AI Chatbot Ads: What Works and What Risks Trust

Why are AI chatbots starting to show ads?

Chatbot responses have a real operating cost. Every conversation needs compute, and longer or more complex interactions can require even more. Subscription fees, enterprise deals, and investor money can cover those costs - but they don't solve monetization for every product. Advertising is one option, especially for free services with large audiences.

There's also a practical imbalance between companies. A big tech company can subsidize an AI product with other revenue streams. A smaller niche chatbot app usually has fewer choices. For a niche product, ads can help pay for infrastructure and ongoing development without forcing every user into a subscription. But adding ads also adds overhead: someone has to manage advertiser relationships, targeting, measurement, policy enforcement, and complaints.

The shift from speculation to reality is happening. In a WIRED snapshot, OpenAI said that after launching ChatGPT ads in February, more than 100,000 advertisers were already using them. Google also expanded advertising in its AI Mode search experience to more countries. Those numbers point to real commercial interest - but they don't prove the experience is "settled," or that users actually like it.

Smaller ad-tech providers are pushing the same idea across specialist apps. Koah Labs, for example, said it served ads in more than 200 apps and worked with brands including Hims and UrbanStems. Its chief executive, Nic Baird, reported that around 3% of users clicked ads across its network - framing that as comparable to Google search ads. It's a useful early signal, but it's not a controlled, independent measure of long-term performance or user satisfaction.

There's another constraint, too: if ad revenue is the goal, the platform has incentives to show more ads or make targeting tighter. Either move can undermine why users came to a conversational assistant in the first place. So the key product question isn't "can ads make money?" It's whether that revenue is worth the trust and attention it costs.

For AI companies, the answer depends on the business model and audience. A paid assistant can reasonably avoid ads. A free consumer product may need a different revenue stream. But the choice has to be explicit: ads can't be treated as a harmless add-on if they change what data gets used, what users see, and how recommendations get interpreted.

What makes chatbot ads effective?

Early experiments point to two possible advantages over traditional search ads: more context and more room for flexible creative. In search, you might type a short query like "help with stress." In a chatbot, the system can see a longer prompt and multiple turns describing what the person has tried, what they want, and what constraints matter. That can help an ad system choose something that fits the moment better than a simple keyword match.

Koah's Baird described spa ads showing up after conversations about stress as an example of this broader contextual matching. But "relevant" doesn't have to mean "based on intimate personal details." An ad can respond to the topic without making a leap into a diagnosis, identifying a vulnerable person, or dragging conversation history into a deeper profile. Those guardrails are design choices - not technical inevitabilities.

Creative also has to feel native in a conversational interface. Baird said ads that used some AI-generated copy got twice as many clicks as ads written entirely by people on Koah's network. That's a company-reported result, not an independently verified rule. One plausible explanation is that AI-generated copy may better match the tone and formatting of an AI response. But that can also blur the line between sponsored content and the assistant's actual answer.

That's why disclosure and placement aren't just compliance checkboxes - they affect performance. A sponsored recommendation should be recognizable before a user acts on it. If the label is easy to miss, a click may just mean confusion. Short-term click-through numbers can reward that confusion, but long-term trust probably won't.

AI-generated creative also needs a reliable source of truth. If the system invents product claims, an ad can become misleading even when the advertiser provided accurate information elsewhere. Koah said its ads were grounded in product details and imagery supplied by advertisers. That approach can reduce the odds of fabricated claims, but it doesn't eliminate the need for review. Regulated advertisers, including drugmakers, may require strict control over what claims appear and what qualifications are displayed.

In practice, you can separate "what the system can generate" from "what must be approved":

• Lock factual claims. Product specs, pricing, health claims, and legal qualifications should come only from approved advertiser materials.
• Allow variation in low-risk copy. Tone, sentence structure, and format can be tested within clear limits.
• Keep sponsorship visible. The interface should make it obvious that an ad is an ad at the moment users see it.
• Measure more than clicks. Track complaints, hides, conversions, repeat use, and whether users can identify sponsored content.

The likely advantage of chatbot advertising isn't that every ad becomes "smarter." It's that the system has more context for deciding what might be relevant. But that advantage only matters if targeting is bounded and the creative stays clearly commercial.

How does conversation-level targeting affect privacy?

Using conversational context can improve relevance - but it can also include information users never meant to share with an advertiser. People may use chatbots to discuss health issues, financial pressure, relationship problems, or work stress. Typing a prompt into an assistant isn't automatic consent for the entire conversation to be used for ad selection, retargeting, or future personalization.

This is where conversation history is different from a single search query. A search can reveal an immediate need. A longer back-and-forth can also reveal the reasons behind that need, plus unrelated details from earlier turns. Sending that context into an advertising system creates a new data flow - even if the system uses it only to pick an ad "right now."

Koah's Baird said the company anonymized queries and stored them in a de-identified way, but the company didn't spell out more in its account about collection and storage practices. Those descriptions aren't enough to judge privacy risk on their own. De-identification may reduce exposure, but it doesn't answer practical questions like what data gets transmitted, how long it's retained, whether it's linked across sessions, or what users can control.

For product teams, the strongest privacy approach is to minimize the data used - not to rely on a promise that sensitive data will be handled carefully. A reasonable decision process looks like this:

1. Define the purpose. Decide whether the system needs the current prompt, the whole conversation, or just a broad topic category.
2. Use the least sensitive signal. If a general topic works, don't pass the full transcript to an ad partner.
3. Set retention and reuse rules. Data used to select one ad shouldn't quietly turn into a long-term profile or retargeting record.
4. Give users meaningful controls. Tell people what's used and provide a real way to limit personalization.
5. Test the boundary. Ask whether the ad would feel acceptable if the user knew exactly which conversation details drove it.

This matters even more for sensitive categories. A spa ad after a general conversation about stress might feel harmless. The same mechanism applied to a discussion of a specific medical condition, debt, or a family crisis can feel exploitative. A system needs rules that recognize those differences - rather than treating every context as a chance to target.

There's also a wider trust issue: users may not know whether the assistant is responding based on its own reasoning, an advertiser's influence, or both. Clear sponsorship labels help, but platforms should also stop paid placement from silently changing what the assistant would otherwise say. If an ad shows up after an answer, users should be able to tell where the answer ends and the paid message begins.

Personalization isn't automatically a bad thing. The standard should be whether users can understand and predict how their conversation gets used. If the system can't explain that clearly, it shouldn't rely on conversational context for advertising.

Can advertising work when AI agents shop?

The next phase of AI commerce could involve shopping agents - systems that compare products, manage tasks, and complete purchases for users. That changes the advertising problem. Today, an ad is shown to someone who decides whether to click. In an agent-mediated purchase, the system may evaluate options on the user's behalf - meaning an advertiser could try to influence what the agent recommends before the user even sees it.

There is a legitimate role for sponsored offers here. A discount, a product that fits the user's requirements, or a faster path to purchase can all be genuinely useful. But an agent shouldn't treat payment as a substitute for quality. If a sponsored listing gets preference, the platform has to disclose that to the user and make it clear to the agent doing the recommendation.

This distinction matters because agents optimize against instructions. If an agent is told to find the best value, a paid placement shouldn't quietly redefine "best" as "most profitable for the platform." The user's criteria - price, suitability, delivery time, accessibility, or whatever requirements they specify - should stay the basis for the recommendation. Sponsored status is something the agent should factor in, not a reason to override those criteria.

Ads may also show up during customer support. Koah estimated that about one in four chatbot queries about a retailer or brand involved returns, exchanges, or other support issues. If that estimate is directionally right, it points to an opportunity: a company could use an assistant to explain a warranty, offer a replacement, or resolve a complaint before the user escalates to a human support channel.

But this is where placements can go wrong quickly. A customer trying to return a product shouldn't be pushed toward a paid offer that obstructs the return or pressures them to spend more. Ads might help if they actually support the resolution - but support has to stay available regardless of whether the user clicks or buys. Otherwise, advertising turns a service interaction into a sales funnel at the customer's expense.

Companies that want to run agent-facing ads should set clear rules:

• The user's stated requirements should come before sponsorship.
• Sponsored products and commercial relationships should be disclosed in a format the agent can understand and pass along.
• Support and cancellation paths can't be made harder to protect ad revenue.
• The platform should record why an option was recommended, so disputes can be investigated.
• Users should be able to decline sponsored recommendations without losing access to the core task.

There's no guarantee that advertising will become the default model for AI shopping. Some businesses may prefer transaction fees, subscriptions, or other revenue sources. Some users may just reject ads in assistants entirely. A specialist chatbot executive, Kevin Baragona of DeepAI, said ad revenue is meaningful, but the experience doesn't yet feel right on privacy and design grounds. That's a useful reminder: a revenue stream can exist before a product has earned the right to scale it.