels read at about 18 pixels beside 17-pixel text, and the Horizon single-loop view is no longer enlarged as far.
  • Editorial cleanup of 7 October 2026, after the author’s final review. The page now opens with a short current-status note; these dated notes moved here, after the method section. The edition label now reads 7 October 2026. Section 31 and claim V3-029 now report the release checks that ran on the final build, with their dates and results, instead of describing them as pending, and they keep the list of what was not tested. Readability was measured again on the final text: 69.4 for the narrative, against a target of 85 that this edition does not meet, and 87.3 for the plain edition. The hardest passages are the enumerations of the nine cost lines and the loop descriptions in parts II and III; they were left as written, because shortening them would drop terms the cost boundary needs. Section 26 now states the delayed-read rule in one sentence, as the author decided: a delayed read is not a failed result, but it does not extend spending authority. There is no continuation after week 8; new execution stays paused until the read arrives or the CMO and budget owner approve a logged exception with its own spending cap, expiry and rationale, and tranche 3 stays held. The retailer figure’s table says the same; the spending-to-proof figure is unchanged. No finding or source changed (FG-61 to FG-63).
  • Figure addition of 7 October 2026, at the author’s request. Section 20 gains one paragraph and one figure, “From PAM to the governed loop”: the present-day loop, here called PAM for plan, activate and measure, beside a possible future of three loops on three clocks around the five rights. It restates the section 20 baseline and the loop rules of section 26; no finding, source, date or claim changed, no source was re-opened and nothing after the 5 October 2026 cutoff was added. It is the paper’s only figure with motion: dots run on its rings at relative speeds, not to scale. The motion stops under a reduced-motion setting, in print and in the download, and a control on the figure pauses it.