What We Measured and Why It Matters
One of the questions we hear most from publishers considering Velocity is some version of "what will I actually earn?" It is a fair question, and a hard one to answer in the abstract. Display advertising took decades to develop the CPM benchmarks that publishers now use as rough planning inputs. Conversational advertising is a few years old at most, and the publishers running it are early enough that there is no established comps market to reference.
That said, we do have data. Over the past 60 days, 14 publishers in our early-access program have been running Velocity across their AI assistant products. Their collective output gives us the first real baseline for what conversational ad revenue looks like in practice, at least at early scale. We are sharing these numbers here because we think the industry is better served by transparency than by protected internal benchmarks. We will also explain what drives variance, because the average hides quite a bit.
The Baseline Numbers
Across our early-access cohort, the average RPM (revenue per thousand conversation turns) on contextually matched turns was $2.40. That number applies to turns where Velocity made a successful contextual match. Not every turn receives an ad, and that is by design. Publishers can set frequency caps that limit how often an ad appears, and the matching engine only inserts an ad when the relevance threshold is met.
Fill rate averaged 68 percent across the cohort. That means roughly 68 out of every 100 eligible turns received a matched ad insertion. The remaining 32 percent either did not meet the relevance threshold for any active advertiser in the auction, fell within a publisher-set category exclusion, or were suppressed by frequency capping.
Translating that into estimated earnings per publisher: a publisher running 500,000 conversation turns per month, with a 68 percent fill rate and $2.40 RPM on matched turns, would earn approximately $816 per month before revenue share. On the Starter tier (70/30 split), that is around $571 to the publisher. On Studio (75/25), around $612.
These are early-cohort figures. Both RPM and fill rate should improve as advertiser inventory on the platform grows and as publishers optimize their topic configurations.
What Drives Fill Rate Variance
Fill rate varied significantly across our cohort, ranging from roughly 50 percent to above 80 percent. The variance breaks down into three identifiable factors.
First, topic coverage. Publishers whose AI assistants focus on topics with strong advertiser interest, such as personal finance, software tools, home improvement, and health and wellness, see higher fill rates than those in lower-monetization categories. This mirrors the display advertising dynamic where lifestyle and business content commands more advertiser demand than, say, academic research tools.
Second, category exclusion settings. Publishers who configured aggressive exclusion lists reduced their effective addressable inventory. There is a real tradeoff here: tighter controls give publishers more confidence about what ads appear, but they do reduce fill. We recommend starting with broader settings and tightening based on what you actually see, rather than pre-excluding categories that may never appear in your specific conversation set.
Third, frequency cap settings. A cap of one ad per five turns produces more fill opportunities than one per ten. Publishers setting very low frequency will see their effective RPM per total turn drop considerably. The right number depends on how much ad density feels appropriate for your product. Testing beats defaulting to the most conservative setting.
How RPM Compares to Display
Display CPM for a typical mid-tier publisher runs anywhere from $1.50 to $6.00 depending on audience, geography, and ad quality. Conversational RPM on matched turns is currently in that same range. That comparison is worth interpreting carefully, though.
A page view in display advertising might serve two to four ads simultaneously, across banner, sidebar, and interstitial placements. A conversation turn in Velocity serves at most one ad, inserted at one moment. So the per-impression revenue is comparable, but the gross volume of impressions per session is lower in conversational advertising.
The counterargument, and the reason we think conversational ad revenue scales attractively, is session depth. A user asking 20 questions over a 15-minute session creates 20 eligible turns, versus maybe 3 page views with a short scroll each. Heavy AI assistant usage tends to produce more turns per session than a typical browsing session produces page views. Publishers with engaged AI assistant audiences may find that total impressions per session compare favorably.
What Early-Stage Numbers Cannot Tell You
We want to be direct about the limits of this data. Fourteen publishers over 60 days is a small sample. The cohort skews toward technically sophisticated publishers who self-selected into an early program and configured it carefully. They are not representative of the full range of publishers who might eventually use Velocity.
Advertiser competition is also thin at this stage. As more advertisers enter the auction, CPMs on contextual matches should rise. Current RPM benchmarks likely understate what a comparable publisher could earn in 12 to 18 months as the demand side scales. We are not saying you will definitely earn more later. What we are saying is that these early numbers should be read as a lower bound, not a ceiling.
Planning Your Integration Around Revenue
If you are evaluating Velocity for a publisher integration, here is how we suggest framing the revenue calculation. Estimate your monthly conversation turns. Apply a 60 to 70 percent fill rate as a conservative planning assumption. Use $2.00 RPM on matched turns as a conservative floor. That gives you a rough monthly earnings estimate to compare against your integration costs and ongoing platform fees.
For a publisher at 1 million turns per month, that is approximately $1,200 per month at conservative assumptions and closer to $1,632 at our observed fill and RPM averages. Neither of those numbers is a guarantee, but they give you a planning anchor grounded in real cohort data rather than theoretical projections.
The goal of publishing these benchmarks is to help publishers make a decision based on realistic expectations. If the numbers make sense for your product and audience, the integration is worth trying. If they do not clear the bar today, we would rather you know that than discover it three months into an integration.
