Knowledge Cutoffs: Why AI Knows Your Old News

A knowledge cutoff is the last date covered by the data a model was trained on. Everything the model knows without searching, its parametric knowledge, stops there.

That has a blunt consequence for brands. If you changed your pricing, launched a product, rebranded or moved headquarters after the cutoff, the model's trained memory does not contain it. Ask the model without web search and it answers from that frozen snapshot. It will often answer confidently, with the old facts.

This is a ceiling, not a setting. No publisher moves it. The vendor moves it when it ships a new model, and the latency is a full model generation. So the useful question is not how to change the cutoff. It is how your current facts reach the answer anyway.

"ChatGPT" is a product, not a single model. Different requests can be routed to different models, and those models carry different cutoffs.

OpenAI's model documentation, as verified in August 2026, lists:

• GPT-5.6 family (Luna, Sol, Terra): knowledge cutoff of February 16, 2026.
• The Instant path (chat-latest): knowledge cutoff of August 31, 2025. OpenAI describes this alias as the latest Instant model currently used in ChatGPT.

That is roughly five and a half months of difference between two paths inside the same product. OpenAI also warns that these snapshots will be regularly updated, so the dates themselves move over time.

What it means for you: two people asking ChatGPT the same question about your company can get answers built on different memories. A brand fact from December 2025 may exist in one and not the other. Any test of "what ChatGPT knows about us" that ignores routing is mixing results from different models. The details are on our page on why ChatGPT's knowledge cutoff varies.

Other vendors have their own cutoffs per model. Treat each model as its own snapshot, and check the vendor's model page before you assume a date.

There are four stages between your website and an AI answer:

1. Parametric: what the model learned in training. Frozen at the cutoff. Nobody controls it this quarter.
2. Access: whether a crawler may fetch and store your page at all. You and your CDN control this.
3. Retrieval: whether your page enters the model's context for a given question. The engine controls this, partly.
4. Generation: which retrieved pages the model uses and cites. The model controls this.

Post-cutoff facts can only travel through stages 2 to 4. The engine has to decide to search, find your page, fetch readable text and use it.

The catch is that search is not always on. As of October 2026, ChatGPT searches at the model's discretion, Claude.ai search is a user toggle, and Google AI Overviews often do not trigger. When an engine does search, OpenAI's help page says ChatGPT typically rewrites the question into one or more targeted queries. You cannot see those queries.

Say you run a hypothetical SaaS company that cut its starter price in May 2026. A search-off answer from any model with an earlier cutoff will quote the old price. A search-on answer can quote the new one, but only if the engine reaches a page that states it. Our guide to search on vs search off answers shows why the two can disagree.

Be precise about the line. The evidence is strong on access and thin on almost everything else.

Two points deserve emphasis.

First, training crawlers and search crawlers are separate. OpenAI's crawler page states that GPTBot governs training use and OAI-SearchBot controls ChatGPT search, and that each setting is independent of the others. Allowing GPTBot does not get you into ChatGPT search. Whether training inclusion changes what a future model recalls about your brand has not been measured by anyone. Our page on how brands end up in training data covers that channel.

Second, rendering matters for non-Google engines. Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but do not execute them, so they cannot read client-side rendered content. If your new price only appears after a script runs, those crawlers may never see it. Googlebot renders JavaScript, so this gap mostly affects other engines.

Work from the gates that fully block you to the ones that only help.

1. List your post-cutoff changes. Write down every brand fact that changed after August 31, 2025, then flag which changed after February 16, 2026. These are the facts a search-off answer cannot know.
2. Check robots.txt per bot. Confirm search crawlers such as OAI-SearchBot, PerplexityBot and Claude-SearchBot are allowed. A blanket AI block often removes search eligibility by accident.
3. Check the network layer. OpenAI's help page says eligibility also requires your host or CDN to allow traffic from its published search bot IP addresses. A firewall can return 403 to a crawler your robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
4. Put the facts in the raw HTML. Fetch the page with JavaScript disabled. If the price, product name or date is missing, fix the rendering for that content.
5. State each fact plainly on a page that can be fetched. A specific sentence, such as a price or a one-line definition, gives the engine something it can quote. This is the only on-page signal with positive experimental support, and it is still not a guarantee.
6. Remove accidental suppression. Look for noindex, nosnippet, max-snippet:0 or data-nosnippet on pages that hold your current facts.
7. Fix contradictions on your own site. If the pricing page, about page and footer disagree, an engine can pick the wrong one.

For wrong facts that persist after this, follow the steps in correcting outdated facts past a knowledge cutoff.

• Treating "ChatGPT" as one source of truth. Routing changes the cutoff. Record which model and which retrieval state produced an answer before drawing conclusions.
• Allowing GPTBot to fix stale answers. GPTBot is a training crawler. It does nothing for current search answers, and any training effect would land a model generation later.
• Blocking all AI bots and expecting fresh answers. Blocking OAI-SearchBot removes you from ChatGPT search answers, which is the only route your new facts have.
• Reaching for schema or llms.txt. Google states structured data isn't required for its generative AI features, and Google Search ignores llms.txt. No engine documents reading llms.txt. Neither moves a cutoff or forces a retrieval.
• Bumping dates to look fresh. A changed dateModified with no new content is self-declared, and no vendor states that page age drives selection.
• Ignoring old coverage elsewhere. If engines cite third-party pages about you, those pages shape the answer. Being named accurately on them is associated with appearing in answers, though no effect is promised. See how old news stories shape AI answers.

Run each question twice: once with search off, once with search on. The search-off answer shows the model's frozen memory. The search-on answer shows what retrieval adds, and its citations show which pages the engine used.

If the search-off answer is stale, that is expected and nothing you do this quarter changes it. If the search-on answer is also stale, something in the access chain is broken, or the engine is citing an outdated third-party page. That second case is fixable.

Control for memory and personalization. OpenAI says ChatGPT may use saved memories when rewriting a search query, so test logged out or with memory off. For a quick snapshot, the free checker on this site shows whether AI engines name you for up to three buyer questions. It is a point-in-time check, not continuous tracking, so repeat it after you ship fixes.