Skip to main content
Case Study · Code AI

CodeBerry shipped sidebar ads in the IDE with zero impact on session length

CodeBerry’s challenge was existential: fund the free tier without the distraction-cost of traditional in-editor ads. They shipped sponsored tool recommendations in the right sidebar — only when commercial intent was high — and cleared $12K/month on 85K WAU without a measurable session-length change.

The problem

Running the GPT-4 and Claude Opus inference behind the free tier cost $28K/month in raw API spend. Existing $8/mo sub tier covered 40% of that. The gap was closing the window for a Series A.

In-editor ads are a category no one has done well. Display units break focus, and the market knows it.

The team refused to ship anything that added a single measurable ms to the autocomplete flow — a constraint their users would notice immediately.

What they did

Integrated the Surfacedd Vercel AI SDK adapter. Sponsored recommendations stream after the main response, never mid-completion.

Placement: the right sidebar, triggered only when the AI detected commercial intent in the query (e.g., "which postgres vector extension should I use"). Non-commercial queries get no ad at all.

Disclosure is a persistent "Sponsored" label + hover tooltip explaining the matching criteria. No pop-ups, no expansion, no autoplay.

Launched as a free-tier feature; paying subscribers see the same sponsored recs but with a 90% reduced frequency, so premium users still get valuable product recommendations.

Results

$12,300
Monthly ad revenue
from 85K WAU, roughly $0.145 RPU
+0.2%
Median session length
not statistically significant
0 ms change
Autocomplete latency
sponsored fetch runs post-response, never in the hot path
+0.3%
Upgrade-to-paid rate
marginal positive effect from premium tier’s reduced-ad experience
44%
Share of CPU costs covered
$12.3K / $28K monthly inference spend
We told the team: if it adds even one millisecond to autocomplete, we kill it. Surfacedd’s post-response pattern was the first ad architecture we saw that didn’t violate that constraint. We now cover 44% of our inference bill with it.
Marcus Chen, CTO, CodeBerry

Key takeaways

  • Commercial-intent-gated ads convert better and complain less than frequency-based ads.
  • Post-response streaming placement protects the hot path and the user’s trust.
  • Ad revenue as an inference-cost subsidy is a legitimate Series-A lever, not a sign of bad unit economics.

Frequently asked questions

How did CodeBerry define "commercial intent"?

They used the Surfacedd intent classifier in the SDK, tuned against a manual labeled set of 500 CodeBerry queries. The classifier flags queries like "which X should I use" or "compare Y and Z", which are roughly 8% of total query volume.

Does the sponsored sidebar affect premium conversion?

Marginally positive. The team saw a +0.3% upgrade rate — statistically weak but directionally correct. The premium-reduced-ads experience is a clear value prop for heavy users.

What happens with the sponsored content on enterprise installs?

Enterprise plans turn off ads entirely. The sponsored layer is consumer-tier and free-tier only; enterprise customers pay a license that covers the inference cost directly.

Want numbers like CodeBerry’s?

Estimate your app’s potential in the revenue calculator, then start the integration.

Try the calculator →See the SDK