Can old news stories shape what AI says about my brand today?

Old stories shape AI answers through two separate channels, and each one needs a different fix.

The first is parametric knowledge. A model learns from text collected up to a knowledge cutoff, then freezes. If a 2024 article said your company was raising prices or being acquired, the model may still repeat it. It has no way to know the story aged badly.

The second is retrieval. Engines like ChatGPT search, Perplexity and Google AI Overviews can fetch live pages before answering. If an old article still ranks for a query about your brand, the engine may pull it in and quote it as current.

The difference matters because the fixes don't overlap. Retrieval responds to what is on the web now. The snapshot responds to nothing you do until the vendor ships a new model. Our guide to why AI knows your old news covers the cutoff mechanics in more depth.

Before you fix anything, confirm the answer is outdated rather than invented. An outdated answer was true once. A hallucination never was. The diagnosis guide for hallucination vs outdated answers walks through telling them apart.

When a model answers without searching, its knowledge cutoff is the only thing that governs the answer.

As of October 2026, OpenAI's model documentation lists a February 16, 2026 knowledge cutoff for the GPT-5.6 family. The Instant path, the chat-latest alias, carries an August 31, 2025 cutoff. So "what ChatGPT knows" is not one value. Routing between models changes the snapshot by months.

What this means for old news:

• Anything published after a model's cutoff can't appear in that model's no-search answers.
• A correction published after the cutoff can't reach those answers either. The old version stays in the snapshot.
• Any change arrives only with a new model generation, on the vendor's schedule.

You can't speed this up. Nothing in the evidence shows that allowing or blocking training crawlers changes what a future model recalls about a brand. The honest label is unmeasured. Treat the snapshot as a constraint and put your effort into retrieval.

When an engine searches, the pages it fetches carry most of the weight. Old articles that still rank are the main way stale news enters a current answer.

Ranking still feeds Google's AI surfaces, but loosely. Ahrefs found that 37.9% of URLs cited in AI Overviews also appeared in the top 10 results (March 2026). The rest came from further down. So an old story does not need to rank first to be cited.

Two things make this route hard to observe:

• The queries are rewritten. OpenAI says ChatGPT search typically rewrites your question into one or more targeted queries. You can't see which ones ran.
• Search is optional. ChatGPT decides per query whether to search, and Google AI Overviews often don't trigger. One question can get a retrieval answer today and a snapshot answer tomorrow.

The good news is that this route reflects the web as it is now. Change what the fetched pages say and the next retrieval can pick that up.

Two moves address the retrieval route. Neither touches the snapshot.

Publish the current fact where crawlers can read it

Stating a specific fact on a fetchable page gives an engine a sentence it can quote. Write it plainly: the current price, the current owner, the date a product was retired. Then check that the page can actually be read:

• Put the fact in the server-rendered HTML. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them.
• Make sure robots.txt allows the search bots, not just training bots. For ChatGPT search that's OAI-SearchBot, not GPTBot.
• Confirm your firewall or CDN doesn't return a 403 to crawlers that robots.txt allows.
• For Google, the page must be indexed and eligible to show a snippet.

Ask publishers to correct the story

When engines cite other sites instead of yours, what those sites say is the route into the answer. The evidence shows a strong association here, not a guaranteed effect. Say a trade site reported a funding round you later called off. A dated editor's note on that article changes what any engine retrieving it will read. It does not change any model already trained on the old version.

For your own old pages, the question on deleting or noindexing outdated pages covers the trade-offs. Google documents noindex and nosnippet as the controls that keep a page's content out of its AI answers.

Work through this in order. Each step depends on the one before.

1. Record the wrong answer. Note the platform, the question, and whether the answer showed citations. Citations mean retrieval. No citations usually means the snapshot.
2. Compare dates. If the old story predates the model's cutoff and the correction came after, expect the no-search answer to stay wrong until a new model ships. The free knowledge cutoff check helps with this comparison.
3. Fix your own pages. Put the current fact in server-rendered HTML and confirm search crawlers can reach it.
4. Contact publishers. Ask for a correction or dated update on cited articles that are still live.
5. Recheck over time. The playbook for correcting outdated facts past a cutoff sets out a schedule.

The free checker on this site shows what AI platforms currently say for up to three buyer questions, which gives you a before snapshot. Nobody can promise when, or whether, an engine will repeat the corrected version.