AI tools give outdated information about your company because most of their knowledge is frozen at a training cutoff, often a year or more in the past. Unless the assistant retrieves live sources for a given answer, it describes the company you were at training time: the old pricing, the discontinued product, the positioning you moved away from two rebrands ago.
The fix depends entirely on which surface is wrong, because some refresh continuously and some only change when a model is retrained.
Training cutoffs: why the model remembers your past self
Large language models learn from a snapshot of the web taken before a cutoff date, then answer from those frozen weights. Everything you changed after the snapshot, prices, features, leadership, even your company name, does not exist inside the base model. Worse, the model states its stale facts with the same confidence as current ones, so buyers have no cue that the description is old.
There is a second wrinkle: even before the cutoff, the model learned from whatever was most repeated. If your old positioning appears in hundreds of pages and your new one in five, the old one wins even in a fresh model.
Which surfaces refresh and which lag
AI surfaces sit on a freshness spectrum. Know where each answer is coming from before you try to correct it:
- Perplexity: retrieves and cites live sources for every answer, so it reflects the current web within days of a change
- Google AI Overviews and Gemini: grounded on Google's index and Knowledge Graph, refreshing about as fast as Google recrawls your sources
- ChatGPT search mode and Copilot: pull live pages from their search indexes when search triggers, current when it does
- Base model answers in ChatGPT, Claude, and Gemini: frozen at the training cutoff, only corrected by retraining
How to correct each surface
For the live surfaces, correct the sources they cite. Ask Perplexity about your company, open every citation, and fix or update the outdated ones you control: your site, profiles, directories. For third-party pages, request updates, and publish a current, crawlable facts page (what you do, pricing, key details) so retrieval has an authoritative source to prefer.
For base models, you cannot edit the weights. What you can do is shift the distribution the next training run will see: make the new facts the most repeated version of your story across your site, profiles, press, and third-party mentions. Consistency is the mechanism, which is why this overlaps heavily with entity optimization for AI search.
Keep it corrected
Corrections drift. New third-party content repeats old facts, and each model release resets the training-data question. Re-check quarterly with the same set of branded prompts across providers, and after any pricing change, rebrand, or major launch. A free scan gives you a repeatable readout of how each provider currently describes you, including sentiment, so drift shows up before your buyers see it.
Check what AI currently says about you
Run a free scan to see how ChatGPT, Claude, Gemini, Perplexity, and Google AI Overview describe your brand today, and catch outdated facts before buyers do.
Frequently asked questions
How do I make sure AI describes my brand correctly?
Publish one canonical, current description of your company and repeat it consistently across your site, social profiles, directories, and press materials, with Organization schema tying them together. Then correct the third-party sources AI cites, since assistants often quote those over your own site. Verify quarterly by asking each assistant about your brand by name.
Why does ChatGPT show my old pricing?
In its default mode ChatGPT answers from training data with a fixed cutoff, so it repeats the pricing that existed when its corpus was collected. It has no mechanism to notice your pricing page changed. Keep current pricing in plain crawlable HTML so search-enabled answers get it right, and expect the base model to catch up only after retraining.
Can I ask OpenAI or Google to correct facts about my company?
There is no editorial correction channel for model knowledge. Providers do not accept fact updates about brands the way a directory would. The working lever is the source layer: update the pages and profiles their systems crawl and cite, and the corrections propagate to retrieval-based answers first, then eventually into training data.
How long do AI corrections take to show up?
Perplexity and other live-retrieval surfaces can reflect corrected sources within days to weeks of a recrawl. Google-grounded surfaces follow its index refresh. Base model answers lag until the next model release trained on data that includes your corrections, which can be many months and is not announced in advance.
Related answers