You’ve renamed the company, moved, or changed what you lead with. You might update the website first, then replace the business cards and the sign, and consider the job done. Open ChatGPT at that point and ask it, using the old company name, where the company is and what it does, and you might still see the old address, along with the product line you stopped leading with long ago.

Suppose you run a metal stamping plant that moved early this year from Shulin in New Taipei to an industrial park in Nantou. The company changed its name from “Example Hardware” to “Example Precision”, and its main line switched from furniture hardware to automotive parts. The new website is live, with the new address, the new name and the new product line all clearly set out. One day a buyer asks ChatGPT which factories in Taiwan make stamped automotive parts, and you aren’t in the answer. Remembering your old name, the buyer asks “Where is Example Hardware?” instead, and ChatGPT gives the old Shulin address and says your main business is furniture hardware.

Updating your website replaces just one of the pages AI holds on you

The old version the buyer heard comes mainly from two places. To put it as an analogy, AI holds two sets of information about you: one is the impression it formed from what it read while it was learning, stored in its memory; the other is the stack of web pages it looks up on the spot when it answers a question.

Updating your website replaces just one page in that stack. The copy stored in its memory hasn’t changed by a single word, and in the stack, the old articles other people wrote about you, the company details nobody updated on directory sites and the old pages you forgot to take down are all still there.

The copy stored in memory stops at the knowledge cutoff

Start with the copy stored in memory. AI like ChatGPT can answer questions because the company that built it first had it read and learn from a vast amount of text on the web. That process is called training, and a trained AI is what the industry calls a model. The date up to which the training data was collected is that model’s knowledge cutoff, and nothing that happened after it is in the model’s memory.

Take OpenAI, the company behind ChatGPT: in its official documentation for engineers, it lists this date on its model pages. For one of its own models, GPT-5.5, the date given is 1 December 2025. The stamping plant only moved to Nantou early this year, so the move isn’t in that model’s memory.

We infer that for this copy to change, the AI company has to retrain on new data and release a new model; how far the new model’s data goes, and whether it takes in your move, only becomes visible once it’s out.

Even then, the new model won’t necessarily remember the new you, because what it remembers depends on how the web described you on its cutoff date. The plant’s old name has been building up online for more than a decade, in news stories, directories and customers’ supplier lists. The new name has only just appeared, and the only place it’s written may be your own website. Our assessment is that if pages carrying the old name still outnumber pages carrying the new one at the next round of training, the new model may well still remember the old version.

The copy it looks up online holds more old pages than you remember

The copy it looks up online doesn’t wait for training. At the end of October 2024, OpenAI added search to ChatGPT, and its own description is that ChatGPT will choose to search the web based on what you ask. Each search picks up web pages afresh, which makes it sound as if, once you update your site, it should find the new version next time.

The catch is that you don’t decide which pages it picks up. Your new website is just one of them, and lying alongside it is everything left over from the Shulin days. What you remember is probably the old job ads on recruitment sites. What you don’t remember may be more, such as the PDF catalogue a sales rep sent out, which a customer then posted on their own website. Every one of those pages carries the old address, the old name and furniture hardware, and any of them could be picked up by AI and used to describe you.

After a rename, the old name’s signal is stale and the new name’s signal is too weak

Once AI has this information, two things can happen. The signal here simply means what’s written about you online.

The first is a stale signal: what AI receives is the way you used to describe yourself, your old address, the products you no longer lead with. The “Shulin” and “furniture hardware” the buyer heard were all true once. By the standards of the time, it hasn’t got anything wrong; it has simply been handed your old version.

The second is a weak signal: there’s too little written about you, and it’s too scattered, so AI can’t assemble a complete description and fills the gaps with fragments and things that don’t belong to you. Where the signal is weak, the noise is loud, for the same reason a radio crackles most when the station is faint. After a rename, the weakest signal of all happens to be the new name. Only a handful of pages mention “Example Precision”, and all of them appeared this year. Unless AI links it to “Example Hardware” as the same company, all it can do is piece together a new company from those few pages, and it may describe you as a small plant founded this year, or even mix you up with a different firm with a similar name. It also helps explain why the buyer who asked about stamped automotive parts at the start didn’t hear your name: there’s still very little online saying you make automotive parts.

So after a rename, the way AI describes you may split in two. Ask with the old name and you hear the out-of-date you, which is the stale signal; ask with the new name and you hear a pieced-together you, or it simply says it doesn’t know you, which is the weak signal. Together we call these signal distortion: the you that AI describes isn’t the you of today. How they come about, AI companies haven’t published in any detail; what’s above is our own inference from the knowledge cutoff and from the way ChatGPT searches the web.

Our approach treats both at once, from two sides: on one side, fewer pages still carrying the old version, so there’s less outdated signal; on the other, more websites mentioning your new name, so the signal gets stronger.

You often can’t find all the old pages AI is picking up

Start with the side about fewer old pages: getting the pages that still show the old name and old address corrected or taken down. To correct or remove a page you first have to find it, yet the old pages are often impossible to find in full, because you can only search using the clues you remember. Search for “Example Hardware” and you’ll find the pages carrying that full name, but the old pages don’t always write the name that way: a directory may list an even older short form, and on forums people call it “Mr Chen’s place in Shulin”. Customers put all sorts of questions to AI, AI searches based on those words, and what it picks up can be an old page using any of those forms of the name; when you search yourself, you only find the forms you thought of.

The ones you do find aren’t necessarily the ones AI is actually picking up, so even deciding which pages to deal with first is hard to judge on your own. The old pages you remember and the ones AI actually picks up may not overlap at all: the old news story you care about most may never have been used by AI, while the page AI keeps picking up may be on a directory site you’ve never heard of. We haven’t found any AI company that tells you, in a list, which pages it’s actually picking up. Different AI don’t necessarily pick the same pages, and the same one may switch to a different set after a while.

Old pages you can’t take down can only be covered over by the new version

Finding a page doesn’t mean you can take it down, because many of these pages aren’t yours to remove. Some directory sites let you request a correction, and some have long since stopped being looked after. With the old catalogue a customer posted on their own website, or other people’s articles and forum threads, all you can do is politely ask for a change, and they’re under no obligation to make it. A news story reports what was true at the time and was never going to be rewritten for you. The ones you can’t take down only stand a chance of being covered over once your current name and address are written up on other websites more often than they are.

If the new version is only on your website, you’re the only one saying it. The new name needs to appear on other people’s websites, and AI will usually need to read somewhere that the old and new names are the same company before it can link them up. Asking a regular customer or someone you know to write a post or two is likely to fall well short against an old name that built up over more than a decade in Shulin. Which places to write it, and how much counts as enough, depends on where customers in your line of business go to look things up and whom AI cites when it answers questions about that line of business, and every line of business is different. And the new name has to keep being written up on other websites; one round of website updates won’t be the end of it.

When AI will change its answer: we couldn’t find a timetable

Once there’s less outdated signal and the new signal is stronger, how long does it take AI to change its answer? As of October 2026, we couldn’t find a single AI company that has published this, nor any academic research that measures it. The copy in its memory and the copy it looks up online are each unpredictable in their own way. The copy stored in memory, as we inferred above, depends on when the AI company retrains, and we haven’t found a public schedule for that either. With the copy it looks up online, which pages it picks up can differ every time: it might find your new website this time and an old directory the next.

So among ChatGPT, Gemini and Perplexity, which has switched to the new version and which is still stuck on the old one is something you only find out by going back and asking; and a correct answer this time doesn’t mean it will still be correct next time. A rename or a move never reaches a day when it counts as done. All you can do is go back and ask every so often, and check whether AI is describing the business you are now.

Can’t find them all, can’t take them all down, can’t predict the change: how managed GEO splits the work

Old pages you can’t all find, pages you can’t all take down, and no telling when AI will change its answer: together that makes long-term work you won’t finish in one go. So who does it? geoweb.tw’s managed GEO service means our team fixes your site’s content, brings what’s said about you elsewhere into line, and regularly checks how each AI engine describes you; you just review the results. The GEO in the name, short for generative engine optimisation, means adjusting your website and the information about you online to suit the way AI looks things up and writes its answers.

Taking those three in turn: we list the different ways your old name has been written together with you, using what you remember, then work through each AI engine’s answers to track down, page by page, the old pages it’s picking up. Of the ones we find, those that can be fixed, we request corrections for or ask the owners to change; those that can’t, we cover with new articles and descriptions, so the new name gets written up more often than the old one. We can write those articles. If you’d rather write them yourselves, that’s fine too: send us the drafts, and we’ll rework them into something AI can read clearly before they’re published on your website and on other sites. Whether AI has changed its answer is what the regular checks are for.

Those regular checks mean putting the questions customers ask to the major AI engines, such as ChatGPT, Gemini and Perplexity, at regular intervals, and looking at whether they mention you, whether they recommend you, and whether they get you right, with every round recorded. Unlike asking once on your own phone, comparing the record over time is what shows whether AI has really changed its answer. We call this AI citation backtesting (it measures first-recommendation rate, recommendation rate, mention rate, answer accuracy, sentiment and message alignment). For a company that has renamed or moved, the most relevant of these are answer accuracy and message alignment: answer accuracy looks at whether facts such as your address and phone number are wrong; message alignment looks at whether, when AI says what you do, it uses your current story (automotive parts) or the old one (furniture hardware).

To find out which name AI knows you by right now, you can confirm it in a demo: a session with one of our consultants where they show you live. The consultant asks ChatGPT and Gemini a few questions the way your customers would, once under your old name and once under the new one: where the company is, what it does, how to reach it. Wherever an answer still gives the old version, or treats the two names as two different companies, the consultant points it out, one question at a time; if you haven’t changed your name yet, they can ask using your current name, so you can see which version of you AI is describing. Your website is the one page entirely under your control, and the new name and address need to be written there in a way AI can follow first, so after the questions the consultant puts your site through a GEO audit. The GEO audit is geoweb.tw’s free website audit, which checks how well AI can understand your site and scores it across 12 areas.

Further reading