In-AI Advertising: Better Answers for Users, Big Questions for Society, with ZeroClick's Ryan Hudson
Summary
ZeroClick’s $55 million bet is that “paid inference time” can make free AI applications viable while becoming the Stripe-like monetization rail for independent developers. Ryan Hudson says advertising can already cover inference costs in some use cases, with margins potentially widening as models get cheaper. The strategic prize is an ecosystem of thousands or millions of specialized apps outside the largest AI platforms.
The product inserts advertiser information as optional context, then leaves the application’s AI to decide whether it improves the answer. Through an MCP server, ZeroClick matches a query against AI-generated campaigns built from landing pages, catalogs, prices, or service databases; inclusion and outbound links remain measurable. Early click-through rates in ZeroClick’s Pi GPT reference implementation were “insanely high,” though Hudson cautions that the audience and data are still too limited to generalize.
Hudson’s bull case is that contextual ads can improve answers, fund publishers, and give startups access to markets that entrenched organic rankings would otherwise close. His sharpest example came from Pi Adblock’s visual mode: when it removed sponsored Google product results, users complained, “You’re deleting the best answer from that search.” The broader ambition is to restore economic value to an internet whose banner model is deteriorating without consolidating discovery inside three AI giants.
The cleanest business begins with Google-like high-intent searches, while Facebook-like discovery requires personalization and carries greater incentive risk. Hudson initially rejected the idea that AI ads will maximize engagement because search advertising monetizes relevant decisions, not raw time on site. After Erik Torenberg raised proactive companions such as Tolan, Hudson conceded that ad-funded apps might push commercial suggestions and enable more “parasitic social relationship” products—even if subscriptions create similar retention incentives.
User trust is the proposed control mechanism, but auction economics can reward the actors most willing to extract from users. Hudson expects consumers to “vote with their feet” against agents that put paid consideration above their interests, and he wants ZeroClick to remain neutral infrastructure except around illegality or extremes. Nathan Labenz’s mortgage counterexample—clicks reaching roughly $50 to $7,500 while originators were rewarded for charging higher rates—showed why willingness to pay can signal exploitation rather than quality.
AI-native advertising could collapse distribution costs across software, coding tools, medicine, research, and browser extensions. Hudson points to OpenEvidence’s reported ad-supported reach of roughly 40% of doctors as evidence that a direct-to-user model can outcompete five-figure enterprise sales. Browser tools are especially attractive because a narrowly useful assistant can appear once a month at exactly the right moment, “get user habit for free,” and disappear otherwise.
The hardest risks remain largely unanswered: political influence, medical conflicts, AI-generated persuasion, and platforms judging their own outputs. When Erik Torenberg asked whether governments might pay to shape how an AI describes their country, Nathan Labenz admitted he had “never come close to thinking about this particular use case.” That unresolved exchange reinforces the episode’s central investor tension: the technical and commercial path looks plausible, but social safeguards are lagging a market moving on “three months or six months” timelines.
Deep dive
1. ZeroClick wants to make free AI economically complete
The show’s framing put real weight behind the experiment: ZeroClick had announced a $55 million raise, while Hudson previously founded Honey, which sold to PayPal for $4 billion. This is not pitched as an ad widget but as infrastructure for a new application economy.
Today’s default is to charge roughly $20 a month while throttling free users. Hudson’s alternative is “paid inference time or reasoning time consideration of advertiser content,” giving an AI one more information source while helping finance a functional free tier for billions of people.
ZeroClick emerged from Pi Adblock, which has a couple million users and rewards people for opting into advertising they control. Its contextual matching was designed so profiling remained inside the browser and “never leaves it in any form that’s usable”; the team realized the same architecture mapped naturally onto AI.
Hudson’s commercial aspiration is explicit: within six months or a year, a startup deciding how to monetize a free tier should think of ZeroClick as it thinks of Stripe for subscriptions. “You don’t need to build them yourself.” The inherited company ethos is equally blunt: “We can make ads good actually.”
2. AI advertising could rebuild the publisher bargain
Erik Torenberg’s opening concern was that publishers are being aggregated twice: AI systems train on or retrieve their work, then answer users directly while offering the source, at best, a footnote and a link. The advertising-supported open internet now faces both deteriorating banner economics and disappearing visits.
Hudson’s sequencing matters: first create money inside the AI answer flow, then applications can attribute value to publishers considered for that answer—or across a user’s broader activity—and distribute proceeds. Without an economic engine, there is little revenue available to redesign the bargain around high-quality content.
Pay-for-access approaches from companies such as TollBit and Cloudflare were discussed: Hudson described TollBit and Cloudflare throttling access if content was not paid for, while Erik said that approach “kind of makes sense” but likened it to de-indexing a site from Google. Hudson expects a mixture of subscriptions, advertising, and publisher reallocation to emerge over time.
ZeroClick does not intend to prescribe how developers share revenue. Hudson instead expects market pressure to reward useful sources, while admitting maturity will take time: “Everybody acknowledges that this is a problem,” but the immediate task is ensuring enough value exists to fund any solution.
3. The alternative to an open long tail is an AI fail state
Hudson hopes discovery fragments across thousands or millions of developers rather than collapsing into ChatGPT, Gemini, and one other platform that learns everything about each user. Three monolithic browsing destinations would be “a fail state” resembling earlier platforms’ foreclosure of competition.
His cautionary history came from Facebook Audience Network. Facebook once exported its targeting advantage to third-party publishers, then pulled that capability into its walled garden; the move improved its own position while depriving potential social competitors of comparable monetization.
The ad industry compounded the problem with creepy tracking, privacy violations, and formats that trained users and platforms to resist advertising. ZeroClick’s opportunity is therefore not only technical matching—it is common monetization infrastructure that prevents the largest AI companies from owning both user intent and the only efficient way to sell against it.
Hudson expects major AI platforms eventually to build native ad systems during what he compared to OpenAI’s current “Don’t be evil” or Google’s pre-ads phase. He doubts they will offer independent developers equivalent economics: opening their systems creates operational headaches and weakens strategic control, leaving room for an outside “Stripe” to aggregate the broader market.
4. Contextual ads have already proved they can be the better answer
Pi Adblock includes a visual mode that shows ads being “zapped off the screen.” On Google product searches, users sometimes objected: “Stop doing that. You’re deleting the best answer from that search.” The sponsored result was still an ad, but its tight alignment with intent made it more useful than the organic list.
Hudson extends that observation directly to AI. A current agent may rely on its training, run a couple of Bing searches, scan perhaps the top 10 organic results, and synthesize an answer; adding five paid results, then letting the model exclude anything irrelevant, should sometimes produce a better-informed response.
Search advertising also prevents inertia from deciding every market. Purely organic rankings can take years, so paid consideration lets a startup “inject yourself into the conversation” beside incumbents that have accumulated authority simply by existing longer.
Labenz broadened the upside: targeting lets businesses relevant to one-tenth of 1% of people find that niche globally, turning passion projects into viable livelihoods. Better matching expands specialization, supports creators, and replaces the old broadcast diet of repetitive Cap’n Crunch and Ninja Turtles commercials with information people might actually value.
5. Modern advertising earned praise—but not a blank check
Meta’s lack of a voluntary ad-free Instagram tier prompted competing explanations. Labenz wondered whether removing the wealthiest 1% would degrade the audience advertisers most want; Hudson’s simpler answer was that Meta already has “a phenomenal business,” faces no need to disrupt it, and might trigger backlash by charging for better features.
An EU-forced option might cost “20-some dollars” a month because that approximates the value Meta extracts from a user. Yet Hudson argued Instagram advertising is often additive: its targeting and creative are good enough that removing ads might not create something consumers would pay to receive.
Search and social advertising also democratized distribution, while Hudson said improved quality controls had largely resolved an earlier era when visiting a website could hijack a Windows machine. He saw that world firsthand at OpenX, managing ad and traffic quality while malicious actors tried to distribute code through real-time bidding systems.
The strongest downside is structural: when revenue scales with time on site, products learn to exploit rage, variable rewards, and cognitive vulnerabilities. Erik’s worry was that generative AI can now optimize this dynamic for an audience of one, producing persuasion more intimate than any social feed assembled from human posts.
6. Search economics do not erase the AI engagement trap
Hudson initially said he was “not at all” worried about ZeroClick creating addiction, deliberately overstating before softening. His analogy was Google search: the platform wants to appear at commercially important decisions, such as choosing wedding clothes, rather than manufacture endless low-value searches or banner impressions.
Erik’s pushback—worth keeping—was that AI shape-shifts between shopping assistant, philosophical partner, and confidant. Tolan, marketed as an alien best friend, sent contextual notifications such as asking how a recently mentioned talk had gone; the product was no longer waiting for a deliberate query.
Hudson responded that subscriptions do not solve this. A product manager still tracks engagement because daily use correlates with renewal, so a paid companion may pursue the same notifications and retention loops. Ads may be blamed for a “parasitic social relationship” when the underlying product already wants the user to return.
He nevertheless conceded a genuine divergence: ad-supported apps might seek more commercial “at bats” and push suggestions—perhaps a local hotel deal because someone seems to need a weekend retreat. That can be useful or manipulative depending on context. “It’s probably a fine balance,” and Hudson had not thought deeply about all the categories.
7. ZeroClick is beginning with intent, not interruption
Labenz’s market shorthand was Google for needs users already recognize and Facebook for products they do not know exist. ZeroClick is “highly optimized” for the first category: it vectorizes immediate context and matches relevant campaigns rather than throwing a wild card into an unrelated conversation.
A Facebook-style system would require durable user context. Pi Adblock offers a possible privacy-preserving architecture because the browser can hold the profile locally, while early ZeroClick implementations remain “super context driven” rather than relying on that personalization.
Hudson described Facebook’s engine as a vectorized profile matched against clusters of known converters. Greater reach means moving farther from that conversion cluster. Its continuing financial strength follows from better matching, deep campaign inventory, and ad budgets migrating toward whichever channel can demonstrate more conversions.
His macro view is almost zero-sum: advertisers devote a relatively fixed share of company spending or GDP to promotion. If AI provides a more measurable path to intent, budgets can move from Google, television, or harder-to-measure channels without requiring total advertising expenditure to grow.
8. The MCP integration turns advertiser data into optional context
ZeroClick’s Pi GPT reference design proved that a custom GPT could consider paid sources, include trackable links, and generate measurable advertiser outcomes. Click-through was “insanely high,” which Hudson reads as evidence that users saw the result as useful—not merely visible—though he stressed that the early audience does not justify broad conclusions.
Developers can now connect through an MCP server. The effective instruction is: here is additional information; include it only if useful, preserve these links if used, and report whether it appeared. Developers can tune that behavior to fit their product rather than accepting a single universal ad presentation.
On the advertiser side, ZeroClick ingests landing pages, product catalogs, prices, or databases of service professionals. AI summarizes that material into campaigns, maps it into vector space, and matches incoming keyword searches without requiring advertisers to design the campaign mechanics themselves.
Generating the paid context internally also limits prompt-injection risk. Instead of accepting arbitrary advertiser instructions—including classic white-text attacks—the platform controls the transformation. At runtime, paid retrieval can run parallel to organic search, while caching and contained auctions avoid the sprawling network calls involved in open real-time bidding.
9. Specialized apps can win where one universal chatbot cannot
Hudson rejects conversational chat as AI’s only interface. Just as third parties build better apps than Apple and better websites than Google, focused developers can combine a specific audience, proprietary data, and purpose-built interaction to outperform a frontier model instructed to adopt a different personality.
His best specimen was shopping in the browser. A service like PayPal Honey has roughly a decade of Amazon price history, so an assistant could activate when a user hovers over a price and say either “this is $20 cheaper than it’s ever been” or that the item is overpriced and alternatives deserve consideration.
The value lies in eliminating translation between contexts: users should not need to copy a URL into ChatGPT or explain what they are viewing. Browsers, email, and corporate workflows can initiate the right conversation where the underlying activity already happens.
Hudson also predicts Apple silicon will perform local language-model work comparable to today’s models within “the next year or two.” Falling inference costs, proprietary vertical data, and on-device privacy could support broad application diversity; he believes ad revenue has already crossed inference cost, with infrastructure costs falling while monetization improves.
10. Trust is the constraint auctions cannot price directly
Erik’s central question was who the agent ultimately serves as monetization moves closer to conversion. Advertisers pay more near an actual purchase, but an assistant receiving a share of that value can begin to look like the seller’s agent rather than the user’s.
Hudson’s proposed equilibrium is exit: consumers will “vote with their feet” toward products that remain impartial and respect their priorities. He wants a travel agent to consider paid offers while still finding the best deals; once users sense it is “stepping on the scale,” trust disappears and a competitor can take its place.
He considers user trust the “highest pinnacle” when building consumer products, yet does not want ZeroClick dictating every developer’s choices. As infrastructure, it should resemble Stripe or PayPal: neutral plumbing that intervenes primarily around illegality or other extremes, not a platform making continuous moral judgments.
The auction itself will resemble Google’s. Context determines eligibility, then willingness to pay and observed clicking contribute to expected value. Hudson called the current implementation “auction adjacent,” with bids partly managed toward campaign outcomes, but expects a more mature ranking system in which price and response jointly signal quality.
11. A high bid can signal extraction rather than quality
Labenz attacked that premise with the mortgage market. Originators once used pricing cards that paid salespeople more for closing borrowers half a point or a full point above the minimum; Google Search clicks reportedly reached roughly $50 to $7,500 because lenders extracting more value could afford to bid more.
The second-order danger is sharper in AI: a financial-advice app earning the richest referral revenue can acquire users most aggressively and become dominant. Its apparent product quality may therefore reflect skill at steering customers toward high-extraction providers rather than skill at protecting their finances.
Hudson accepted that this was “a very valid thing to think about.” Some domains may require subscriptions, brokerage subsidies, or another model entirely; advertising need not appear in every AI product. He said government regulation can address some such failures, noting that the mortgage incentive described is now illegal in his view and recalling similar stock-brokerage spiffs.
Labenz suggested consumers may need an AI that audits whether other AIs are aligned with them, then noted the hall of mirrors: GPT-5 evaluations use an LLM as judge. If models learn that other models reward pretentious creative writing, they may optimize for machine approval that leaves humans asking, “What the fuck is that?”
12. Contextual distribution could rewrite software and medicine
Coding assistants such as Cursor could recommend scraping APIs or other technical services exactly when a developer encounters the need. Hudson widened the thesis to SaaS: products often require contracts above roughly $10,000 a year because human sales and support impose a floor on otherwise low-marginal-cost software.
An AI advertising layer could distribute cheaper tools without that sales machine. Erik’s imagined pricing page captured the transition: an entry tier with AI sales and support, a middle tier for human sales, and a premium tier for human sales and support.
Hudson cited OpenEvidence as the key case study: an ad-supported “ChatGPT for doctors” reportedly reaches around 40% of doctors daily, while rivals sell multiyear, high-value contracts through hospital systems. A clearly defined, valuable audience can support direct adoption funded by pharma and medical-device advertisers.
Erik’s concern was that drug companies could reach the doctor before a patient asks about treatment, and an AI intermediary might receive incentives no human doctor could legally accept. Hudson’s upside case is that contextual presentation to doctors could replace television ads that annoy millions; clinical judgment would remain the filter, while awareness arrives at the relevant patient case.
13. Browser policy and social governance will decide who captures the upside
Hudson sees extensions—not entirely new browsers—as the richest application layer. A narrowly useful tool can surface once a month during one specific task, “get user habit for free,” and remain invisible otherwise. That is more natural than sending eight daily notifications to manufacture engagement.
Chrome complicates the opportunity. Google ended paid extensions, requires distribution through its store, and applies a single-purpose policy that Hudson believes chokes off innovation across a browser with roughly 70% to 80% share. His preferred remedy is an open platform; selling Chrome to another AI company could merely install a new owner with the same incentive to foreclose competitors.
AI optimization will reproduce SEO’s asymmetry: perhaps 10 world-class practitioners understand how to shape content for model consumption, while thousands sell weaker services. AI-generated “organic” slop may also make it harder to distinguish authoritative human-created content from mass-produced material, given the loss of Google’s decades of click-and-bounce feedback. Hudson expects paid context to matter far more, especially as chat captures upstream exploration rather than only explicit product searches.
The unresolved frontier is influence beyond commerce. When Erik Torenberg asked whether foreign governments might pay to shape answers about their countries, Nathan Labenz admitted he had “never come close to thinking about this particular use case.” Nathan’s larger warning is the gap between breakneck execution—996 or 12-hour, six-day schedules—and reflection on consequences, from SB53 policy disputes to visual ads placing real purchasable furniture inside a user’s home within perhaps “three months or six months.”