Who's Buying Ads Inside ChatGPT? | A Conversation with Nexad COO Harry Zhou

👦🏻 Interview: Crossing

🧑‍🎨 Layout: NCon

🚥 Earlier this year, ChatGPT began small-scale ad testing. Six months in, we were curious: what have the people who actually put budget into it actually seen?

Crossing spoke with one of the first to take the plunge on ChatGPT ads — Harry Zhou, co-founder and COO of Nexad. Nexad is a Silicon Valley-based AI-native marketing company backed by top-tier VC firms including a16z, Prosus, and Point72, and was among the earliest globally to build an AI-native chatbot advertising platform.

After ChatGPT officially launched ads, dozens of Chinese and US advertisers participated in testing through Nexad. These firsthand experiences were compiled into a 31-page ChatGPT Advertising White Paper. There will also be a closed-door exchange in Shanghai on July 7 — you can click "Read More" to register for the event and claim the white paper (invitation code: SZLK).

What's interesting is that this conversation started with the ChatGPT ad experience, but quickly moved to bigger questions: agentic commerce, vertical-domain agent harness, and — can AI actually make money through advertising?

🚥

🎙 Crossing

What made you decide to release the ChatGPT Advertising White Paper at this moment?

👨‍💻 Harry Zhou

Back in the ancient GPT Store era, before we had even raised funding, we were already inserting ads into many GPTs — probably among the earliest people in the world to do that, haha. So when OpenAI officially opened testing, we joined immediately. Over the past few months, we've invited dozens of Chinese and US advertisers to test with us, covering industries like AI SaaS, AI content platforms, DTC brands, mobile apps, and gaming. Everyone's been very interested, but the official self-serve platform is currently only open to US-based advertisers. With the elimination of the tens-of-thousands-of-dollars minimum spend requirement, smaller advertisers can now participate in testing too. So we compiled our experiences from the past six months into this white paper to share openly, and we're also opening test access through our own marketing agent platform Soku.ai, so this is no longer mysterious — anyone can participate self-serve.

🎙 Crossing

Where do ChatGPT ads mainly appear right now?

👨‍💻 Harry Zhou

Free users and Go users see sponsored ad cards below the main response when using ChatGPT, clearly separated from the content. Advertisers can set their own image and text creatives. Geographically, you can currently target users in the United States, Canada, Australia, and New Zealand. Japan, South Korea, Brazil, and Mexico are opening soon.

🎙 Crossing

What does the data look like so far?

👨‍💻 Harry Zhou

CTR is roughly between 1.5% and 4%, CPC mostly falls between $2-$4, with AI tools — which have high scene overlap with ChatGPT — on the lower end. CPM started around $60 early on, and we've seen many recently drop to around $20.

To be honest, we were initially pretty cooled off by these surface-level metrics, because they weren't dramatically better than traditional Google/Meta performance — but ChatGPT's user quality quickly became the surprise.

For example: many clients connect PostHog, Stripe, and other data into Soku.ai, so our agent can track more complex user behavior and long-term LTV than traditional agencies. One AI video product had decent front-end data on both Google and Meta — clicks, signups, first video generation all had volume. But once Soku.ai started running ChatGPT ads for it, the quality difference at deeper funnel stages became clear. Of registered users from Google/Meta, roughly 60-70% would complete their first generation, 30-40% would enter the pricing page, and final payment conversion typically landed in the single digits. ChatGPT users, however, had a higher same-day first-generation completion rate, and the proportion who generated continuously, returned multiple times, clicked on pricing, tried exporting watermark-free versions, and purchased generation credits was roughly two to three times that of Google/Meta. This difference is critical for AI video products! Because what they fear most isn't lack of signups — it's hordes of users who come in, try once, and leave.

We've observed similar patterns across multiple AI and e-commerce clients (including when factoring in longer-term renewal and repurchase rates). Interestingly, I saw that Adobe just published data this month showing that AI-driven traffic on their clients' retail sites converts 54% higher than non-AI traffic — unprecedented. All of this gives us strong confidence in the user quality that ChatGPT ads can deliver.

🎙 Crossing

Why is the user quality from ChatGPT so high?

👨‍💻 Harry Zhou

Google matches keywords, Meta matches audiences, ChatGPT matches a decision someone is actively making. Reaching people at the right timing is crucial.

In search, users compress their needs into a few words — "noise cancelling headphones" or "best monitor for gaming and work." Advertisers know they're looking for headphones or monitors, but not why they're buying, their budget, what they've already ruled out, or what's most important to them. Meta is even further upstream, inferring what someone might buy from interests, behaviors, and lookalike audiences.

In ChatGPT, users might describe their full situation: flying long-haul next month, want headphones that are lightweight, have good noise cancellation, battery lasts round-trip, budget under $300. The model keeps asking follow-ups, turning fuzzy needs into a set of specific constraints. By the time the ad appears, the user has often already completed a round of research, comparison, and elimination — at a "high-intent moment" very close to action.

ChatGPT Ads has a unique targeting method called "Context Hints." Simply put, instead of just filling in keywords or selecting audiences, advertisers use natural language to describe "what kinds of conversation scenarios my product fits into," letting OpenAI's ad system understand what decision contexts this product belongs in, and whether to show it based on the entire conversation — something Google and Meta's rule engines can't do.

ChatGPT Ads makes real purchase intent within AI conversations targetable, trackable, and scalable for the first time. For high-ticket products with longer decision chains, this may be one of the rare quality new growth channels to emerge in recent years.

🎙 Crossing

Many people are looking forward to agentic commerce — AI directly making decisions and completing purchases for people. Where does it stand in reality?

👨‍💻 Harry Zhou

I'd split it into two parts: AI helping you choose, and AI buying for you. We've already covered the former; the latter is actually completing the transaction — money paid, goods delivered, after-sales handled if something goes wrong. On the user psychology side, when it actually comes to payment, returns, after-sales, privacy, account security, and so on, people become noticeably more cautious. Walmart tested instant checkout on ChatGPT, letting users complete purchases within the conversation, but the conversion rate for in-conversation checkout was only about one-third of sending users back to Walmart's own website — showing that trust is still very low.

On the platform and merchant side, I think a major difficulty is how responsibility actually lands. When OpenAI and Stripe pushed the Agentic Commerce Protocol, a key design was that even if checkout happens within the ChatGPT interface, the merchant of record remains the merchant. AI mainly coordinates and initiates the transaction; fulfillment, returns, customer service, chargebacks, tax, payment risk control — AI can't magically assume these responsibilities for merchants.

So what's working first in agentic commerce is "AI helping you narrow down candidates." This also explains why advertising has taken off first.

🎙 Crossing

Back to ChatGPT ads specifically — how do advertisers actually run and manage these accounts?

👨‍💻 Harry Zhou

Advertisers we work with mostly go fully managed through Soku.ai, Nexad's self-developed marketing agent. Is the term "digital employee" a bit old-fashioned now? Haha. But it genuinely feels like a "real digital employee" to me. Just sticking to the advertising use case: from thinking to proactively taking automated action, it can manage dozens of campaigns across ChatGPT/Google/Meta/TikTok for an advertiser end-to-end, on a month-long time scale. It checks account status daily, spots event stream breaks, creative fatigue, budget spending too fast, learning phases getting interrupted, or abnormal conversion in certain regions, and surfaces the issues. Low-risk actions it can take automatically; for anything involving spending, scaling, shutting off, or budget changes, it proposes a plan with reasoning first, gets human confirmation, then executes itself. It's essentially a growth operations colleague that's been battle-tested here and has been on the job long-term.

This is especially important in the early days of ChatGPT Ads. It's not as mature as Google or Meta yet — many pieces are still changing: which markets can be targeted, what format creatives need, how creatives get approved, how conversion events are passed back, what counts as a valid conversion — all require continuous monitoring and operation.

🎙 Crossing

Why specifically emphasize the month-long time scale?

👨‍💻 Harry Zhou

Because managing ads is a classic long-horizon agent task. It naturally fluctuates daily, learns weekly, and gets reviewed monthly.

Running an account well isn't about adjusting a bid today and swapping an image tomorrow. You need to look at it on a month-long cycle: which campaign is still in learning phase and shouldn't be touched; which one spiked yesterday — real momentum or noise; where to shift budget this week, which to kill; whether to refresh fatigued creatives; knowing what fluctuation a given client can tolerate, which actions need pre-approval; and remembering why you didn't kill a certain ad three weeks ago. A skilled growth team can only watch a few dozen ads simultaneously before hitting their limit, and it's very dependent on intuition. This information is scattered across ad platforms, GA4, PostHog, Stripe, CRM, creative libraries, and client communications. Human growth teams hold it in their heads, or scatter it across Slack, Lark, Notion. If an agent is going to actually take over part of the work, it can't start from zero each time — it needs to accumulate this context so the next round of decisions can build on it.

🎙 Crossing

What's the hard part in building an agent like this?

👨‍💻 Harry Zhou

Agents "quietly degrade." The longer the task horizon, the bigger the silent degradation challenge.

Visible failures are easy to handle — bugs or obviously wrong results, humans catch quickly. The trouble is model judgment slowly drifting off course over time: audiences gradually drifting to people who shouldn't be targeted, bids getting more and more conservative, an ad that should have been killed long ago still running, conversion events broken in a region but budget still spending. Each step looks justified, but stacked together, money is quietly leaking.

This isn't just our problem building marketing agents — it's a universal challenge for long-horizon agents. For example, METR proposed a "time horizon" metric measuring how long a model can stably complete tasks. A key distinction inside it: the task length a model can "occasionally pull off" versus what it can "reliably complete with high confidence" — these aren't the same order of magnitude. Doing it once doesn't mean it can take over long-term. We've also seen interesting public experiments — Anthropic tried having an agent run a vending machine, which looked fine short-term, but over time entered "tangential meltdown loops" getting increasingly chaotic.

🎙 Crossing

So how do you prevent it from degrading?

👨‍💻 Harry Zhou

Our team wrote an agent harness lessons-learned post in early April (called "Three Months of Pitfalls: Our Agent Harness Practices and Reflections", searchable online), and one core takeaway was: "Constraints are only real constraints when they're machine-executable." Writing them in documentation or prompts isn't enough.

In real accounts, many critical rules can't be left to the model's discretion. Things like how time windows are calculated, how conversions are counted, budget caps, which actions can be automated, which require human confirmation — these need to become hard constraints in the system. Otherwise the model can easily look like it's analyzing carefully while actually using the wrong metric, or crossing boundaries it shouldn't cross.

So a large part of our work is designing constraint systems: what it can do, what it can't, where it must stop. Read-only diagnosis is one category, low-risk operations another, and irreversible actions involving spending, killing campaigns, scaling, or budget changes yet another. Different risk levels correspond to different permissions, approvals, and logging.

I think this applies to all long-horizon agents. For an agent to truly take over work, being smart isn't enough. It needs to be governed by a system that can correct it.

🎙 Crossing

Back to a bigger question: can AI actually make money through advertising?

👨‍💻 Harry Zhou

Yes, but I wouldn't think of it as "selling a few more ads in a chat box." The past internet turned attention into a business; AI is closer to turning decisions and execution into a business.

Looking at the value chain: chips, cloud, and inference infrastructure get paid first, because they sell scarce compute. The model layer sells capability — subscriptions, APIs, enterprise contracts will all be huge, but capabilities will keep getting cheaper, and open source will catch up. Further up, applications and agents are closer to real business: they know what users want, can connect to data, permissions, and result feedback. Advertising, transaction fees, qualified leads, outcome-based billing, managed services — they're all essentially charging for things closer to results.

Why did advertising appear first? Because it's the easiest way to monetize "a user is making an economically valuable decision." The value of ChatGPT Ads is here: purchase intent within AI conversations can now, for the first time, be targeted, tracked, and optimized. But it's just an early form. Once agents truly start doing things for people, money will continue flowing toward transactions, outcomes, and managed services.

So I don't think AI will simply end up as "advertising model" or "subscription model." The truly long-term, compounding money will flow to whoever is closest to decisions and actions. Whoever can stand beside the user at the moment of decision, and actually get the follow-through done, will be closest to the richest returns in AI commercialization.

The full Nexad ChatGPT Advertising White Paper and registration for the July 7 Shanghai closed-door exchange can be accessed by clicking "Read More" below.