I Have a Thesis

Full analysis

Two founders spent an hour arguing that programmatic never delivered on its founding promise: accurate, per-impression valuation. Adam Heimlich (Chalice) says the fix is per-advertiser algorithms instead of pooled ones. Gareth Glaser (Gamera) says the fix is a neutral, on-page signal layer that publishers can't game. Both are pitching their own book, but the diagnosis underneath is worth taking seriously.

What's actually being decided for an operator: whether the "pooled DSP algorithm" and the "trust the bid-stream fields" era is genuinely ending, and whether you should reallocate budget, roadmap, or vendor spend toward custom bidding and independent inventory labeling. This is a Type 2 decision for most: you can run a pilot, measure, and walk away. Cheap to test, so the bar for deliberation is low and the bar for action is also low.

Timeline / forcing function: none hard. The John Wren non-answer and the Meta/Microsoft capex split are the timely hooks, but the theses themselves are slow-burn. No renewal clock forces your hand this quarter.

The Market Analyst Two vendors both backed by The Trade Desk are on a podcast explaining why the incumbent DSP model is broken. Read the incentives and it still holds up: Trade Desk is the one large buy-side platform whose pitch has always been "we're on your side, not the pool's," so custom algorithms and neutral labeling feed its narrative against Google and Amazon. For an operator, the plain-English version: the money is starting to bet that generic, one-size-fits-all bidding loses to advertiser-specific bidding. The Meta-versus-Microsoft point Ian Whitaker made is the tell for the whole space. Microsoft gets paid for compute because it resells it; Meta gets punished because nobody believes more GPUs automatically buy a bigger ad moat. That skepticism is coming for agency AI claims next.

The Skeptic Steelman the case against these two, because they're selling. Heimlich asserts agencies build nothing and just repackage Google and Microsoft tools. His evidence is one evasive earnings answer. That's inference dressed as fact, and Chalice competes directly with agency in-house stacks, so weight it accordingly. Glaser's claim that Moat's death left publishers with "basically no tools" is overstated. IAS and DV run publisher-side products, and Confiant and others exist. The defensible version is narrower: free, proactive, placement-level monitoring is thin. Fine, but "narrower and more defensible" is not the headline they sold. The theses are directionally interesting. The specifics are self-serving.

The Operator Strip the theory and ask what breaks Tuesday morning. The prisoners'-dilemma read on why publishers game fields is the most useful thing in the episode: pooled algorithms concentrate bids on a thin slice of inventory, so publishers manipulate metadata to look like that slice. If you run a supply business, you already know this. You do it because your fill depends on it. A neutral on-page signal layer only helps you if buyers actually bid off it, and today they don't. So a publisher adopting Gamera-style labeling is betting on a demand-side behavior change that hasn't happened yet. On the buy side, custom algorithms are real work: your own outcome data, your own page metrics, your own maintenance. Not a switch you flip.

The Customer / End User For the advertiser, the per-advertiser architecture argument lands. If your bidding is pooled with fifty other brands chasing the same "premium" impressions, you're bidding against yourself and eroding your own brand equity. Heimlich ties slowing brand growth to exactly this. In plain terms: shared algorithms make everyone chase the same inventory, which raises what you pay and narrows what you reach. But the advertiser has to supply clean outcome data to make a custom stack work, and most don't have it organized. The attention-vendor swipe lands. If you can correlate what created attention directly to outcomes, you don't need attention as a middle metric at all.

The CFO Custom-per-advertiser sounds great until you price the labor. Pooled algorithms exist because they're cheap to run across many accounts. Bespoke stacks mean per-advertiser build and upkeep, which is why this stayed a niche for years. The payback only works for advertisers big enough to amortize that cost against real budget, which is the enterprise tier Heimlich names. For everyone else the math doesn't clear. On the publisher side, Gamera is free, so the cost question flips: what's the opportunity cost of exposing objective labels that might rate your inventory lower than the fields you currently set yourself? That's a revenue risk, not a line item.

The tensions:

Heimlich and Glaser agree programmatic's valuation promise failed, but they disagree on the cure. Heimlich says fix the buy side with custom algorithms. Glaser says fix the signal by making inventory labels un-gameable. Those are different bets. One says the demand side is broken, the other says the supply signal is.

The Operator and the Customer split on feasibility. The advertiser wants per-advertiser bidding; the operator points out it requires clean first-party outcome data most brands haven't organized. The thesis is right and unusable at the same time for most of the market.

The Skeptic and the Market Analyst split on how much to trust the messengers. The money is genuinely moving toward custom and neutral. The men describing that move happen to sell both, and both are funded by the same DSP.

What it hinges on: whether buyers actually change behavior. Both theses die if demand keeps rewarding gamed fields and pooled bidding. Custom algorithms only pay off for advertisers with organized outcome data and budget to amortize the build. Neutral labeling only pays off if buyers bid off it. The council leans toward the diagnosis being correct and the timeline being slower than either founder implies. Test it cheaply: if you're a large advertiser, run one custom-algorithm pilot against your pooled baseline on your own conversion data. If you're a publisher, sample the free labeling before you expose it programmatically, and see whether your metadata survives contact with an objective read.

Prediction: At Omnicom's next two quarterly earnings calls through Q1 2027, John Wren or his CFO will again decline to disclose a specific split of AI efficiency gains between agency margin and client savings.

Confidence: Medium. Agencies have every incentive to keep the AI margin question vague.

Why: Wren already punted once when an analyst asked directly, and the reason is structural: if Omnicom admits AI savings flow to margin, clients demand price cuts; if it admits savings flow to clients, investors question the growth story. There's no answer that helps both audiences, so the rational move is to keep repackaging generic capability as proprietary build without disclosing milestones. The opposite outcome, a clean numeric disclosure, only happens if a competitor forces it or a client contract makes it public, and neither pressure is visible yet. The pending Omnicom-IPG integration gives Wren every additional reason to stay vague.

Revisit by 2027-02-28: We're right if Omnicom's next two earnings calls contain no specific agency-versus-client AI savings split. We're wrong if leadership quantifies where the AI efficiency goes.