Not every prompt sends an AI to the web

Ask Google what time it is in Singapore and you get the answer instantly. Nothing gets crawled. No page gets fetched. The system already knows how time zones work and what time it is, so it does the arithmetic and hands you a number. Ask the same system which running shoe is best for flat feet in 2026 and something different happens. It has to go look, because the answer depends on products that exist right now, reviews written recently, and prices that moved last week. Large language models behave the same way, and this is the part most content strategies miss. A model is not a search engine with a chat interface in front of it. It is a system that decides, prompt by prompt, whether it already knows enough. Every piece of content you publish is competing only for the prompts where that answer is no.

Why models skip the search step most of the time

Models are trained on enormous amounts of text, and for a large share of what people ask, that training is enough. Definitions don't change. Most procedures don't change. General advice questions rarely hinge on this week's news. When someone asks what a word means or how a basic process works, the model already has a reasonable answer sitting in its weights. Search only gets triggered when internal knowledge is stale, incomplete, or simply not confident enough. It's the same logic Google has run for years against its own index: a define: espresso query doesn't need a crawl, but best espresso machines 2026 absolutely does, because the answer is tied to inventory, pricing, and reviews that change constantly.

The data: which prompt types actually force a fetch

Research from Profound looked at how often different categories of prompts triggered a live web search in Claude, and the spread is wide. Worth noting up front: Claude only reaches for search on roughly 37% of prompts overall, versus roughly 90% for ChatGPT. The intent behind a query matters even more on Claude, where the model decides case by case whether it's confident enough to skip the fetch.

Search rate by prompt intentClaude

Recency 81%

Ranking 67%

Local 55%

Comparison 51%

Personal advice 38%

Definitional 24%

Procedural 11%

The gap between the top and the bottom of that list is the whole story. A procedural page competes for a slice of prompts where the model reaches for the web one time in nine. A recency page competes where it reaches four times in five.

The four intents worth building for

Four categories sit clearly above the pack, and they share a trait: the correct answer today isn't necessarily the correct answer next month. Here's what makes each one a fetch trigger, using running shoes as the working example.

Search rates from Profound's prompt research on Claude. Example queries are illustrative.

Where Prism Optimizer fits in

Knowing that four intents drive most of the fetch behavior is useful in theory. It gets a lot more useful once you know how much of your own site actually falls into those buckets. That's the part we automated. Prism Optimizer scans your entire URL inventory and sorts every page into a fetch-probability bucket, using the same lens an assistant applies when it decides whether to answer from memory or go get fresh information. Instead of guessing which pages have a shot at surfacing in an AI answer, you get a straight count: how many pages sit in Recency, how many in Ranking, how many in Local, how many in Comparison, and how many are stuck in the low-probability zone answering definitional or procedural questions the model already knew. The classification isn't a guess. It runs on the same scoring framework we use everywhere else in the product: Run that scan across a full site and a pattern usually shows up fast. Most of the content sits in the low-probability buckets, quietly answering questions the model already knew, while the handful of pages that could actually win a live citation are thin, outdated, or missing the structure an assistant needs to pull from them. That gap is the opportunity. It's also the first thing a Prism Optimizer audit puts in front of you.

Sources

Related reading

• What actually gets a page quoted in an AI Overview — once the fetch happens, this is what decides whose sentence gets lifted.
• The citation is the conversion — what to count as the outcome when the click never arrives.
• Real movement versus measurement noise — why answer surfaces move on their own, and how to tell that apart from progress.
• Live answer-surface data — the fan-out queries and AI Overview data fetched on every scan.