The honest answer: it depends on the model, and the gap is almost always longer than you'd hope.
Some AI tools refresh business information in near real-time. Others are working from data that is six months to two years old. Most local business owners have no idea which category the model currently answering questions about them falls into — or that both types exist at the same time.
Training cutoffs vs. live retrieval
Large language models like GPT-4o and Claude learn from a static snapshot of the internet taken at a specific point in time. This is called a training cutoff. After that date, the model knows nothing new unless it's given a separate tool to go look things up.
For a local business, this means a model trained on data through, say, late 2024 will still recite your old phone number in 2026 — because that's what it learned, and it has no mechanism to check whether anything changed.
Retrieval-based models work differently. Perplexity, Bing Copilot, and ChatGPT with Browse enabled can fetch live web pages when answering a question. If your website and Google Business Profile show the correct hours, a retrieval-based model can surface updated information within days of the change going live — as long as it actually searches for you.
That last part matters. Most AI sessions don't trigger a live web search. ChatGPT in its default state answers from memory first. Only certain question types, or an explicit user prompt to "look it up," push the model toward retrieval.
How long the lag actually is
For pure training-data models, the gap between current reality and what the model knows runs 6–24 months. A plumber who updated their service area in March might still get routed the wrong calls in October if a model is confidently answering from stale weights.
For retrieval-based answers, the lag is shorter but not instant. Search engines need to crawl and index your updated pages first — a process that takes anywhere from a few days to several weeks depending on your site's crawl frequency. A Google Business Profile change can appear in search within 24–48 hours but may take longer to surface inside AI-generated answers that pull from search results.
One more wrinkle: even retrieval-based models sometimes cache or summarize their findings rather than re-fetching on every query. So "live retrieval" doesn't reliably mean today's data.
What types of information update faster
Not all business data ages at the same rate.
Days to weeks: Your Google Business Profile — hours, phone, address — flows into retrieval-based answers relatively quickly once search indexes it. Your website's structured data moves at a similar speed if you've added LocalBusiness schema. A news mention or press release can surface within days.
Months: Review sentiment and star ratings — models often summarize these from a training snapshot rather than fetching live counts. Service descriptions pulled from directory listings and aggregator sites that get crawled infrequently also sit in this band.
A year or more, sometimes never: Information baked into model weights from training data doesn't update at all until the model is fully retrained and redeployed. Old forum posts, cached pages, and data aggregator profiles that stopped being crawled are in this category too. So is a business's categorical identity — if the model learned a practice as a "general dentist," it can be slow to reclassify it as a cosmetic specialist even when every recent signal says otherwise.
Why a "new" model doesn't fix old data
A newly released model doesn't automatically carry current information about your business. A model that ships in mid-2026 might have a training cutoff of late 2025 — and even then, the web data ingested from that period skews heavily toward high-traffic pages and established domains. A single-location bakery in a mid-size city may appear lightly in the training data, or not at all. When that happens, the model fills gaps with assumptions or simply declines to answer.
This is a core reason why AI gets your business info wrong even when the correct information is published somewhere online — being published isn't the same as being learned.
What you can do to shorten the lag
You cannot force a model to retrain. But you can make your information easier for retrieval-based models to find, and more credible when they do find it.
Keep your website authoritative and crawlable. Name, address, phone, current hours, and an up-to-date service list should all live as readable text — not buried in an image or a PDF. Retrieval models that visit your site need to parse it without interpretation.
Add LocalBusiness schema. Structured data tells crawlers exactly what your hours, address, and service area are. This is one of the higher-leverage changes a local business can make, and it takes effect as soon as the updated page is crawled.
Update your Google Business Profile the same day anything changes. For a family dental practice or a plumbing company, the GBP is the fastest path from a real-world change to a correct AI-generated answer. Same-day edits can surface in retrieval-based responses within the week.
Publish an llms.txt file. This convention signals to AI crawlers which pages on your site are authoritative. It won't fix stale training weights, but it helps retrieval-based models prioritize your own pages over third-party descriptions of your business. Here's how it works for local businesses.
The lag only costs you money when the wrong answer reaches a real customer. A free scan shows you what major AI models are currently saying about your business — so you know whether you're dealing with a stale-training problem, a retrieval problem, or no problem at all before a potential customer gets the wrong answer first.