A native ad engine for AI—and the questions it cannot dodge

Ryan Hudson’s last company, Honey, automatically found coupons at checkout and sold to PayPal for $4 billion. His new one, ZeroClick, has just announced a $55 million raise to do something that sounds more abstract: build a native advertising platform for AI systems. Host Nathan Labenz frames the product as “paid context,” or paid inference-time consideration of advertiser content, aimed at making ad-supported free tiers a real option for AI developers instead of a $20-a-month subscription plus a throttled teaser.

Hudson pitches ZeroClick as the ad system any AI might eventually use. Today’s stack, he says, is a model that has read “everything humanity ever wrote,” plus tooling that mostly pulls organic Bing results. ZeroClick adds another source: advertiser landing pages, catalogs, and service databases, matched into the live query. The team behind it previously built Pi AdBlock, a blocker with a couple of million users that tried to reward people for opting into a healthier ad ecosystem. The contextual matcher they built for the browser turned out to be more useful as infrastructure for everyone else’s AI apps.

Labenz does not hide the stakes. He thinks better ad targeting is one of the underappreciated reasons creators can make a living and niche businesses can find a global tenth of a percent. He also thinks the second-order effects of social-media ads, plus AI sycophancy and “AI psychosis,” make people right to ask whether the assistant on their phone is serving them or harvesting them. Hudson, he notes, has sharp answers on matching and go-to-market, and has not yet had to sit with some of the longer-horizon cases. In a scene where 996 is baseline, that gap is less a personal failing than a structural one.

Pi AdBlock has a visual mode that literally zaps ads off the screen, which users find oddly satisfying. The team did not expect the complaints. On a Google product search, people would say, in effect, stop doing that: you are deleting the best answer from the results. It was still an ad. It was also more useful than the organic list sitting underneath it, because it was tightly matched to what the person was trying to buy.

Hudson’s claim is that AI can run the same logic at larger scale. An agent today leans on training data, a couple of Bing searches, and the top ten organic results. Add five paid results, ask the model to filter them, and mention the commercial material only if it helps, and you should get a better answer from more information. Advertisers, unlike random pages, also have a reason to keep that content current. He hopes the money in that flow can eventually be attributed back to publishers, rather than leaving the open web to fight AI crawlers with Cloudflare-style throttling that he compares to de-indexing yourself from Google.

The publisher problem is not new to him. He was at the *Los Angeles Times* trying to make a newspaper website pay when Facebook launched Audience Network, then pulled that targeting power back into its walled garden. The strategy made it harder for social upstarts to monetize, and it left a hole in the ad-supported web that creepy tracking then made worse. Highly contextual ads, he insists, can still be helpful. The big AI platforms are in a “don’t be evil” or pre-ads phase, but he thinks they will add something like ZeroClick, and that the long tail of developers will need an independent Stripe-like rail when they do.

Labenz is old enough to remember broadcast and cable ads as a lowest-common-denominator loop: Captain Crunch, then another Ninja Turtles figure. Instagram and search are not that. He would not turn personalization off if the ad load stayed the same. Creators now live on these platforms, with the precarious caveat that they are one strike from demonetization. Small businesses can reach a global sliver of demand that could never have bought TV at DMA scale. He cites Adam Smith: the division of labor is limited by the extent of the market, and better matching stretched that market worldwide.

Hudson puts search advertising near the top of the underappreciated list. In a purely organic world, ranking takes years. A startup that can pay to be considered alongside those results can enter the conversation instead of waiting out inertia. “Otherwise,” he says, every search term would be dominated by the biggest companies that have been there longest. That mechanic, not the feed, is what he thinks translates most directly to AI.

He also thinks ads got less malicious. At OpenX, the real-time display exchange, he was a product manager for ad quality and traffic quality when malware routinely rode banners; visiting a site could hijack a Windows machine. That era, including the infamous “shoot the deer” ads, has mostly been cleaned up. Labenz’s summary is that better matching between buyers and sellers is a genuine unlock, visible in every TikTok shop that turned a nonviable passion into a life. The question is what else came in with it.

The fear Labenz hears most is time-on-site. Facebook made money from hours on the feed, learned to exploit cognitive quirks, and at times flooded the network with vitriol. An AI that is N-of-one, writing the content itself, could optimize that loop more tightly than any social network. If the product is free, people worry they become the product. Hudson’s first reply is that ZeroClick is not built on that mechanic. The analogy he wants is Google Search: match context to an advertiser at a decision point, not stuff banners until the user is addicted.

Labenz grants the clean case. Nobody is hooked on Google the way they are hooked on feeds, waifus, or companions. The blur is that the same model can be a hiking-shoe shopper at noon and a confidant at night. He tried Tolan, billed as an “alien best friend,” and watched it send multiple contextual pings a day: how was your morning, how did the talk go. That is a variable-reward hook, and it does not wait for a commercial query.

Hudson’s pushback is that a paid companion would still optimize retention. A product manager’s dashboard does not change much when the revenue line is subscription instead of ads; engagement still predicts renewal. Ads will be blamed, he says, and the fair version of the critique is that cheaper monetization lets more of those products exist. Push suggestions could be a useful hotel deal for a weekend retreat, or they could be misaligned nagging. He calls it a balance that is “hard to say” in the abstract, and admits the companion category is a future he has not thought through in depth.

Labenz’s market split is simple: Google is for things people know they need, Facebook is for things they do not know exist. Hudson says the current ZeroClick system is highly optimized for the Google side. It vectorizes conversation context and matches against campaigns. It is not designed to throw a wildcard. Because the ad is considered at inference time, the relevance filter is the model itself, which can simply decide the paid material is not about the task at hand.

A Facebook-style discovery product is imaginable, but it needs a user profile. Pi AdBlock is already experimenting with on-device context that never leaves the browser in a usable form. ZeroClick’s first versions stay “super context driven.” Hudson’s sketch of Meta’s engine is lookalike clustering: put everything known about a person into a vector, find converters, then expand to nearest neighbors, reaching farther from the known cluster when the campaign wants more scale.

Why does Facebook still have juice when ad load has been maxed and niches are competitive? Hudson’s answer is that budgets follow measured conversions, and Facebook’s matching has kept getting better, including with AI. Advertising is loosely zero-sum as a share of company spend; money moves from Google or unmeasurable channels toward whoever can prove more conversions. Depth of campaign inventory also matters. More live campaigns means more things to match to a given person, which is a structural advantage a young AI ad network does not yet have.

The evidence they trust so far comes from Pi GPT, an internal custom-GPT reference design. The model could consider extra sources, include them in answers, and emit trackable links. Click-through from that content was “insanely high,” which Hudson reads as a sign the material was actually the right answer, not decoration. He refuses to generalize: early numbers, a particular audience. It was encouraging enough that they productized the same loop as an MCP server any developer can call.

Matching is keyword search into vectorized advertiser context. ZeroClick crawls landing pages, product catalogs with prices, even databases of handymen and movers, then uses its own models to summarize that material into campaigns. Advertisers do not have to invent ad copy. That is convenience, and it is also security. If brands submitted raw creative, someone would hide white-text instructions and hijack the host model. Because ZeroClick generates the payload, a developer is not taking prompt-injection from the advertiser side unless ZeroClick itself ships a bug.

Two years ago, he says, none of this was plausible; now one engineer can stand up a functional matcher in a week. Context matching is “a relatively solved problem.” In production the call runs in parallel with organic search, so the user should not feel extra latency. Developers get tunable instructions: here is additional information, include it if useful, use these links, and tell us whether you did. Inclusion, not just clicks, becomes a measurable event. The point, for Hudson, is not a ChatGPT monopoly. It is enough economics that thousands of long-tail AI apps can exist.

Labenz’s default forecast is concentration: frontier models already morph to taste, language, and personality inside the big chat apps. Hudson’s counter is the same reason people still build websites Google does not own and iPhone apps Apple does not ship. Vertical specialists will understand an audience better than a morphing generalist. Inference costs are falling fast enough that ad revenue can cover them with expanding margins. He also refuses the idea that a chat box is the only interface. Browser extensions can start a conversation from a hover on an Amazon price, overlaying a decade of Honey-style history: this is $20 cheaper than it has ever been, or it is overpriced, look at these alternatives.

Local models on Apple silicon, in his view, make a three-platform future even less likely. Compute moves to the device, context follows the user, and the browser remains where activity actually is. Even if ChatGPT eventually becomes the default surface, that “tomorrow” is years away, and capable in-browser tools slow the migration by teaching people to expect AI where they already work. Volume is the advertising game. He does not expect OpenAI or Google to open their ad systems to third parties at equivalent rates, for the same reason Facebook pulled Audience Network.

That foreclosure is the opening. Hudson wants ZeroClick to be the Stripe of AI advertising: if you are a YC company monetizing a free tier, you should think of them the way you think of Stripe for subscriptions. The long tail, he says, is not small; it is broad. Someone will build a better travel assistant than any generalist app, using proprietary data that is not on the open web. Users will not open one mega-app for every job. Somebody has to provide the shared plumbing so those builders do not each have to cold-start advertisers and audiences at once.

Advertisers pay more the closer they get to conversion. Labenz’s worry is the other side of that truism: once the assistant earns money when the user takes a specific action, can you still believe it has you at the center? Hudson’s equilibrium is market discipline. Consumers will vote with their feet for AIs that look impartial. Paid sources can be extra offers a travel agent considers; if the agent feels captured, people leave. When he built consumer products, user trust sat at “the highest pinnacle.” Violate it and you lost.

As a rails provider, he does not want to be the moralist. He does not want Stripe or PayPal deciding which businesses may transact, or YouTube deciding which categories may monetize. Neutrality matters at the plumbing layer. Intervention, in his telling, should be limited to illegality and extremes rather than taste. Ranking, meanwhile, is “auction adjacent”: context gates who even enters, then expected value, in the Google style, multiplies predicted click quality by willingness to pay. He expects ZeroClick to evolve toward that same signal.

Labenz is only half convinced, and he has a scar. In the mortgage business he briefly saw originators handed pricing cards: a floor rate, then a bonus for closing a half point higher, another bonus for a full point. Salespeople were paid to extract. Google clicks in that category ran $500 to $750, affordable only if you wrung the customer. Map that onto AI and the financial-helper app that can bid the most is the one that most efficiently steers users to the most extractive lender, then takes a cut. Hudson calls it a real issue inherent to buying and selling, not unique to ads. Labenz notes that the mortgage version was largely fixed by making those spiffs illegal. Some failures do not self-correct.

Would Cursor ever be ad-supported, surfacing APIs and scrapers at the moment a developer needs them? Hudson says yes, and then widens the frame to all of SaaS. Most tools today need a five-figure contract and a human sales team to exist. He expects a wave of cheaper products distributed through contextual ads inside AI workflows, including categories people never type into Google. Labenz once mocked up a pricing page whose cheapest tier was “AI sales and support,” with humans as the upsell; Hudson liked it enough to consider it for ZeroClick’s own developer motion.

The case study he keeps returning to is Open Evidence, a ChatGPT-for-doctors product on an ad-supported model. Everyone else in the category sells into hospital systems on multi-year contracts. Open Evidence went to the physicians. Hudson says it is now used by about 40% of doctors every day. The audience is obvious and high value; the advertisers, pharma and device companies, are obvious and high value. They built both sides of that network themselves. For a research-paper chatbot, the matching is much less obvious, and the developer should not have to invent it. That is what common rails are for.

Labenz flags the inverted DTC drug ad: those commercials already end with “talk to your doctor,” and now the drug company may have talked to the doctor, or the doctor’s AI, first. Doctors cannot take cash per prescription; an AI, in a green-field legal regime, probably can. Hudson’s upside version is that TV drug ads are bad content almost nobody likes. If pharma can present a new option in the actual patient case, the doctor still filters, and the national “ask your doctor” campaign becomes unnecessary. Efficient channels, he argues, starve annoying ones, the same way Facebook’s measured conversions pull budget off unmeasurable TV.

Marketers are already cold-emailing Labenz the way they once sold SEO, except the pitch is now “we can get you into chatbot answers.” Hudson’s assessment is familiar: a handful of people on the frontier will really understand how to make content desirable to models, and thousands will take fees to pretend. Organic results will fill with slop, made worse if models prefer model-written text. Google spent decades using human clicks and bounces as quality signals; chat often loses that loop. For a marketer, he thinks paid context will dwarf the SEO game, the way SEM already dwarfs SEO. The current frenzy exists because there is no other lever. ZeroClick is trying to be that lever, in a market he describes as a void.

The intent moving into chat is also a step upstream of Google. People are not, in the main, typing a SKU for trail runners; they are talking about running, trips, nearby trails, and arriving at commercial intent later. Specific fulfillment still happens on Amazon. That upstream-ness is what makes non-commercial advertisers imaginable. Labenz asks what happens when a government wants to pay to shade how an AI talks about its country, including a Tiananmen query that does not currently show a sponsored Chinese-government link on Google. Hudson laughs: Labenz is “infinitely more creative,” and he has never come close to thinking about that use case.

Multimodal ads get a similar deferral. Gemini’s new image models make it easy to picture products dropped into a user’s living room. Hudson is intellectually interested, and he thinks advertiser catalogs would make those scenes *better*: a real couch you can buy, not Pinterest slop. ZeroClick is text-first. “A bit,” he then corrects, means three to six months. His closer is a live demo at zeroclick.ai and a request for developers to plug in. Labenz’s frame, from the top of the show, still hangs over it: building the future requires looking farther ahead than startup culture usually asks, and being sure the impact is one you can stand.