Google's retrieval edge over OpenAI
Kavin Stewart, Partner at Tribe Capital, on Reddit's 10x opportunity
The real moat in AI search is not the model alone, it is the system that can find fresh, relevant information fast and feed it into the model cheaply. Large language models are good at turning inputs into fluent answers, but they still depend on retrieval when a question needs current facts, niche pages, or exact sources. That makes web indexing, ranking, filtering spam, and securing access to closed content a separate infrastructure advantage that can be harder to rebuild from scratch than training another frontier model.
-
Google already had the pieces that matter for live information, a web crawl, an index, ranking systems, and distribution through Search. OpenAI later built search features and its own search crawler, which shows how model labs eventually have to add retrieval infrastructure around the model instead of relying on the model by itself.
-
The market that emerged after GPT 4 made this concrete. Companies like Exa grew by selling search APIs that give models better web results, because raw browsing often pulled in weak or stale pages and lowered answer quality. In practice, the retrieval layer became its own product category, not just a minor add on.
-
This is also why content partnerships matter. As more valuable information sits behind platform walls like Reddit, publishers, or apps, winning AI search requires both technical indexing and legal or commercial access to the data. The advantage shifts from who has the smartest model to who can assemble the best retrieval network.
Going forward, the strongest AI products are likely to look less like standalone chatbots and more like tightly integrated retrieval systems with a model attached. Model quality will keep converging, while the harder long term advantage will sit in owning the pipes to fresh data, the ranking layer, and the user surface where retrieval happens thousands of times a day.