ChatGPT Remembers Last Year

01The biggest audience asks the oldest questions

ChatGPT has more users than any other assistant, and by the measurements available it is also the least likely to name a brand in an answer. Industry research puts its brand citation rate at roughly 0.6%, against around 13% for Perplexity. When it does cite, it leans encyclopedic: Wikipedia and Reddit together account for more than a quarter of what it quotes. The engine with the most users is the one least interested in your news.

Two paths into the same answer

There are two ways your company gets into a ChatGPT answer, and they run on different clocks. The first is memory, what the model absorbed during training, collected by the crawler OpenAI calls GPTBot. The second is retrieval, pages found while the answer is being written, through the search index OAI-SearchBot builds or a live visit from ChatGPT-User. Each of those agents does a different job, and confusing them is how people fix the wrong thing. Memory is slow and sticky. Retrieval is fast and forgetful.

02You get studied long before you get quoted

On this blog, over fifteen days in August, the training crawler made 106 requests while the live fetcher made 24, better than four to one. I ran the same count again for the first half of September and got 10 against 1. The ratio is not a quirk of a small site. It is what being early looks like: the model is reading you, and nobody is asking it about you yet.

Busy sites show the opposite ratio

Across the other sites I run on the same server, the picture flips. In that same September fortnight, requests identifying as ChatGPT-User came to 10,867 against 8,093 for GPTBot. Those are declared names rather than verified addresses, so read them as the shape of the thing rather than an audit. My reading is that live fetches overtake training visits at roughly the point where people start asking about you by name. Training arrives first. Quoting arrives when people start asking.

What an announcement is worth to memory

Almost nothing, for months. A launch post published today is a fact about this week, and the memory path will not absorb it until a future training run that you do not schedule and cannot see coming. If the launch reaches an answer sooner, it gets there through retrieval, which means it has to be indexed and fetchable rather than merely published. Meanwhile the question somebody actually types is usually the old one, which tool is best for this, and memory answers it with whatever was said about your category most consistently, most often, for longest. Your launch post competes with a competitor's three year old definition, and the definition usually wins.

Long memory cuts the other way too. A price you changed, a feature you removed or a name you dropped can keep turning up in answers for as long as that model stays in use, because memory holds on to what was true when it was collected. I removed a feature from one of my products this summer, and every page that ever described it is now a sentence a model may keep repeating after the feature itself is gone. Correcting that is not a single edit. It is saying the new fact plainly and consistently until a training run catches up.

03Write the page an encyclopaedia would quote

Durable facts, stated once and the same way everywhere

The memory path rewards claims that stay true and stay consistent in every place they appear. What you are, which category you belong to, what you cost, who you are for, when you started. Say each one in a single plain sentence, and do not let your own pages, your profiles and your listings disagree with each other. When the next training run collects you, agreement is what turns a mention into something the model treats as settled. Write the sentence that will still be true when the next model is trained.

Your own pages are only part of what gets collected. Memory is built from what the lists, the threads and the encyclopaedia said about you during the window a model was trained on, which is why yesterday's point about pages you do not own matters most here. The consistency you control is the seed. The consistency strangers repeat is what sets.

Publish floors, not peaks, then test what stuck

A figure that is out of date in a month is worthless to memory and harmful when it gets repeated a year later. Round down to a number that stays defensible, which is the whole case for publishing the floor instead of the peak. Then check what the model actually kept by asking about yourself with browsing switched off, the same method as looking at your brand through ChatGPT's eyes. If memory is empty or wrong, no launch post fixes it this quarter.

None of this means ignoring news. It means sending news down the retrieval path, where it can work this week, and sending your definitions down the memory path, where they keep working for years.

Tomorrow: the engine that does care what you published on Tuesday.