Core concept

What is the knowledge cutoff, and can a new brand even be mentioned?

Last updated: September 14, 2026

The knowledge cutoff is the date a model's training data stops. Anything after it, including your company if you launched after it, is not in the model's memory. So a brand-new brand cannot be recalled from training at all: it did not exist to the model. It can only appear if the model runs a live search and lands on a page that names you. Two things make this sharper than the spec sheet implies. First, the date you actually get is earlier than the one advertised: Anthropic lists Claude Haiku 4.5 as trained on data through July 2025 but reliable only through February 2025, a five-month gap between what the model saw and what it knows well. Second, coverage right before the cutoff is thin, because recent events are underrepresented in the text a model trains on. The fix for a post-cutoff brand isn't to wait for the next model. It's to get onto the searchable third-party pages the retrieval step reads, now.

Jul '25 → Feb '25Claude Haiku 4.5: training data cutoff vs. reliable knowledge cutoff, a 5-month gap (Anthropic docs)

Oct 2023GPT-4o's stated knowledge cutoff. A brand that launched in 2024 is invisible to its memory (OpenAI docs)

0ways a post-cutoff brand appears from training. Live retrieval is the only door.

What the cutoff actually is

A model learns from a fixed snapshot of text collected up to some date. That date is the knowledge cutoff. After it, the model knows nothing from its own memory, no news, no launches, no new companies, until it is retrained and a new snapshot is taken. OpenAI publishes a hard date for each model: GPT-4o's is listed as October 1, 2023. Anthropic publishes one per model in its models overview. These are real, checkable dates, not vibes.

The blunt consequence for a new brand: if you launched after the cutoff, you are not in the model's head. Full stop. You can have a great product, a real website, happy customers, and still be a total blank to a model that was trained before you existed. This is time-based invisibility, and almost nobody flags it when they sell you "AI visibility."

The trap: the effective cutoff is earlier than the printed one

Here's the part that surprises people who read the spec sheet and stop. The date a model saw data and the date it reliably knows that data are not the same. Anthropic makes this distinction on its own docs page, with two separate columns. Claude Haiku 4.5: training data cutoff July 2025, but "reliable knowledge cutoff" February 2025. Anthropic's own note explains the reliable cutoff is "the date through which the model's knowledge is most extensive and reliable," while the training cutoff is just "the broader range of data used." Five months of data went in that the model doesn't know well.

Why the gap? Text about a period keeps getting written for months and years afterward. Analysis, retrospectives, roundups, wikis. So an event from early in the training window is described in thousands of documents the model saw; an event from the last few weeks before the cutoff is described in almost none yet. The model's knowledge doesn't stop at a clean line. It fades out, and the fade starts well before the printed date.

Research backs the fade, and it's uneven

This isn't just a vendor footnote. A 2024 study, "Dated Data: Tracing Knowledge Cutoffs in Large Language Models" (arXiv), probed models to find their effective cutoff per topic and found it routinely lands earlier than the reported one, and varies resource by resource. Some subjects freeze months before the official date; others track closer to it. There is no single crisp line, there's a jagged edge, and it sits inside the stated window.

How we got here

Anthropic models overview, 2026: publishes two columns, training data cutoff and reliable knowledge cutoff, and for Haiku 4.5 they differ by five months (Jul 2025 vs Feb 2025). The vendor itself says the later data is there but not reliable.

"Dated Data," 2024 (arXiv:2403.12958): measures the phenomenon independently. Effective cutoffs are generally earlier than reported, and uneven across topics, because recent periods are thinly represented in training text.

Put together: one is a vendor admitting the gap on its spec sheet; the other is outside research measuring the same gap across models. They agree, from opposite incentives, that "the model knows everything up to date X" is false. It knows things up to X unevenly, and knows the last stretch before X barely. For a brand, that means being mentioned once, recently is close to worthless for the training path, you need to have been described widely and for a while to be solidly in memory. The honest caveat: exact effective dates vary by model and topic, so treat the printed cutoff as a ceiling on what's reliably known, not a promise.

So can a new brand be mentioned at all? Yes, but only one way.

If training memory is closed to you, retrieval is your only door, and the good news is it's an open one. When the model runs a live search, it reads pages as they are today, cutoff be damned. A company that launched last month can be quoted in an answer this afternoon, if it's sitting on a page the retriever pulls. That is the entire game for anyone newer than the models they want to be found in.

Which means the post-cutoff playbook is not "wait for GPT-6 to learn about us." It's:

Situation	Training path	Retrieval path
Launched before the cutoff, mentioned widely	You can be recalled from memory	Also works
Launched before the cutoff, barely mentioned	Thin or absent in memory	Your realistic door
Launched after the cutoff	Impossible, you didn't exist	Your only door

Get onto the third-party pages a search for your category returns: the roundups, the listicles, the coverage, the forum answers. Not for a backlink. For the mention that a live retrieval can land on. Do that and your launch date stops mattering, because you're being read, not remembered.

The one thing to stop doing

Stop treating a model's memory as a target you can influence quickly. You can't. It's frozen behind a date you don't control, and even the part before that date is uneven. Pour the effort into the searchable web instead, the part that updates in days. This is the same lesson as retrieval-versus-training, seen from the calendar: the frozen path is closed to newcomers, and the live path doesn't care when you were born.

Not sure whether AI even knows you exist yet? Pressfront's free AI Visibility Report runs the live engines against your category so you can see the gap, then go get on the sites AI reads.

Related entries

Prerequisite Retrieval vs training: which one decides whether AI mentions you?
Explains Why is my brand invisible to AI even though I rank on Google?
Related What is co-occurrence, and why do mentions beat backlinks for AI?
Related How do I write a first paragraph that AI will lift verbatim?
Compare with How consistent are AI recommendations, really?

Changelog

September 14, 2026: First published. Built on OpenAI and Anthropic's published cutoff dates and the "Dated Data" arXiv study, read together into the training-versus-reliable-knowledge gap and the post-cutoff playbook.

Sources (primary): Anthropic, "Models overview," Claude documentation (training data cutoff vs. reliable knowledge cutoff), 2026; OpenAI, GPT-4o model documentation (knowledge cutoff Oct 1, 2023); "Dated Data: Tracing Knowledge Cutoffs in Large Language Models", arXiv:2403.12958, 2024. Cutoff dates are model-specific and change with each release; this entry is dated and revisited.