Search | The Moat of the Search Index
A year or two ago, I was a bit nervous about whether GPT could handle fresh search well, given it can’t tap Google. If Google is the “better” search engine, why does ChatGPT – which is forced to use Bing – still return such high-quality results?
The key is the difference between Searching (what humans do) and Retrieving + Reading (what AI does). In many ways, the moat of the traditional search index is gone.
When you GOOGL 0.00%↑ something, you normally type one query and look at the first or second result. If the top result is bad, you immediately think “Google failed.”
Google’s edge has always been in:
Ranking the single best link
Dealing with SEO spam, typos, long-tail queries
UI and instant answers, etc.
ChatGPT doesn’t operate like that. It has an “agent” advantage – it basically cheats by working harder (i.e. going through as many pages as needed). It uses multi-hop reasoning:
Ask Bing with query rewriting
Pull the top N results
Read them deeply and extract the relevant pieces
Synthesize an answer
Because it pulls from multiple sources, it naturally hides the failures of any single search result.
The gaps probably still show up in true edge cases: very long-tail or oddly phrased queries, ultra-fresh events (minutes or hours old), or heavily spammed / SEO-gamed topics.
I think this is an underappreciated point - for most queries, the moat of today’s search index is largely gone.
—
ChatGPT uses Bing.
Claude uses Brave Search (an independent, privacy-focused search engine).
Perplexity uses a combination of Bing, Google (via third parties), and its own index.
Gemini uses Google (obviously).
I might take the opposite side here that search is actually becoming a moat for agents but I'm like 50/50. Two possible reasons:
1. Ad monetization - seems clear that openai is in a crunch monetizing itself enough to cover massive data center costs. The two models seem to be subscriptions or ads. OpenAI seems to be moving towards ads which makes these search indices (and their underlying ads infra) more important.
2. Model routing - We are living in this era of subsidized model consumption where we throw amazing models at even dumb tasks. These models can hoover up suboptimal search results and then put out a decent response. But once these companies need to start worrying about cost, I suspect we will see less options for model selection and more platform owned routing to small models that default to search when questions are outside their training. If this happens, solid search indices are still critical to outputting good results. You even see this problem now with enterprise search tools that try to add Agents on top of enterprise search -- because the underlying search index isn't great the output of the agent is oftentimes not useful.
so the its only chatgpt as the title implies or apply to all other chats.