To understand what was happening — and what could change — we had to be honest about how these four engines actually work. They're not the same product.
Three of them lean primarily on training-data substrate: the snapshot of the internet the model was trained on, sometimes months or years old. Retrieval is layered in as augmentation when a query needs it, but it isn't the foundation. Whichever brands had the biggest content footprint at the time of the last training cycle get baked in as the default answers.
One of them, Perplexity, is built the other way around. Retrieval is the substrate. Every query starts with a live web search; the model synthesizes from whatever is currently indexed and structured well enough to be quoted.
| Engine |
Primary substrate |
Updates on |
| ChatGPT |
Training-data first, with retrieval layered on for some queries |
New training cycles + occasional web search |
| Claude |
Training-data first, with retrieval triggers |
New training cycles + web search tool |
| Gemini |
Training-data first, augmented with Google index |
New training cycles + Google index |
| Perplexity |
Retrieval-first synthesis on every query |
The live web, continuously |
FedEx Office, Staples, and ARC won training-data-era AI for the same reason they won pre-AI search: they had the most content, the most citations, the most brand mass at training time. That advantage is locked in until the next training cycle.
But it isn't locked in on Perplexity. Perplexity sees the live web on every query. The instant Azul Prints had better structured content, cleaner schema, deeper category authority, and more topical pages than the incumbents — Perplexity would surface it. The other engines would follow, but only on their next training refresh.
Perplexity is the leading indicator. Winning Perplexity now means winning ChatGPT, Claude, and Gemini at their next training cycle.
Read the full thesis: the training-data lag →
