The Trust Problem in AI Monetization
When publishers first launched AI assistants on their properties, most treated monetization as a later problem. The product priority was getting the assistant working well, building user habit, and figuring out the revenue model once the product had traction. That sequencing made sense, but the later problem is now present for many publishers who have built genuine user adoption and are wondering how to generate revenue without disrupting what made the product worth using.
The risk is real. An AI assistant that users trust to give honest answers and good recommendations is a valuable product. An AI assistant that injects commercial messages without clear disclosure, or that seems to favor sponsored answers over accurate ones, loses that trust quickly and is difficult to rebuild. The history of the web is full of examples of publishers who burned down a good content product by treating monetization as something that could be layered on aggressively after the fact.
At Velocity, we have been thinking about this as a product design problem from the beginning. The principles below are not aspirational guidelines. They are constraints we actually built into the platform, because we think trust-preserving monetization is the only model that works at scale in conversational contexts.
Principle 1: Disclosure Before Content
The sponsored label on an ad must appear before the user reads the ad content, not after. This sounds obvious, but a lot of native advertising on content platforms violates it in practice: the headline or opening text appears to be editorial, and the "Sponsored" tag is in small type below, or tucked into a corner where it is easy to miss.
In conversational advertising, disclosure-after-content is a more serious problem than it is in display. When a user asks a question and receives what appears to be an assistant recommendation, finding out after reading it that it was a paid placement feels like deception. The FTC's guidance on native advertising is clear that disclosures need to be conspicuous, but beyond legal compliance, the user experience is what matters here: the user should know what they are reading before they read it.
Velocity's sponsored card format places the label as the first visual element. It is not negotiable and it is not configurable. Publishers who want to style it differently can change font color and card background within a defined range, but the label position is fixed. This is a case where publisher flexibility would work against publisher interests.
Principle 2: Relevance Is Not Optional
An ad that is irrelevant to the conversation is not just useless for the advertiser. It actively undermines user trust in the assistant's judgment. If a user asks about budgeting apps and receives a sponsored recommendation for a home security product, the mismatch signals that the assistant is not actually paying attention to what the user asked. That impression sticks even when subsequent answers are excellent.
Velocity does not fill with low-relevance ads to chase impressions. When the matching engine cannot find a contextual match that meets the relevance threshold, it returns no-fill. An unfilled turn is better for the publisher in the long run than a mismatched sponsored card. Publishers should evaluate fill rate in context: a 65 percent fill rate with tight relevance is better for user trust than an 85 percent fill rate with poor matches.
We think this is the right call, though we also want to be honest: no-fill turns mean fewer revenue opportunities in the short term. Publishers who want to maximize immediate fill should know that we are prioritizing relevance quality over fill rate, and that tradeoff is a deliberate product decision, not a capability gap.
Principle 3: The Assistant's Answer Comes First
The ad insertion happens within the response, not instead of it. The assistant still answers the question. The sponsored card is supplementary: a relevant commercial suggestion that extends the response, not a replacement for it. A user who asks "what project management tool should I use for a small remote team" gets an assistant answer and, if there is a relevant advertiser, a sponsored recommendation alongside it. The answer is not replaced by the ad.
This sounds obvious but it is worth stating because some approaches to in-chat monetization work differently: the assistant's response is actually generated or shaped by commercial considerations, blurring the line between organic answer and paid placement. We do not do that. The response is generated first. The ad is considered independently. The user gets both, clearly separated.
Principle 4: Frequency Capping at the Conversation Level
One sponsored card per conversation session is the default frequency cap. Publishers can adjust this up to one per five turns as a maximum, but the default reflects our view that a conversation should not feel like it is being interrupted every few exchanges by commercial messages.
The psychology of conversation frequency is different from the psychology of display ad frequency. When someone is scrolling an article page with multiple banner ads visible simultaneously, a second or third banner does not feel like an interruption because they all coexist in the same static field. In a conversation, ads appear sequentially in the flow of the exchange. Two ads within five turns feel intrusive in a way that two banners on a page do not. Publishers who override the default to maximize turn-level impressions tend to see faster engagement drop-off than those who keep frequency low.
Principle 5: Publisher Category Controls, Not Platform-Wide Defaults
Not every category of advertising is appropriate for every publisher context. A mental health assistant should not serve ads for alcohol. An educational product for younger audiences should not serve ads optimized for adult consumer categories. The publisher is the one who understands their audience and their content context. The ad platform should give them the controls to enforce that judgment, not just a set of platform-wide filters determined by the highest-bidding advertisers.
Velocity gives publishers a category exclusion system with 180 IAB taxonomy categories. You can configure up to 20 category exclusion rules, selecting any combination from across those 180 categories. The exclusions apply at the publisher level and cannot be overridden by advertisers bidding on your inventory. A publisher who has excluded gambling advertising from their assistant will not see a gambling ad appear even if that advertiser bids at a premium on a contextually matched turn.
This is one of the areas where we think the publisher-control model matters most. Programmatic display advertising has historically given publishers limited visibility into what ads run on their properties, and publishers have paid the trust cost of that opacity. Conversational advertising needs a different approach, because the ads appear inside a product the publisher's users are actively engaged with, not adjacent to content they are scrolling past.
What These Principles Cost, and Why We Think It Is Worth It
These principles are not cost-free for us or for publishers. Tight relevance requirements mean lower fill rates than a platform that prioritizes impression volume. Mandatory disclosure formatting means some creative options that advertisers prefer are not available. Frequency caps mean fewer ad opportunities per session than a platform that runs ads more aggressively.
We think those costs are the right investment in a sustainable market. Conversational advertising is early enough that the norms being established now will shape how users and regulators view the channel for years. A market where early publishers maximize short-term revenue at the cost of user trust will produce a regulatory and user-behavior backlash that hurts everyone. A market where early publishers demonstrate that ads in AI conversations can be relevant, disclosed, and non-intrusive creates a durable channel.
Publishers who care about the long-term value of their AI assistant products should be skeptical of monetization approaches that are not transparent about these tradeoffs. The choices you make in the first year of AI assistant monetization will be harder to undo than the choices you made in the first year of display advertising, because user trust in an AI assistant is more personal and harder to rebuild than trust in a website that shows banner ads.
