The model ChatGPT answers with stopped learning in August 2025
You asked ChatGPT for the best tools in your category and it named five companies that aren't you. The instinct is to publish something this week. First, work out which part of the answer a new page could possibly reach.
An answer comes from two places. One is what the model absorbed in training, frozen at a knowledge cutoff. The other is what it looks up on the web while it answers. OpenAI's model docs list the Instant model used in ChatGPT with an August 31, 2025 knowledge cutoff, and its flagship API model stops at April 30, 2026. Anthropic lists June 2026 for its newest Claude models. Anything you publish this month sits after every one of those dates.
A page gets into a model's memory only when a model trained after it ships, and no vendor publishes that schedule. So a thirty day plan is a plan for the lookup half. It's also the half you can watch.
In one audit, 409 of 481 answers never searched
OpenAI's help page says ChatGPT may search the web automatically when a question would benefit from current information. May. Its API docs are plainer: when search is offered as a tool and left on auto, search is optional.
We saw what optional means in our own data. An audit of our production runs on 26 September 2026 found that 409 of 481 answers from our ChatGPT engine, where search was offered but not required, made no web search at all. Those answers came from memory alone, so nothing published the week before could have appeared in them. We changed our engine the same day to force a search on commercial and recency questions, the kind buyers ask.
The lesson travels. Before you judge a month's work, check whether the answers you're reading searched at all. An answer that didn't search can't reflect anything you changed.
Days 1 to 3: find out which kind of absent you are
Three checks before you change a word on your site.
- Ask the category question the way a buyer would, several times, in each engine you care about. One answer is one draw.
- Then ask about your brand by name. If the engine describes you correctly, it knows you and isn't choosing you. If it gets you wrong or draws a blank, fix that before worrying about any shortlist.
- When an answer searched, write down every page it cited. Those are the pages the lookup half is reading for your category today, and they're your target list for the month.
Weeks 1 to 2: edit the pages that already rank
Google's documentation on recrawling says crawling can take anywhere from a few days to a few weeks, that a request doesn't guarantee inclusion, and that asking again for the same URL won't get it crawled any faster. That's Google Search, not a promise from any answer engine, but it sets expectations. A brand new page has to be discovered, crawled and indexed before anything can retrieve it. A page already in the index only has to be recrawled.
So the first two weeks go to pages that already exist: your category page, your comparison pages, your pricing page. Say what the product is, who it's for and what it costs in plain text a crawler can read. Request the recrawl once and leave it alone.
Weeks 2 to 6: correct what the engines get wrong about you
If the by-name check turned up a wrong price, a wrong category or a confused identity, fix the source. Start with your own about and pricing pages, then the directory profiles and third-party pages the engine cited when it got you wrong. A correction needs two things to land: the source has to change, and the engine has to fetch the changed version. Expect this work to run past day 30.
New pages are month two and three work
A page that answers a question nobody in your category answers yet is often the biggest opportunity on the list. It's also the slowest. It has to be written, published, discovered, crawled, indexed and then picked by the retrieval step for a real query. Start it in week one if you like. Don't grade it at day 30.
This is the order our own plan uses. The "Next 90 days" section of a Proofsource report puts prompts where a competitor wins and you already have the content in weeks 1 to 2, corrections to how engines describe you in weeks 2 to 6, and new content for unanswered questions in weeks 4 to 12. Those windows are rules we wrote down from the mechanism above. They aren't lead times we've measured.
Nobody can honestly promise you a lift by day 21
The guides that rank for this question are confident. One of them says structured data brings a 28 to 34% lift within 14 to 21 days and that content updates take 30 to 45 days to register. We couldn't trace either figure to a published measurement. Crawl queues and retrieval rankings sit inside the engines, and none of them publishes a lag time.
You can measure your own. Before changing anything, record a baseline on a fixed set of buyer questions in the engines you care about. Log the date each page changed. At day 30, rerun the same set the same way and compare. You'll finish the month with a lag figure for your own category, which beats anyone's benchmark.
Where this could be wrong
The memory and lookup split is cleaner on paper than inside an engine. A model blends what it remembers with what it retrieved, and nobody outside can see the weighting. A strong page can still lose to a brand the model already trusts.
Cutoffs move. OpenAI says the snapshot behind its ChatGPT model is regularly updated, so the August 2025 date is right as published today and may not be next month. The dates also describe API models, and the consumer apps can differ.
Our 409 of 481 is one audit of one engine, through the API and one model, before a fix we shipped the same day. Read it as proof that search is optional, not as the rate at which ChatGPT searches for your buyers.
Google's few days to a few weeks describes Google Search. We didn't find a comparable published figure from any answer engine, so treat it as a reference point.