On the All-In Podcast, AppLovin’s CEO drew a structural line between two kinds of commerce that most agent coverage conflates — and the line lands directly on the business model of advertising.
Editorial note: This article summarises a podcast conversation. Every claim below is Adam Foroughi’s statement, attributed accordingly. Nothing here constitutes investment advice, and no claim is treated as established fact unless it is independently sourced.
The Two Poles of Agentic Commerce
Replenishment
Answer already known. Execution problem only. Automation removes effort without requiring judgment.
Discovery
No pre-existing answer. Demand must be created before anything can be executed. Advertising lives here.
In a conversation on the All-In Podcast, AppLovin co-founder and CEO Adam Foroughi offered a distinction that cuts through most of the ambient noise around agentic commerce. His core observation, in his own framing: an agent can take over a supplement subscription and optimise the delivery cadence every month, but that is a categorically different task from whatever made you want the supplement in the first place. In his words: “I might put my supplement subscription into an agent and have it optimized every single month and deliver on time. But these discovery platforms aren’t that.”
Foroughi’s characterisation of who uses these platforms is his view of his own market — not a measurement. He described the typical shopper as, in his telling, closer to the New York Times reader than to an early-adopter on social media still experimenting with the latest agent tooling. That framing is offered as competitive context, not as audience data, and it should be read that way.
The conversation arrived in a week when two large platforms had taken opposite public positions on whether external agents should be permitted to transact on their surfaces at all — a policy divergence that makes Foroughi’s structural argument about the underlying task types all the more relevant to follow.
The key insight: Replenishment and discovery are not two points on the same automation curve. They are structurally different tasks. Replenishment is a solved problem being re-executed; discovery is an unsolved problem that requires demand to exist before execution can begin. The commercial consequence is that advertising revenue sits on the discovery side of that line — and an agent optimising subscriptions is not competing for the same dollars as the system that created the desire in the first place.
The Structural Read
The standard agent-versus-advertising narrative assumes a clean displacement: as agents handle more purchasing decisions, the advertising machinery that influenced those decisions loses relevance. Foroughi’s framing does not deny that agents matter. It questions whether the two systems are actually competing for the same territory.
The reason replenishment falls to automation first is not that it is technically simpler. It is that the answer is already known. When someone re-orders the same product on the same cycle, there is no latent preference to surface, no new consideration to introduce, and no discovery moment required. Automation removes effort. Discovery cannot be automated in the same way because the thing that needs to happen — creating demand that does not yet exist — is not a re-execution of anything. It is an origination event.
Foroughi’s concession on language models is the part of this conversation worth holding onto. He described advertising as, in his words, “ML 1.0 but really was the first implementation of all these technologies that now are driving AI today.” He did not claim that advertising is more economically important than language models — he explicitly granted the opposite. But he identified advertising as the domain where large-scale learned models on commercial traffic were running profitably for years before anything was publicly called artificial intelligence. That pattern, where the first profitable deployment of a technique happens in an unglamorous domain and the technique gets renamed when it reaches a more visible one, is real enough to be worth tracking regardless of whether his specific history is the most precise account of it.
Adam Foroughi — All-In Podcast (his claim)
“Now recommendation systems are structured differently than large language models but in a lot of ways they follow the same trajectory. So a lot of the research that’s being done in the space in the large language model space can port to recommendation systems.”
That is Foroughi’s claim, not an established technical result, and it should be held as such. But the implication, if it holds even partially, is awkward for the standard narrative. If the same body of research improves both LLMs and recommender systems, then the agent and the advertising machinery are not in a zero-sum capability race. They are neighbouring applications of overlapping methods. “Agents will kill advertising” as a thesis requires a one-sided transfer of capability that the underlying research — if Foroughi’s read is even directionally right — does not respect.
FDE Framework — Distributor Lens
The Outcome-Pricing Asymmetry
Foroughi described a pricingstructure where the platform sells outcomes rather than impressions: “The better it works, the better advertiser return is on our platform and everything is performance based. So we’re selling revenue to advertisers, the more they scale.” That is his description of his own business, not a verified metric. The structural principle underneath it is general: when a contract is written around outcomes, a model improvement converts directly into revenue because the customer simply buys more of something that now works better — no renegotiation required. Under impression-based or seat-based contracts, the identical improvement has to be argued for at the next renewal, and some of its value is captured by the buyer in the interim. Two platforms can ship the same technical advance and see materially different financial effects, entirely because of how the contract is written.
Three Implications
IMPLICATION 1 — TASK TYPE DETERMINES DISPLACEMENT RISK
The question of whether agents threaten a given revenue stream is not answered at the category level (“commerce” or “advertising”). It is answered at the task level. Replenishment-oriented revenue — subscription logistics, re-order fulfilment — faces a different automation gradient than discovery-oriented revenue. Foroughi’s framing, if accepted, means that the correct unit of analysis for any competitive assessment of agents is the specific task, not the industry vertical. That is a more granular frame than most current commentary applies.
IMPLICATION 2 — SHARED RESEARCH CUTS BOTH WAYS
If Foroughi’s claim that LLM research can port to recommendation systems is directionally correct, then the capability advance driving agent adoption also sharpens the advertising machinery agents are supposed to displace. This does not resolve which side wins — nothing here predicts that — but it does mean the threat model that assumes a one-directional capability transfer is underspecified. Platforms running recommender systems at scale are not passive recipients of disruption; they are operating on adjacent technical terrain.
IMPLICATION 3 — CONTRACT STRUCTURE IS A CAPABILITY MULTIPLIER
Foroughi’s pricing observation generalises beyond his own business. The same model improvement produces different revenue outcomes depending entirely on whether the contract is written around outcomes or inputs. Businesses selling outcomes capture model gains automatically; businesses selling impressions or seats have to negotiate them into the next renewal cycle. As AI modelquality continues to improve across the board, contract structure becomes a meaningful differentiator in who actually captures the economic value of that improvement — independent of who built the model.
Business Engineer Framework
The Map of AI — Where Recommenders Sit in the Stack
Foroughi’s argument is partly a claim about layer positioning: recommendation systems are not below LLMs in a hierarchy waiting to be replaced, but beside them as a parallel application of overlapping methods. The Map of AI framework maps 200+ companies across nine layers of the AI stack — infrastructure, models, tooling, applications — and makes it possible to read these adjacency arguments with more precision than the “disruption” frame allows. If you want to locate where the replenishment-versus-discovery distinction actually lands in the commercial AI architecture, that is the right tool.
The most useful thing Foroughi said on the All-In Podcast was not a claim about his own platform’s position — it was the underlying structural cut: agents take replenishment before discovery because replenishment is a re-execution problem and discovery is an origination problem, and advertising revenue is on the origination side of that line. Whether that holds as agents develop further, and whether his read on shared LLM-recommender research is directionally correct, are open questions. But the frame itself — that task type, not industry category, determines displacement risk — is more analytically precise than most of what is currently circulating, and it came with an honest concession about where language models sit economically relative to advertising. That combination of structural clarity and acknowledged limitation is what makes it worth taking seriously.
Sources: Adam Foroughi on the All-In Podcast (YouTube). All claims attributed to Foroughi are drawn from that conversation and represent his statements, not established data or verified figures. This article does not constitute investment advice.
This summarises a podcast conversation, not data or a filing, and it is not investment advice.Every claim above is Adam Foroughi’s own statement, including his description of “the typical shopper”, which is his view of his own market rather than a measurement of consumers generally. His statement that research from large language models “can port to recommendation systems” is his claim, not an established technical result. No AppLovin revenue, margin, growth, share-price, market-size, user or advertiser figure appears above, no model-launch date, no competitor numbers, and no measurement of agent adoption or of what share of commerce is replenishment rather than discovery. Nothing above predicts an adoption rate, a timeline, or whether advertising or agents prevail.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
