An Interview with Eric Seufert About Models and Ads, and AI’s Upside for Humanity

Below top-k · 118

• claimA quiet revolution in RecSys is underwayc 0.95

  Recommender systems are transitioning away from the traditional DLRM two-tower paradigm toward generative recommendation, and this shift is the most important thing happening in the space right now.

• claimAI shifts the binding constraint from production to distributionc 0.95

  When AI causes an explosion of output, the scarce resource is no longer making things but sorting through them to find what people actually want. Distribution becomes the principal allocation mechanism.

• claimBeing pro-ads may correlate with AI optimismc 0.90

  The closing thesis asks whether being pro-advertising is correlated with being optimistic about AI's impact on humanity — framing ad-friendliness and techno-optimism as linked dispositions.

• claimZuckerberg is a poor communicator, especially about adsc 0.90

  However ruthlessly effective he is as an executive, Zuckerberg does a bad job articulating the vision for Meta's ads platform — when he bothers to talk about it at all, which often he doesn't.

• claimPlatform-native agentic commerce is already workingc 0.90

  When the agent is built into the retailer — Amazon's Rufus, Walmart's Sparky — agentic commerce works well right now. The interface isn't the problem; the conflicting incentives of cross-platform integration are.

• claimGoogle's Gambit was a forced response to chatbot habituationc 0.90

  Once consumers got accustomed to chat-like interfaces, Google had to adapt or watch users defect to ChatGPT. But adapting too fast would itself feel like a foreign product, so they had to move in steps.

• claimChatbot advertising demonstrably worksc 0.90

  Skepticism about chatbot ads is misplaced — the format works, and OpenAI's early traction proves there is real demand even before the product is good.

• claimBy being late, OpenAI ceded the right to define the rulesc 0.90

  OpenAI should have launched ads a year earlier. Because they were late, Google got to set advertiser and consumer expectations for what chatbot ads should look like, and OpenAI is now playing on Google's playing field.

• mechanismGenerative recommendation flips the problem aroundc 0.90

  Instead of scoring every possible piece of content against a user's history, generative recommendation simply predicts the next item the user is likely to engage with, like next-token prediction. Google's 2023 TIGER paper introduced this with Semantic IDs.

• implicationOwning a frontier LLM may become existential for ad platformsc 0.90

  If the best LLMs deliver decisive advantages in ranking, frontier labs will increasingly keep them in-house rather than sell access. Platforms that don't own their own models risk being locked out of the next phase of ads.

• implicationApple will run an 'LLM-washing' playbook mirroring its ad strategyc 0.90

  Instead of building, Apple will likely create a model garden — letting users select a default LLM the way they pick a default browser — and collect distribution payments from frontier providers, just as Google pays for search default placement.

• claimThe OpenAI–Amazon deal likely includes Amazon serving ads on ChatGPTc 0.90

  The investment looks like a product alliance where Amazon supplies the ad inventory and fulfillment inside ChatGPT, building on OpenAI's existing bid-in arrangements with Criteo and others.

• claimHuman desires are effectively infinite and combinatorialc 0.90

  Optimism about AI plus ads rests on the belief that wants are not a fixed pie but expand combinatorially per individual. The job is to match infinite need to newly capable supply.

• implicationPre-testing claws back the 10% testing budgetc 0.85

  Advertisers routinely spend around 10% of budget on testing creatives with no ROAS accountability. A distilled pre-testing model lets you screen creatives before spending real dollars, recovering most of that loss.

• mechanismSandberg was Meta's standing champion for the advertiserc 0.85

  On every earnings call Sandberg evangelized for advertisers, citing the ten-million-strong base and concrete success stories. That role disappeared when she left, and no one at Meta has filled it.

• mechanismAffiliate fees are economically inferior to ad biddingc 0.85

  An affiliate fee is an arbitrary percentage off the top line, while ads let advertisers submit their true value as a bid. Bidding surfaces higher-value inventory and captures more economic surplus than a fixed cut ever can.

• implicationThe transformation cuts off the open web and captures more valuec 0.85

  By absorbing answers into the Search surface, Google converts itself from a distribution mechanic that sent traffic outward into a destination that retains users and value internally. The open web loses its referral lifeline.

• implicationAn engagement surface opens a new monetization layer beyond keyword adsc 0.85

  Once Search has dwell time and a richer canvas, it isn't limited to monetizing the click — display-style and conversation-contextual ads become possible on top of search ads. The transformation unlocks new revenue forms, not just defends old ones.

• evidenceOpenAI booked $100M in ad revenue within three weeksc 0.85

  OpenAI generated roughly $100 million in ad revenue in the first few weeks after launching their beta, with a target of $1 billion this year — even though most observers consider the current ads terrible.

• claimGoogle's interconnected ecosystem is its real advantage in chatbot adsc 0.85

  Google can piggyback chatbot ads off existing advertiser relationships, data, and placements across its entire product suite. That means more information AND more ways to deploy a budget — something OpenAI structurally cannot match.

• implicationGenerative ranking forces ads and recommendations into one solutionc 0.85

  Ads and recommendations have always been the same problem; generative ranking is the pivot point that finally makes them the same solution, and the optimal org structure collapses into a single unified team.

• implicationProprietary models for ranking will stay proprietaryc 0.85

  If a model delivers 5-6% quarterly gains in time-spent or conversion rate, the owner has no incentive to expose it for competitors to distill or fine-tune. The most valuable ranking models will be kept entirely in-house.

• mechanismBuilding frontier models would force Apple to break its privacy brandc 0.85

  Training competitive frontier models requires harvesting unique user data, which directly contradicts Apple's privacy marketing and its heavy commitment to Differential Privacy. They've boxed themselves in.

• mechanismAmazon's identity spine covers nearly all US consumersc 0.85

  Amazon has logged-in identity on roughly 90% of the United States, paired with actual retail purchase data rather than engagement signals. That's the most valuable advertising dataset ever assembled.

• implicationProduct ads are the natural fit for ChatGPT, and Amazon owns thatc 0.85

  The ad format that actually works inside a chatbot is product recommendations with a buy link. Offloading that entirely to Amazon solves OpenAI's ad-format problem in one move.

• claimMatching is indifferent to what people want, and that is the pointc 0.85

  The ad-driven matching system has no preference about consumer desires; it simply finds each person the thing they will most enjoy. That neutrality is what makes everyone — advertisers, consumers, platforms — gain value simultaneously.

• implicationAI makes ad-driven matching dramatically more powerful, not obsoletec 0.85

  Generating infinite homogenous output is not what humans want; AI's real upside is making the matching of unique desires to specific products vastly better. The ad-pessimist read of AI gets this backwards.

• contextAd platforms are converging on total end-to-end automationc 0.80

  Every major digital ad platform is moving toward a paradigm where the platform handles targeting, bidding, and increasingly creative itself. Advertisers are losing visibility into how their money is being spent.

• implicationKeyword-based search is going away in a multimodal worldc 0.80

  Pure keyword querying doesn't fit a conversational, multimodal interaction model. De-tethering monetization from keywords means Google can use the whole buffalo — extracting value from full context and conversation rather than just query terms.

• implicationIf forced to pick one, advertisers will pick Googlec 0.80

  Both Google and OpenAI require their own integrations, data modeling, and creative — but with Google an advertiser also gets the rest of the Google ecosystem for free. Given the choice, advertisers will default to Google.

• implicationWorld knowledge from LLMs is an untapped growth leg for adsc 0.80

  The semantic and world knowledge an LLM brings to predicting the next thing a user will engage with represents a major leg of growth that platforms have barely started to capture.

• claimGEM is a foundation model, not a toolc 0.80

  Meta's Generative Ads Model (GEM) is genuinely a foundation model — the output of a research team — and people misread it as just another product feature. That misreading hides the scale and seriousness of Meta's infrastructure bet.

• mechanismAI expands the product space into finer-grained nichesc 0.80

  AI makes creators more efficient at research, packaging, and execution, producing more granular products that satisfy increasingly specific needs. This enables more precise matching between supply and demand.

• claimSilicon Valley's blind spot on advertisingc 0.80

  OpenAI's long resistance to advertising — and Zuckerberg's apparent ambivalence — reflects a failure to grasp that ads discover price and meet consumer needs. Ads are not just a revenue model, they are a value-creation engine.

• claimAdvertisers are left with one or two levers of controlc 0.75

  Behind the platform's opaque event horizon, advertisers can only meaningfully influence the creative and the stated objective. And with generative AI, even the creative may no longer be theirs.

• claimMany advertisers will refuse to hand creative to the platformc 0.75

  Brand conflicts and the need for consistency across multiple platforms mean serious advertisers won't surrender creative production to Meta or anyone else. The advertiser is the only party who works across platforms.

• mechanismExternal checkout breaks the recommender flywheelc 0.75

  Retailers make money by expanding the cart at checkout through recommendations. A chatbot that buys one item at a time bypasses that machinery and crushes average order value.

• mechanismNative agents preserve the data, ad, and cart stackc 0.75

  When Rufus sits inside Amazon, Amazon keeps the data, the ad surface, and a consistent cart. The reason native works and bolt-on doesn't is that the retailer doesn't have to give anything up to enable it.

• claimOpenAI is executing on ads infrastructure blazingly fastc 0.75

  OpenAI has rapidly shipped a CAPI, a pixel, and self-serve, putting the foundations of a modern ads platform in place far quicker than skeptics expected.

• claimThe embarrassing first launch was the pointc 0.75

  OpenAI's initial ads product was meager and unimpressive, but shipping it created an embarrassment function that forced rapid iteration. The bad first version was the mechanism for shipping a useful one faster.

• mechanismRQ-VAE codebooks compress catalogs into tokensc 0.75

  TIGER uses Residual Quantized VAEs to train hierarchical codebooks — four codebooks of 256 each cover a billion-item catalog by combinatorics. Each codebook entry becomes a token an LLM can predict, turning recommendation into next-token prediction over a tiny vocabulary.

• mechanismGEM is distilled into ~40 ranking models across surfacesc 0.75

  GEM itself is too large for inference, so Meta distills it into roughly 40 point solutions for different ranking scenarios — as described in their 2024 SUM paper — spanning both organic and ads surfaces.

• claimAmazon's ad strategy is serving more surface area, not owning morec 0.75

  Amazon doesn't need to own the destinations where its ads appear. With a scaled DSP and unmatched identity data, it just needs access to inventory wherever it lives.

• contextGalbraith's Affluent Society as the foilc 0.75

  Galbraith argued advertising existed to manufacture demand for surplus output. The Prosperous Society thesis inverts this: matching, not manufacturing demand, is what advertising actually does.

• claim'Human labor will be worthless' is the lump-of-labor fallacyc 0.75

  The belief that AI will make human labor valueless is the same broken mental model as the lump-of-labor fallacy — both assume a fixed quantity of work or value rather than an expanding frontier of new needs.

• contextInterview centers on creative measurement as the unlock for generative AI in adsc 0.70

  Seufert's master's thesis describes how companies can build models to measure the effectiveness of ad creative using ad platform black boxes, which is what unlocks the real value of generative AI in advertising.

• claimZuckerberg is failing Meta by not defending advertisingc 0.70

  Mark Zuckerberg is hurting Meta by declining to publicly advocate for ads as a legitimate and valuable business model.

• mechanismMeta is using foundation models to transform ad targetingc 0.70

  Meta is applying foundation models to overhaul how ads are targeted, moving the targeting stack onto modern AI infrastructure.

• contextThe cost of producing creative is trending to zeroc 0.70

  Whether platforms generate the creative or advertisers use third-party generative tools, the marginal cost of producing an ad asset is collapsing. This changes what's economically rational at the creative-production stage.

• implicationPerformance marketers can fire-hose the entire creative surfacec 0.70

  Because production cost is near zero and pre-testing filters out losers before they touch the testing budget, DTC marketers can generate every plausible messaging-and-visual combination and let the model triage. Exploration is no longer bounded by spend.

• claimMeta's "just give us your credit card" pitch only resonates with DTC, not brand advertisersc 0.70

  Zuckerberg's vision of Meta handling creative, measurement, and everything else has been the bull thesis for direct-to-consumer advertisers, but brand advertisers don't buy it — and that gap is a real part of Meta's problem.

• claimTargeted ads are a societal good that enables new industriesc 0.70

  Meta-style advertising makes entirely new businesses and niches viable by letting them surface to relevant customers — a positive case that contrasts with the slightly icky feel of search ads.

• evidenceChatGPT's Instant Checkout has effectively floppedc 0.70

  The Instant Checkout product appears deprioritized or gone, and Walmart publicly expressed dissatisfaction with its participation — confirming the skepticism that the integration would not work.

• evidenceWalmart said clickthroughs from ChatGPT converted terriblyc 0.70

  Walmart's complaint was that conversions through the integration were worse than just getting fewer users to land on their own site, where they control the funnel — a direct articulation of why bolt-on agentic commerce fails.

• evidenceAmazon's refusal kills the third-party agent thesisc 0.70

  Third-party agentic commerce needed Amazon to participate, and Amazon has not — it's actually blocking more chatbots in its robots.txt, removing the largest possible retail surface from any independent agent layer.

• claimDisplay ads work anywhere attention is appliedc 0.70

  The objection that AI is too 'lean-in' for advertising misreads attention. Any context where users are paying attention is a candidate for display advertising, regardless of whether they're being productive or passive.

• caveatMost chatbot conversations won't be monetizablec 0.70

  If ads are anchored only to conversational context, a large share of sessions will carry no commercial intent — historically around 80% in search — so chatbots will need to monetize users across their broader internet activity to capture full value.

• mechanismSearch is constrained by the MAYA principlec 0.70

  Google Search is a victim of Most Advanced Yet Accessible: change it too much and users no longer recognize the product they know how to use. That forces incremental evolution rather than wholesale redesign.

• evidenceThe gambit has already paid off on Google's own metricsc 0.70

  Google reports monetizing AI results at parity with traditional results, queries are up rather than down, and Search revenue grew 19%. The transformation didn't cannibalize the franchise.

• caveatConsumer pre-training by chatbots was the necessary catalystc 0.70

  Without ChatGPT first normalizing the conversational interaction mode, Google's pivot would have felt disorienting and probably wouldn't have worked. An external behavior shift gave Google permission to change.

• claimProducts and ads are the same system informed by the same researchc 0.70

  At Meta, the recommendation and ads stacks aren't literally unified but they're driven by the same underlying research, which is why advances in one immediately propagate to the other.

• evidenceMeta, Google, and Kuaishou are rebuilding around LLM rankingc 0.70

  The largest ad platforms are publishing papers about rebuilding ranking architectures on LLMs. Meta now mentions LLMs for ad ranking in earnings calls, signaling how high in the org this shift has reached.

• exampleMeta Spark is the endorsement of Meta building its own modelsc 0.70

  Spark launching the same week as Mythos, even slightly behind the frontier, is actually the strongest possible argument that Meta must own its model stack — and the fact that Spark is competitive means they're on the right path.

• claimApple engages in 'ad-washing' by outsourcing its ad businessc 0.70

  Apple is highly reliant on advertising revenue, but routes it through Google's search rev-share and other partnership deals so it doesn't look like an ad business. The margin on these deals is essentially 100%.

• mechanismSocial inventory has structurally better unit economics than gamingc 0.70

  Gaming requires constantly buying users for a fixed 90- or 180-day LTV. A social network offers viral distribution and multi-year retention per acquired user, which is why AppLovin would experiment here after selling its games studio.

• claimAppLovin's edge is algorithmic targeting on non-logged-in usersc 0.70

  Their core advantage is figuring out who anonymous users are and what they want better than anyone else. That capability is portable to any surface — CTV being the most obvious next expansion via their Wurl acquisition.

• mechanismDigital ad systems are already extremely efficient allocatorsc 0.70

  Because digital ad platforms are highly efficient at matching, shifting more of the economy through them automatically creates growth and value.

• contextDegrowth as the dominant default discoursec 0.70

  Much of the AI commentary is shaped by people who have internalized a degrowth mindset, with Galbraith's Affluent Society serving as a kind of degrowther manifesto. Without active rebuttal, that framing wins by default.

• implicationOptimism about ads predicts optimism about AIc 0.70

  Those who understand advertising's role in price discovery and need-matching see AI as a force multiplier; those suspicious of commerce see it as a threat. The split runs along the same axis.

• exampleMobile gaming proves lean-in contexts monetizec 0.65

  Mobile gaming is intensely active rather than passive, and ads work very well there — direct counterevidence to the claim that chatbot users are too engaged or productive for advertising to land.

• caveatOpenAI's CAPI is for measurement only, not optimizationc 0.65

  Modern ad platforms use conversion signals to optimize ad serving toward specific events, but OpenAI only lets advertisers measure those events. Bidding is still CPC, and the advertiser has to judge profitability themselves.

• exampleCheckout vs. Shopify shows OpenAI losing the protocol warc 0.65

  OpenAI got over their skis with Checkout and related shopping protocols, then Google arrived with Shopify and apparently won the standard. Stripe now supports both — concrete evidence that OpenAI no longer defines the playing field.

• claimCapitalism is objectively better and must be defended explicitlyc 0.65

  Capitalism creates more satisfaction and fairer societies than its alternatives. Optimists need to make this case analytically and tied to canonical works, or the pessimists' framing dominates by default.

• claimGoogle has reoriented search around AIc 0.60

  Google has shifted its search product toward AI-driven experiences rather than the classic ten blue links model.

• implicationPlatforms may resent losing the free testing-budget revenuec 0.60

  That 10% testing spend was effectively free money for platforms, absorbed as COGS with no ROAS scrutiny. Widespread pre-testing turns it into a visible cost saving for advertisers and a visible revenue loss for platforms, which could make platforms hostile to the methodology.

• evidenceNobody can name Meta's current COOc 0.60

  Sheryl Sandberg's successor, Javier Olivan, has so little public profile that even close industry observers can't recall the name — a stark contrast with how visible Sandberg was.

• exampleGoogle's repetitive advertiser anecdotes are doing real workc 0.60

  Google's habit of trotting out five advertiser success stories per call can feel eye-rolling, but the repetition of "look at this real-world result" builds a narrative around the platform — something Meta now conspicuously lacks.

• mechanismAffiliate models bias toward cheap, low-value goodsc 0.60

  Because affiliate fees are a fixed percentage, the system tends to prefer low-cost items with higher conversion rates, degrading the user experience and depressing revenue compared to ad auctions.

• exampleAI Overviews and AI Mode as the staged rolloutc 0.60

  AI Overviews displaced links with summaries; AI Mode made conversational interaction a core part of Search. Together they are the concrete steps of the Ship of Theseus swap.

• claimDisplay is not a sideshow to search in digital advertisingc 0.60

  It's a misconception that display is a tiny slice of digital ad spend — it's actually a huge part of the economy. The instinct that keyword search is de facto superior to display is wrong.

• caveatZuck's Superintelligence framing obscures the real storyc 0.60

  By talking about Superintelligence rather than GEM and ad ranking, Zuckerberg invites Metaverse-style skepticism and makes people dismiss as cosmetic what is actually a concrete, commercially decisive bet.

• contextApple already builds and ships its own models on-devicec 0.60

  It's a misconception that Apple doesn't build models. They have a 3B-parameter LLM running on the iPhone, plus vision transformers and BERT, all accessible to developers via Core ML.

• mechanismAutomated ad platforms onboard ever-smaller SMBsc 0.60

  End-to-end automation in ad platforms brings in smaller and smaller businesses that could never have advertised before, broadening participation in the commerce economy.

• claimBubble-or-not depends entirely on whether the matching worksc 0.60

  If the AI-plus-distribution thesis works, current CapEx is a growth trajectory; if it doesn't, it's a bubble that bursts into recession. There is no middle outcome.

• evidenceThe deep network beat XGBoost on this prediction taskc 0.55

  Seufert empirically showed that a deep neural network outperforms a tree-based baseline — XGBoost being the canonical comparison — at predicting context-conditioned creative ROAS from historical platform data.

• caveatThere's not much platforms can do to stop itc 0.55

  The approach piggybacks on the very opacity platforms have created — it only needs inputs and observed outcomes. Retraining will be needed regularly but is cheap, so the approach should improve over time.

• mechanismEvery integration hop costs conversionsc 0.55

  If millisecond delays measurably hurt conversion on a site where the user is already trying to buy, any additional integration step between intent and checkout is almost definitionally destructive.

• contextCAPI exists because Apple killed browser-based trackingc 0.55

  CAPI is a server-to-server replacement for the pixel, invented after Apple's ITP blocked third-party and then first-party cookies in Safari. Facebook responded by moving tracking out of the browser entirely, doing probabilistic matching on IP and other signals on their backend.

• mechanismShipping ads early may have been an internal political movec 0.55

  Pushing out a primitive ads product and immediately landing $100M in commitments may have been a way for ads advocates inside OpenAI to prove demand and unlock real engineering resources.

• caveatYou don't want to make advertisers compare you to Googlec 0.55

  Forcing advertisers into a head-to-head comparison with Google on ads is a losing position. OpenAI's late entry has put them in exactly that situation.

• caveatUpgrading to a new model isn't free for existing toolsc 0.55

  Once a tool is built on a model that works, upgrading means resampling from a different output distribution, rewriting system prompts, and retesting. If the old model is cheaper and stable, there's little reason to upgrade.

• implicationMost retail is still offline and ripe for efficient matchingc 0.55

  83% of retail commerce is not yet online; bringing more of it into efficient digital distribution would unlock large gains in both consumption value and producer revenue.

• claimChat checkout failed as a commerce paradigmc 0.50

  The conversation examines why chat-based checkout flows did not take off as a viable commerce surface.

• claimChatGPT ads are evolving rapidlyc 0.50

  The advertising layer inside ChatGPT is developing quickly and is treated as a significant emerging surface.

• contextAd serving and content recommendation are the same technologyc 0.50

  The retrieval and ranking methods used to pick ads are identical to those used to pick content in recommendation systems — only the objective function differs. They became separate tech trees for historical reasons, but the underlying machinery is the same.

• contextEnd-to-end automation is the direction of travel regardless of preferencec 0.50

  Whether or not an individual advertiser likes total platform automation, it's working for the platforms and is where the market is heading. Strategy has to start from that assumption rather than fight it.

• implicationBetter AI-generated creative is win-win for platforms and advertisersc 0.50

  If advertisers deploy creatives that scale and reduce dead weight from ads that don't work, they'll spend more, which makes platforms happy too. Advertisers bearing the compute cost of generation is, at minimum, a bonus for the platforms.

• implicationForced ad breaks may be more valuable than feed adsc 0.50

  AppLovin's pitch is that an enforced pause where the user must watch the ad is more attractive than social feeds. Chatbot ad breaks tied to conversation context could inherit the same dynamic.

• contextSearch is the wrong mental model for chatbot adsc 0.50

  Seufert's broader argument, made in a recent piece, is that chatbot advertising should not be modeled on search advertising — a position that frames the entire discussion of how AI monetization should be structured.

• implicationConversion optimization is just a matter of timec 0.50

  Given how fast OpenAI is executing, conversion optimization on top of the CAPI is likely coming soon — the current measurement-only setup is a stopgap, not a permanent limitation.

• contextThe old DLRM two-tower paradigmc 0.50

  Since YouTube's 2016 DLRM, recommendation has worked by building a user tower and a content tower with large embedding tables and scoring candidates via dot product similarity over user history and possible content.

• claimFrontier LLM value is currently captured at the frontierc 0.50

  All the economics of LLM adoption are being absorbed by the latest frontier models, as workflows integrate them for the first time — which seems to support the idea that value will keep moving to whoever holds the frontier.

• caveatOn-device frontier models are blocked by memory, not by intentc 0.50

  The real near-term blocker for putting a frontier LLM on the phone is memory — the KV cache would be too big. But ongoing work on quantization and cache compression should eventually relax this.

• caveatAtomization and social cohesion are real but secondary risksc 0.50

  Hyper-personalization could produce alienation and weaken social cohesion, but consumers will reject the extreme version. The bigger risk is that the bet simply fails to pay off economically.

• evidenceMissing the Meta upside as evidence the matching already worksc 0.50

  Many investors and commentators missed years of Meta (and Google) returns precisely because they underestimated how well ad-driven matching of expanding needs to expanding supply already functions.

• implicationBrand marketers keep control of messaging across platformsc 0.45

  For brand-conscious advertisers, pre-testing means they can centrally produce and govern their creative everywhere without surrendering brand voice to any single platform's generator.

• evidenceInnovation across the model stack is happening at lightning speedc 0.45

  Techniques like TurboQuant are only months old. The bubble is funding parallel work on algorithms, KV cache, memory, and quantization, so breakthroughs that change what fits on a phone are likely.

• contextSearch remains one of the great business models even nowc 0.40

  Google Search grew 19% last quarter on a massive base, making it arguably one of the best business model innovations ever. Any critique of Google's strategy has to start by acknowledging that performance.

• contextMeta closing the ad-revenue gap reframes the stakesc 0.40

  Headlines that Meta is about to eclipse Google's ex-TAC ad revenue make Google's pivot look defensive as well as offensive. Seufert thinks the eclipse claim is overstated, but the competitive pressure is real.

• mechanismHow the pixel and cookie combine to measure conversionsc 0.40

  The pixel is a 1x1 image embedded in a webpage that fires on load, signaling where the user is in the shopping journey. Paired with a cookie carrying the click ID and UTM, it lets advertisers tie post-click behavior back to a specific ad campaign.

• contextPixel is old-school, CAPI is new-school, but you need bothc 0.40

  Pixel-based tracking represents the pre-Apple-crackdown world and CAPI the post-crackdown world. Many advertisers still aren't equipped to set up CAPI, so the two coexist.

• contextAppLovin launched a social app called Gistc 0.40

  AppLovin quietly launched Gist, a Pinterest-meets-RedNote social product gated by referral codes. Early usage isn't compelling but it's an explicit trial balloon.

• contextAppLovin is still overwhelmingly a gaming businessc 0.40

  Despite the e-commerce initiative growing fast and not yet in GA, AppLovin's revenue is still almost entirely gaming — a category that is finally growing again after years of ATT-induced stagnation.

• contextWhy DeCANT exists at allc 0.35

  DeCANT — Deep Creative Attention-based Network for pre-Testing — is the thesis project from Seufert's Harvard Master's in Applied Computation, motivated by his sense three years ago that he couldn't credibly comment on digital ads without understanding ML from the inside.

• contextThe series was a direct rebuttal to the Citrini short reportc 0.35

  Seufert abandoned a planned book to produce a four-part podcast series specifically because he felt the bearish Citrini report needed an immediate, pointed rebuttal in public discourse.

• contextApple, Amazon, and AppLovin also in scopec 0.30

  The interview also covers Apple, Amazon, and AppLovin as part of the broader landscape of platforms reshaping ads and AI.

• contextMeta is treated as the stand-in because it's best in classc 0.30

  The conversation focuses on Meta not as an outlier but because it's the most advanced ad platform, making it the natural proxy for talking about the industry as a whole.

• contextSeufert's 'Agentic Commerce Is a Mirage' piece drew unusual hostilityc 0.25

  Seufert wrote a piece called Agentic Commerce Is a Mirage that generated more hostile feedback than anything else he's published, including his harshest Apple criticism.