All writing

How fast do LLMs pick up new information from the web?

, 2 min read

How fast do LLMs pick up new information from the web? I ran a simple AI search experiment to find out.

The setup

  1. I invented several new English words that didn’t exist anywhere online.
  2. For each word, I wrote a definition-heavy page.
  3. I published them in different formats: a news article, a homepage, a product page and a blog post on existing sites; a public Notion doc; and brand-new sites with exact-match domains, like zemuretomal.info. In total: 30 pages across 3 sets, each featuring a unique term.
  4. Some sites included llm.txt, some didn’t.
  5. No backlinks, no social signals, no indexing requests.

Then I waited. Every day I searched for each term in ChatGPT, Gemini, Perplexity, DeepSeek, Google and Bing, both as plain queries and with the site: operator.

Five days later

  1. Gemini showed one definition on day one, before Google showed the page as indexed. Later it stopped showing it entirely, even though the page is now indexed. Possibly a content filter, or resource optimization suppressing low-signal terms.
  2. ChatGPT gave full definitions with source links, even when the pages weren’t visible on Google or Bing. It’s clearly using other indexing datasets.
  3. Perplexity performed surprisingly well. It consistently surfaced correct definitions and linked to the source pages quickly.
  4. DeepSeek was the last to catch on.

Which pages worked best

  • News articles were picked up first by several models, possibly because of their semantic structure or assumed freshness.
  • Exact-match domain pages also did well. Focused, on-topic and self-contained: LLMs seemed to like that.
  • Homepages did better on visibility.
  • Public Notion docs were invisible. I didn’t expect much, but I needed something to compare with.

Other findings

  • Search engines lagged behind. Google indexed most pages by day five, Bing only one. LLMs started showing content before Google did.
  • llm.txt doesn’t help.
  • Definition-style clarity matters. Pages with a single semantic purpose, explaining one concept, seem to get surfaced faster.
  • Perplexity and ChatGPT showed the strongest “indexing” behavior, and they can clearly be influenced by what we put online.

This wasn’t about traffic. It was about visibility. Next, I want to test whether synthetic pages at a larger scale can influence real search topics.