Why AI Assistants Do Not Know About Your New Product

A model knows things in two entirely different ways, and the difference decides whether your latest product exists as far as a buyer is concerned. There is what the model absorbed during training, frozen into its weights on the day training stopped. And there is what it fetches while it is answering you, live, from the web. The first is memory. The second is research. They fail for different reasons and they are fixed by different things.

A product you launched three months ago is not in the first one. It cannot be: the weights were finished before the product existed. So the entire question of whether an assistant can tell a buyer about it collapses into a narrower one, which is whether retrieval will find it and use it.

The cutoff is a wall, not a gradient

Every model has a training cutoff, a date after which nothing entered its weights. It is not a soft edge where recent things are known a bit less well. It is a wall. A company founded after the cutoff, a product line renamed after it, a factory certified after it: none of that is in there, no matter how much the web has said about it since.

Two things about that wall are worth knowing, and both are checkable rather than taken on faith.

The first is that assistants are usually poor at reporting their own cutoff. Ask one directly and you will often get a date that is wrong, sometimes by many months, because the model is producing a plausible sentence rather than reading a value out of a register. So do not take the answer as a specification. If you want to know what a model knows about a period, ask it about things that happened in that period and see what it produces.

The second is that the wall moves without warning. Providers ship new versions, and a question that returned nothing about your company in one release can return a paragraph in the next. That is not something you cause and not something you can schedule. It is a reason to keep measuring rather than to check once and file the result.

Retrieval is the only route in, and it is conditional

Retrieval is what closes the gap. The assistant runs a search, pulls back some pages, and writes its answer partly from those. If your launch is going to be mentioned, this is how.

The catch is in the word conditional. Perplexity retrieves on essentially every question, which is why it is the most legible of the 4 assistants and why its answers cite pages you can open. The others decide. ChatGPT and Claude reach for the web when something about the question suggests they should, and answer from memory when it does not. Nobody outside those companies knows the rule, and it changes.

So the same question about your new product can produce a confident, current, sourced answer on Monday and a blank on Tuesday, from the same assistant, because one run retrieved and the other did not. People read this as the model being unreliable. It is more useful to read it as two different mechanisms answering the same question.

Being fetchable is necessary and nowhere near sufficient

If retrieval is the route, the obvious move is to make the route passable, and that part is genuinely mechanical. The product page has to be reachable by the fetchers, which means checking that your robots.txt and your bot-management rules are not turning them away. It has to say plainly what the thing is, who it is for, and what distinguishes it, in the first screen rather than after three paragraphs of positioning. And it has to use the words a buyer would type, not the internal product name your catalogue uses.

That gets you eligible. It does not get you named. Retrieval returns a handful of pages and the model writes from them, and a single vendor page is a thin thing to write from, particularly for a question phrased as a recommendation. When someone asks which suppliers make a certain component, an assistant assembling an answer from one manufacturer's own page and four independent pages about other manufacturers will produce a predictable list.

Which is the uncomfortable part: the thing that most reliably puts a new product into answers is other people writing about it. Distributor catalogues, technical documentation, industry press, forum threads where somebody asks what to use for a given job. None of that is under your control on any schedule, and anyone who tells you they can guarantee it is selling something they cannot deliver.

What to do with a launch, concretely

Treat the first months as measurement, not as a campaign with a deadline.

Ask the buying question, not the brand question. "What is your product name" tells you nothing, because the model will find something to say and you will feel reassured for no reason. Ask what a buyer asks before they know you exist, and see whether the new line comes up at all.

Ask it in every language you sell in. A launch propagates through the English-language web at one speed and through the German or Spanish one at another, and the gap between them is usually wider for something new than for something established.

And repeat it, because a single run cannot distinguish between a model that does not know you and a run that did not retrieve. Doing that by hand across 4 assistants and 5 languages stops being feasible quickly, which is where PSentry comes in: it runs the same prompt set on a schedule, 2 scans a month, and reports where the new product is named and who is named instead. It measures. It does not make a model learn faster, and nothing does.

Frequently Asked Questions

How long until an assistant knows about my product?

There is no answer to this, and treat any specific number you are given as invented. Retrieval can surface a page within days if the question triggers a search and your page is fetchable. Entering a model's weights depends on a training run you have no visibility into. The two are unrelated timelines.

Can I submit my product to ChatGPT or Claude?

No. There is no submission form, no index to be added to, and no ranking factor to set. What exists is the ordinary web, read by fetchers, plus whatever a training run absorbed.

Should I ask the model what its knowledge cutoff is?

You can, but do not build anything on the answer. Models frequently misstate their own cutoff. Testing what it actually knows about a period is more reliable than asking what it claims.

My old products come up but the new one does not. Is something broken?

Probably not. The old ones have had time to accumulate mentions across the web and may sit inside the training data as well. The new one depends entirely on retrieval, which is the harder and more conditional of the two paths.

Does a press release help?

Only in as much as it produces pages that others publish and that fetchers can read. A release sitting on your own newsroom adds one more first-party page to a set that already had one.