Google's search ad machine is now bleeding $9 billion a quarter
Issue #070 · 4 min read
Google's search ad machine is now bleeding $9 billion a quarter
Google and Reddit are both losing money to the same thing: AI answers that skip the click.
The big story
Google just told investors something uncomfortable. AI search makes less money than the search it's replacing, and the hole is getting bigger.
Negative cash flow hit $9.06 billion last quarter. The stock dropped after the earnings call — that's Ars Technica.
Google's fix is more ads inside the AI answers themselves. Which tells you the first version of AI search wasn't built to make money — it was built to keep people from leaving for ChatGPT.
The knock-on damage is showing up somewhere else entirely. Reddit's traffic is drying up because fewer people are clicking through from search at all, so Reddit is pulling back toward a logged-in product instead of an open one — that's TechCrunch.
Platformer frames Google and Reddit's bad weeks as the same problem wearing two faces: the open web ran on people clicking through, and AI answers are built to make the click unnecessary.
If you run a product that lives on search traffic or on being cited by an AI answer, this is the number to watch. The company that invented the ads-on-search business model just admitted the AI version of it doesn't pay for itself yet — and its answer is to put more ads in front of you, not fewer.
What shipped
GitHub Copilot Fusion is now generally available — instead of picking a model per task, you hand that decision to Copilot and it routes and orchestrates models for you, per GitHub's changelog. If your team has been maintaining a mental chart of "which model for which task," this is GitHub betting that chart becomes obsolete. Worth a trial run before you renew any per-model tooling spend.
GitLab Orbit is GA, aimed at a real problem: AI code review agents that say "safe to merge" after seeing only the diff, with no view of the pipeline, the ownership, or the vulnerability history — what GitLab calls "artificial confidence, not artificial intelligence," per their launch post. A companion post shows Orbit finding who owns an exposed vulnerable component without the usual spreadsheet-and-Slack scavenger hunt — see GitLab's writeup. If your eng org is already drowning in AI-reviewed PRs nobody fully trusts, this is a direct answer to that complaint, not a feature nobody asked for.
OpenAI published an updated Model Spec, its public document on what it wants its models to do when instructions conflict — now built around five values aimed at "sound judgment, human agency, and faithfulness to human intent," per OpenAI. Alongside it, OpenAI detailed how it actually watches for models going off the rails in production — layered monitoring meant to catch unauthorized agent behavior before it escalates, per this second post. Neither is a feature you'll touch this week. Both are worth five minutes if you're the one explaining to legal or security why your vendor's agents can be trusted with write access.
What I'd actually do this week
- Check your traffic sources. If any part of your product depends on search-driven discovery, Google's ad-heavy AI answers and Reddit's pullback from the open web are both signals the funnel is narrowing. Model what a 20% drop in organic click-through does to your growth plan.
- Pilot GitHub Copilot Fusion or GitLab Orbit on one real repo, not a toy project. The claims — lower token cost, less "artificial confidence" — are only worth believing once you've watched them handle your actual codebase's mess.
- Pull up OpenAI's updated Model Spec and compare it against whatever your own team's AI usage policy says. If you can't tell where your policy diverges from the vendor's, that's the gap someone will find for you eventually.
Hit reply if you've already run Copilot Fusion or Orbit against a real repo — I want to hear what broke.