Mara ReyesChatGPT can know far more about a person’s problem than a search box and still offer worse advertising inventory. Extra context can improve a recommendation. But a conversation may begin with drafting, tutoring, or troubleshooting, and every turn consumes compute before an ad appears. ChatGPT can make the interface richer while making the commercial unit poorer.
Key takeaways
- OpenAI’s projected $102 billion in ad revenue by 2030 requires ChatGPT to become a major commerce gateway, not merely add sponsored messages to conversations.
- Chat creates abundant engagement but limited monetizable inventory: drafting, tutoring, and troubleshooting sessions consume inference without necessarily bringing users near a transaction.
- The strongest conversational-ad formats establish an explicit commercial boundary, such as a brand agent, product comparison, or merchant handoff, instead of blending sponsorship into an assistant’s general advice.
- Inference costs make commercial intent per inference dollar the key metric; cheaper serving helps, but it also lowers barriers for rivals and may weaken frontier-model margins.
- Advertising can complement subscriptions, enterprise contracts, APIs, and commerce fees, but it is unlikely in the near term to underwrite broad chatbot usage on search-like economic terms.
$102 billion describes a gateway that does not yet exist
OpenAI’s internal model projected approximately $2.4 billion in advertising revenue in 2026, nearly $11 billion in 2027, and $102 billion in 2030. At that final figure, ads would supply 36% of total projected revenue. A side business cannot deliver that share; OpenAI would have to turn ChatGPT into a commercial gateway broad enough to support a frontier lab’s capital structure.
To reach $102 billion, OpenAI would have to more than quadruple ad revenue in the first year, then multiply it again through 2030. ChatGPT would need to create commercial inventory that does not yet exist at the necessary scale: repeatable moments when a user is close enough to a purchase, booking, subscription, or other transaction that an advertiser can value the placement and measure the result.
OpenAI first offered chatbot ads to dozens of advertisers with commitments below $1 million and charged for views rather than clicks. It later added cost-per-click buying with bids between $3 and $5, then introduced a U.S. Ads Manager beta for smaller advertisers. As OpenAI moves from impressions to clicks to self-service buying, it is still establishing what an assistant ad is worth.
eMarketer estimated less than $1 billion in chatbot advertising revenue for 2026 and placed the ceiling for the U.S. chatbot-ad market at $5.41 billion. That estimate and OpenAI’s $102 billion target are not like-for-like: one describes a U.S. market, while the other is a company revenue target. Together, the figures show what OpenAI’s target requires—assistants must rapidly create a market far larger than the one analysts can measure now.
Search begins near a decision; chat often begins with work
Search advertising captures a compact declaration of intent. A person types a product, destination, service, or problem into a box; the system classifies the query; advertisers bid through an established auction; and the result page sends the user toward merchants equipped to complete the transaction. Not every search is commercial, but the commercial subset is legible, repeatable, and surrounded by attribution infrastructure.
ChatGPT can hold far more context than a query box, but context is not commercial intent. A twenty-turn debugging session contains abundant information and no necessary purchase. A short shopping query can contain little information and nearly all purchase intent. A platform that counts both as engagement hides the variable that pays.
Google approaches this problem by attaching AI answers to machinery it already owns. Its planned search and shopping ads in AI Overviews drew from advertisers’ existing campaigns, and ads in AI Mode resemble citations while carrying a sponsored label. The model generates a new interface, but the merchant relationships, campaign data, auction system, and transaction paths remain underneath it.
Google strengthens that mechanism when AI Mode moves from answer to action. It can link to and interact with services including Instacart, Canva, and YouTube Music. In those moments, the assistant becomes a work surface that hands the user into a service capable of fulfilling the task.
An advertiser pays not for a conversational turn but for proximity to a decision the system can help complete. The assistant’s commercial value therefore concentrates in the workflow around the answer—merchant data, product availability, identity, payment, fulfillment, and attribution—not in the answer alone.
ChatGPT can inflate apparent inventory by producing more turns without producing more buying intent. The metric rises while the scarce event—the decision an advertiser can value—stays fixed. Search built its margins around queries that could be sorted into commercial and noncommercial classes. ChatGPT cannot manufacture more of the first category merely by making the second category longer.
The formats that work put the sponsor inside a boundary
Snap’s AI Sponsored Snaps let users converse with brand-specific agents for product recommendations. The sponsor is explicit, the subject is constrained, and the interaction begins inside a commercial context. A user who opens a brand agent has already crossed a threshold that a general-purpose assistant session may never approach.
Snap’s format proves that a commercial conversation can carry sponsorship. It does not prove that broad assistant usage can absorb search-scale ad load without changing the product users believe they are using.
ChatGPT faces a structural trust problem. A search page presents multiple links, and a sponsored result remains one option among others. An assistant synthesizes an answer in a single voice. When sponsorship enters that voice, users have more trouble separating relevance from influence, particularly outside an explicit shopping task. Because the assistant knows why the user is there, an irrelevant or intrusive ad looks less like a targeting error and more like a decision.
Explicit boundaries contain that risk. A merchant handoff makes the commercial role visible. A brand-specific agent identifies whose interests it represents. A product comparison gives the placement a task to serve. A generalized sponsored answer removes those boundaries, asking one response to act as adviser, publisher, auction surface, and salesperson. Search separated those roles on a page, while conversation compresses them into a sentence.
Cheaper inference lowers both the bill and the moat
Conversational advertising is also an inference-economics problem. OpenAI and Anthropic told investors that inference costs exceeded half of revenue in their projections. Their burden does not end when they train a frontier model because they keep paying to serve it.
Each long answer, follow-up, tool call, or agentic action can add work before the system produces an impression or conversion. The bill arrives in GPU time, cooling, fiber, and electricity inside data centers whose capacity is increasingly secured as contracted megawatts, not as an abstract cloud. OpenAI must cover those variable serving costs while leaving enough margin to support the fixed infrastructure behind them.
Google can place several monetizable spots around one search query. ChatGPT may need several inference steps before it reaches a commercial surface, if it ever does. That makes commercial intent per inference dollar—not ads per user or session—the critical ratio for OpenAI.
OpenAI engineers reportedly found a technique that could more than halve inference costs. Lower serving costs would let OpenAI expand usage and make previously uneconomic commercial interactions viable. Cheaper answers can create surfaces that were once too expensive to serve.
Yet lower costs would also let rivals offer capable models more cheaply. Users and businesses could route work toward the least expensive adequate system. Perplexity’s Computer feature splits tasks between local and cloud models to preserve private data on-device and maximize token efficiency, cutting inference per useful task.
Cheaper inference can enlarge the assistant market while making it harder for any one frontier lab to preserve the margins assumed by its infrastructure commitments. Advertising then rewards companies that control distribution and transactions, not the lab that merely produced an answer.
Advertising belongs in the stack, not under the whole structure
OpenAI projected that advertising could help double consumer ChatGPT revenue to $17 billion in 2026. Its model treats ads as one layer in a consumer revenue stack. Subscriptions charge for sustained access and higher usage. Enterprise contracts charge for organizational deployment. APIs charge for consumption inside other products. Commerce fees and targeted ads capture value when the assistant influences or completes a transaction.
Users reveal why those layers must remain distinct. A drafting session may justify a subscription but no advertisement. A corporate workflow may support a contract without exposing an employee to a sponsored answer. A shopping session may support a merchant handoff, a click payment, or a transaction fee.
Amazon starts with an advantage because its chatbot-style assistant sits on top of merchant relationships, product data, customer identity, and places where purchases already occur. AI can improve production and targeting without inventing the transaction system beneath it.
OpenAI’s forecast matters because it shows the system the company is trying to build: a gateway with enough control over discovery and action to claim part of the commerce moving through it. To reach $102 billion, OpenAI must monetize broad uses that generate trust and engagement without necessarily producing a transaction, even though the infrastructure serves both kinds of use at real cost.
ChatGPT can become a storefront. The $102 billion bet is that every GPU rack behind it can also become a billboard.
Frequently asked questions
Why can’t ChatGPT monetize conversations the way Google monetizes searches?
Search queries often begin close to a purchase and plug into mature auctions, merchant systems, and attribution. Many chatbot sessions begin as work—such as drafting or debugging—and can require multiple costly turns without producing commercial intent.
What kinds of ads are most likely to work in conversational AI?
Ads work best inside clearly commercial contexts, including brand-specific agents, product comparisons, shopping tasks, and merchant handoffs. These formats make the sponsor’s role visible and connect the interaction to a measurable action.
Why is putting ads directly into assistant answers risky?
An assistant speaks in a synthesized, authoritative voice, making paid influence harder to distinguish from neutral advice. Outside an explicit shopping task, sponsorship can therefore damage trust more than a labeled ad presented among multiple search results.
Would lower inference costs solve the chatbot advertising problem?
Lower costs could make more assistant interactions economically viable, but they cannot create purchase intent where none exists. They also enable competitors to serve capable models cheaply, shifting advantage toward platforms that control distribution, merchant relationships, and transactions.
Can advertising still become a meaningful ChatGPT business?
Yes, particularly in the minority of sessions involving shopping, bookings, subscriptions, or other transactions. The piece argues that ads should remain one layer alongside subscriptions, enterprise contracts, APIs, and commerce fees rather than finance the entire system.