When we started designing Velocity's ad format, the first question we had to answer was what problem we were solving for. The easy framing is advertiser reach: build a format that gets ads in front of chat users. But if that is the only lens, you end up with something users learn to distrust and publishers regret deploying.
The harder framing, which is the one we came back to repeatedly, is: what does an ad look like when it is genuinely useful? When a user asking a question about refinancing their mortgage sees a relevant lender's offer appear in the response, is that an interruption or is it a resource? The answer depends entirely on format design, match quality, and placement timing. Get all three right and the ad belongs there. Get any one of them wrong and the whole thing feels wrong.
What We Ruled Out First
We went through several ad format concepts before arriving at the sponsored card. Understanding why we rejected the alternatives explains the reasoning behind what we built.
The first concept was inline text insertion: embedding a sponsored phrase or link directly within the assistant's response text, styled differently to indicate sponsorship. We ruled this out quickly. Users read assistant responses as a coherent voice. Injecting a sponsored phrase into that prose, even with clear labeling, creates a confusing hybrid where the assistant's authority and the advertiser's message blur together. Users who notice it feel manipulated. Users who do not notice it are being misled. Neither is acceptable.
The second concept was a banner-style format that appeared below the full response. This is closer to how display advertising works in editorial contexts. The problem is that it is invisible at the wrong times: long responses push the banner below the visible viewport before the user has finished reading, and short responses cause the banner to appear before the user has fully engaged with the content. Timing and placement are impossible to control.
The third concept was a follow-up prompt injection: a sponsored suggested question that appears below the assistant's response, alongside the regular conversation interface. This has some appeal because it does not interrupt the response flow, but it creates a different kind of deception: the suggested prompt looks like the assistant recommending a next step, when it is actually a paid placement. Even with labeling, the interface affordance creates an expectation that the suggestion is organically generated.
The Sponsored Card: What It Is and Why
The sponsored card is a distinct, clearly bounded element that appears within the response flow at a natural paragraph break. It has a consistent visual structure: a slightly differentiated background, a left-border accent in the publisher's configured style, a "Sponsored" label as the first visible element, an advertiser product name, a single-sentence description, and a call-to-action link.
The visual distinction from the surrounding response text is deliberate and required. We do not allow publishers or advertisers to style the card in a way that makes it visually blend with the assistant's prose. The border, the background differentiation, and the sponsorship label are non-negotiable. The card should be immediately recognizable as a different kind of content from the response it is adjacent to.
At the same time, the card is typographically consistent with the rest of the chat interface. It uses the same font family, similar sizing, and a clean layout that does not introduce visual chaos into the reading experience. The goal is a format that is clearly labeled but not aggressively disruptive, present but not demanding.
Placement Timing in the Response Stream
Where in the response the card appears is as important as what the card looks like. We experimented with several insertion strategies before settling on first-paragraph-break insertion as the default.
The logic: the assistant's first paragraph typically establishes context and directly answers the user's question. Inserting an ad before this paragraph would interrupt the answer before the user has received it. Inserting the ad after the full response means the user has already moved on cognitively. The natural break after the first paragraph, the point where the response transitions from direct answer to explanation or elaboration, is the moment when the user has received the core information and is most receptive to a relevant additional resource.
This is not a universal rule. Very short responses with no natural paragraph break do not inject an ad at all. Responses where the entire content is a single continuous statement, like a quick factual answer, are not interrupted. The insertion logic evaluates response structure and passes on the placement if no clean break point exists. A turn that does not get an ad is preferable to a turn where the ad appears in a place that disrupts the information flow.
Publisher Customization Within the Format
Publishers can adjust the sponsored card's left-border color within their brand palette, the CTA button label, and whether the card includes or omits the advertiser's logo. They cannot remove the "Sponsored" label, cannot make the card visually indistinguishable from response prose, and cannot inject additional content into the card body.
This is a deliberate constraint. We have had publisher requests to allow richer card formats, additional images, and more flexible layout. We have pushed back on most of these because richer formats increase the cognitive load of the placement and make it harder for users to quickly understand what they are looking at. The current card is minimal by design. Every element that is present is present because it carries information the user needs; nothing is there for engagement optimization alone.
Measuring Format Quality
We track several metrics to evaluate whether the format is performing well beyond standard click rates. The one we weight most heavily internally is session continuation rate. Revenue implications of format decisions are explored in the publisher revenue benchmarks: do users continue the conversation after seeing a sponsored card, at the same rate as turns where no card appeared? A format that causes users to drop off the conversation is failing regardless of its click rate.
Across our early-access publisher data, session continuation rate after a sponsored card insertion is within a few percentage points of the baseline for non-monetized turns. This is not a claim that users do not notice the cards. They do notice them. The data suggests they are not treating them as a reason to stop the conversation, which is the bar we set for ourselves when we started this design process.
The sponsored card is not a finished design. It is the version that meets our current standard for contextual fit and transparency. We are paying attention to how users interact with it as the publisher base grows, and we expect the format to evolve as we learn more about what works across different assistant types and topic contexts. What will not change is the baseline requirement: clear sponsorship labeling, visual distinction from response content, and placement timing that respects the user's information-seeking flow.
