Training data

Training data is the text an AI model learned from before release. The knowledge cutoff is the date that text stops, so the model knows nothing newer unless it searches.

AI searchUpdated 2 min read

Also called parametric knowledge · pre-training data

What it means

A language model learns from a huge amount of text before it's released. That text is its training data. The date the text stops is its knowledge cutoff. Anything that happened after that date, the model only knows if it searches the web.

How it shapes recommendations

Brands that were written about a lot before the cutoff are often "known" to the model. It can name them without searching.

We saw this in our study. For 109 of 210 buying questions, ChatGPT put brand names into its web search before it had read anything. It started from a shortlist it already knew, then searched to check and fill in the rest. 98% of the brands it named in advance and then recommended were also on a list it found in the search.

So training data gives big brands a head start. But the pages the AI reads during the search still decide most of the answer.

What you can do

You can't edit a model's training data, and you can't see when your brand gets into it. What you can do:

  • Get covered on many sites over time. Future models learn from what's widely written.
  • Get onto the pages AI reads today. This works now, and you can measure it.
  • Don't block AI search crawlers, so tools can read your site when they search. See AI crawlers.

Example

A bakery chain was well known in 2022 and wrote about often. A newer competitor opened in 2025.

Ask a model with search turned off for the best bakeries in the city, and it will likely name the older chain, because that's what was in its training data. Turn search on, and the newer one can show up, if it's on the pages the model reads.

Frequently asked questions

My company is new. Can AI still recommend it?

Yes. AI search tools look things up on the web, so a new brand can be recommended if it's on the pages they read. In our study, brands the AI didn't name in advance still got into answers through the lists: 90% of them were on one.

How old is a model's training data?

Usually several months to over a year older than the model's release. Models with search fill the gap by reading the web.

Does blocking AI crawlers remove me from training data?

It stops new training crawls from those bots, if they respect robots.txt. It doesn't remove what was already collected, and it doesn't stop other sites from writing about you.

Can I pay to be in training data?

No. But being written about on sites that get crawled a lot, like big review sites and forums, makes it more likely.

Go deeper

All posts →
Research

AI product recommendations are mostly a summary of listicles

We fetched every 'best X' page ChatGPT, Gemini and Google AI cited in 839 answers. 90% of the brands they recommended were on one of those lists. Brands on none were recommended 2.5% of the time. Here's how it works and what to do about it.

Make AI search your next revenue channel

Start free trial