What RAG Actually Is, for People Who Sell Machines

Someone in your market has started saying RAG in meetings. It stands for retrieval-augmented generation, and behind the acronym is a mechanism worth understanding if you sell technical products, because it decides whether an AI assistant describes the machine you actually build or a plausible average of everything it read about machines like yours.

Two kinds of memory

A language model has something like two memories, and they behave nothing alike.

The first is what it absorbed during training. Enormous, frozen, and undated. The model does not store your product page as a page. It stores statistical regularities about how words go together, which includes the shape of your industry's vocabulary and, if you were written about often enough, your company name in the company of certain products. This memory has a cutoff, and it cannot be updated by you. It also cannot be inspected: nobody can point at a spot and say your tolerance range is stored here.

The second memory is what the system fetches at the moment of answering. That is retrieval. Before writing a word, the machinery goes and gets a handful of documents that look relevant to the question, pastes them into the model's working context, and asks the model to answer using those. It is closer to handing someone a folder of printouts and asking them to summarise it than to asking them what they remember.

Retrieval-augmented generation is simply the combination: the model's absorbed memory, plus a folder of fetched documents assembled for this specific question. Whether your new product line appears in an answer usually depends on the second half, not the first.

What gets fetched, and how

Here is the part that surprises people who sell machines. Retrieval does not fetch your website. It fetches passages.

Documents are broken into chunks: a few paragraphs each, sometimes less. Those chunks get indexed by meaning rather than by exact keyword, which is why a query about pumps for aggressive media can pull a passage that never uses the word aggressive. When a question arrives, the system pulls the chunks that sit closest to it in meaning, and those chunks alone go to the model. Everything else on your page, including the sentence three screens up that gives the passage its context, stays behind.

The practical consequence is uncomfortable if your documentation is built like most industrial documentation. Imagine a product page where the model designation sits in a heading, the material appears in a table two sections lower, and the operating temperature is only in the caption of a diagram. Each of those facts may be retrievable on its own. None of them travels with the others. A chunk containing a temperature range and no product name is nearly useless to a system trying to answer which of your models tolerates that temperature.

Why the assistant describes a product you discontinued

When retrieval fails to find anything usable, the model does not stop. It answers from absorbed memory, which is where the plausible-sounding wrong answers come from: a discontinued model name, a specification that belonged to the previous generation, a certification that your competitor holds and you do not. The output reads with exactly the same confidence as a retrieved fact. There is no visual difference between an answer built from your current datasheet and one assembled from statistical residue.

This is also why the fix for wrong information is rarely to argue with the model. If the retrieved folder contains a clear, current, self-contained statement of the fact, the model has a strong reason to use it. If the folder comes back empty, the model falls back on memory, and memory is where the old version of your company lives.

What makes a passage retrievable

Nothing here requires new software, and none of it is exotic.

• Self-contained paragraphs. A passage should make sense to someone who has read only that passage. Repeat the product designation inside the paragraph that states the specification, even when it feels redundant to a human reader who can see the heading.
• Facts as sentences, not only as table cells or drawing labels. Tables are retrievable when they are real text, but a plain sentence stating material, range and standard survives chunking far more reliably.
• Text that can be fetched at all. A specification that exists only inside a downloadable file, or only behind a login, is a specification the retrieval step will usually not reach. The same applies to content that appears only after a script runs in the browser.
• The buyer's vocabulary, not only your internal one. Retrieval matches meaning, but it works from the words on the page. If your industry says slurry pump and your catalogue says only the internal series name, the two are further apart than you think.
• One page per thing. A single page carrying an entire product family gives retrieval a pile of similar chunks with no way to tell them apart. Separate pages produce separate, cleaner passages.

A test you can run today

Pick a specification that is genuinely current and reasonably specific: an operating range, a material, a compatibility. Ask an assistant that searches the web a question whose answer requires that specification, without naming your company. Then ask the same question naming your company. Then ask it to state the source.

If the answer is right and the source is your own page, retrieval reached you. If the answer is right but the source is a distributor or a marketplace listing, retrieval reached your specification through somebody else's page, which is worth knowing: they are the ones being cited, and their description of your product is the one in circulation. If the answer is wrong and no source is offered, you are looking at absorbed memory with nothing to correct it.

What this does not do

Making your pages retrievable is a prerequisite, not a lever. Nobody, including us, can guarantee that an assistant will cite you, and no supplier of any tool has privileged access to how these systems rank what they fetch. What can be done honestly is measurement: PSentry checks whether your brand appears in answers across ChatGPT, Claude, Gemini and Perplexity, in each language and market you sell in, and which sources those answers lean on. It measures and reports. It does not edit your pages, it does not touch the retrieval index of any assistant, and it does not promise that a change on your site will produce a change in an answer.

Understanding the mechanism is still worth the afternoon it costs. It converts a vague anxiety about AI into a concrete list of things about your own documentation that you can check, fix, and check again.

Frequently Asked Questions

Is retrieval just a search engine with extra steps?

It uses search, but the goal is different. A search engine returns links for a person to open. Retrieval returns passages for a model to write from, which is why the quality of the individual paragraph matters more than the quality of the page as a whole.

Does my website content get used to train the models?

That depends on the provider and on what your site allows, and it is a separate question from retrieval. Training is slow, opaque and outside your control. Retrieval happens at the moment of the answer and depends on things you can inspect: whether the page is fetchable, whether the text is real, whether the passage stands alone.

Do I need a vector database or an internal RAG system?

Not for this. Building your own retrieval system is about answering questions over your internal documents. Being cited by public assistants is about the retrieval that already happens on the open web, where your only input is the pages you publish.

Will an llms.txt file make my content retrievable?

Publishing one costs little and may help systems that choose to read it, but treat it as a hint rather than a mechanism. A well-formed HTML page with real text is what retrieval consumes today, and no side file compensates for a specification that exists only inside a scanned document.

Why does the assistant quote my distributor instead of my own site?

Usually because the distributor states the specification as plain text on a fetchable page while the manufacturer keeps it in a downloadable file. Retrieval takes the passage it can reach, and attribution follows the passage.