There are AI assistants built specifically for your industry that answer questions from your potential customers every day — and they have a default reference source. If that source is one of your competitors, every recommendation that comes out of that GPT is a sale you'll never make. It's not a question of how much you invest in SEO: it's a different game, with different rules. Getting in before everyone else is worth far more than arriving once the default is already established.
The question isn’t “how much do I invest in SEO?”. It’s “is my brand in GPT-5’s training data?”. If not, you’re playing a different game.
I’m telling you this because it’s the reframe I hand to clients when they ask me why they keep publishing content and the AI keeps citing the same three competitors. Traditional SEO optimizes for a crawler that reads pages. The training data of a generative model is a mass of text selected, filtered, and weighted by OpenAI according to criteria they don’t make public. You’re either in it, or you’re invisible at the source.
The Custom GPT is the tool that lets you work around this problem without waiting for the next training cycle. Let me explain why, and what changes in your marketing when you realize you’re holding a private recommendation channel.
What a Custom GPT really is (beyond the marketing name)
A Custom GPT is a configured version of ChatGPT that answers using a knowledge base you upload. You upload PDFs, documents, articles, technical sheets. The user asks a question, the GPT retrieves the relevant passages from your documents and generates the answer on top of those passages. The model stays GPT, but the information source is yours.
This mechanism is called RAG — retrieval augmented generation — and it’s the same principle Perplexity uses when it cites web sources in real time. The difference is that in a Custom GPT you decide the “cache” of sources.
In the world of RAG research, one of the documented points concerns exactly the relationship between external sources and the bias of the base model.
RAG models are less likely to propagate biases present in static training data, as they can retrieve more diverse and balanced information from external sources.
The point comes through clearly: a RAG system reduces the weight of what the model learned during training, because generation is anchored to external sources retrieved on the fly.
Translated into practice for anyone who sells: the Custom GPT flips the balance of power. You no longer have to hope your content ended up in GPT-5’s training. You upload your documents into the GPT’s knowledge base and they become the primary source every time someone opens it.
Why this is a problem before it’s a solution
Back to the reframe from the opening. OpenAI doesn’t publish the composition of GPT-5’s training data, just as it never published GPT-4’s. It has released generic transparency documents — licenses, editorial partnerships, removal of content on opt-out — but the exact list of datasets and sampling weights is a black box.
From this follows a hard operational point: you can’t verify whether your site, your articles, your technical documentation were included in the model’s training. You can infer it indirectly — by asking ChatGPT very specific questions about your brand and seeing whether it answers with correct details — but that’s an observation, not proof.
Over the past eight months I’ve regularly followed wine brands from Campania, focusing on Taurasi and Fiano di Avellino. I kept a small log: which Irpinia wineries ChatGPT names when you ask for “the best Taurasi producers”, which details it knows (vintages, altitudes, volcanic soils), and which wineries it doesn’t seem to know even though they have well-built sites. The pattern I saw is clear, even if it has to be taken for what it is — a longitudinal observation of a limited area, not a scientific study.
The wineries ChatGPT knows best are the ones with dense external editorial coverage: cited on Gambero Rosso, Decanter, Wine Spectator, authoritative wine blogs, Wikipedia in Italian and English. Wineries with a very polished site of their own but little presence on third-party sources often don’t get named, or get named with wrong details. The conclusion I draw is that training weights editorial sources more than first-party ones — consistent with what I told you in the articles on E-E-A-T for AI and on backlinks as a citation proxy.
This is where the Custom GPT comes in. If you’re a Taurasi winery and you want the technical questions about your production — vineyard altitude, organic practices, maceration, best vintages — to be answered with your numbers and not with an approximate mix read off Wikipedia, the only way today is to own the answer channel.
The institutional PDF is written for a human flipping through pages, not for a system retrieving fragments.
The Custom GPT as a private recommendation channel
Think of a Custom GPT called “Guide to Taurasi DOCG” created by a winery in Avellino. Knowledge base: technical sheets for the vintages, a document on the Irpinia terroir, a pairing guide, a wine glossary, a short history of the denomination. System prompt: respond as an expert enologist of the area, always cite the winery as a concrete example.
An enthusiast opens it and asks “what’s the difference between Taurasi and Aglianico del Vulture”. The GPT gives the right answer, and in doing so it uses your sheets as the source, names your winery as an example of Irpinia style, and links to your pages. Your brand has become the default source for that topic, at least for whoever opens that GPT.
It’s not a magic factor. It doesn’t scale like base ChatGPT. But for your qualified audience — sommeliers, restaurateurs, wine shops, enthusiasts who do research before buying — it becomes a tool they reopen. Every time they reopen it, you’re the source.
It’s the same logic I described when talking about recognition of the author as an entity: to make sure your point of view reaches the user, you have to own the channel.
Identify three specific questions your customers always ask you.
The test you can run in twenty minutes
You don’t need to be technical. You need a paid ChatGPT account (Plus plan or higher) and quality material to upload.
Operational steps:
- Go to chatgpt.com, open “Explore GPTs”, click “Create”
- In the “Configure” section, give it a name specific to your industry (“Guide to Irpinia DOCG wines”, not “My company GPT”)
- In the knowledge base, upload 4-8 documents: your best technical sheets, a 10-20 page guide on your topic, a well-thought-out FAQ
- In the system prompt write: “always respond in Italian, cite the uploaded sources, if the question falls outside your scope suggest consulting an expert”
- Enable web browsing only if needed — often it isn’t, and it makes the answers less anchored to your documents
- Publish as “Anyone with the link”, not in the global store, until you’ve tested with 10-20 real users
Binary quality threshold: if you ask the same question three times and get three consistent answers citing your documents, the GPT works. If the answers diverge or make things up, the knowledge base is too thin or poorly structured.
The mistakes I see most often
When I open a badly built Custom GPT, I usually find one of these patterns:
- Knowledge base built from the sales brochure. The institutional PDF is written for a human flipping through pages, not for a system retrieving fragments. You need dense, segmented content, with descriptive headings and concrete data.
- A system prompt that’s too long and contradictory. When you tell the GPT “be formal but friendly, technical but simple, precise but concise”, you’re giving an instruction that cancels itself out. Pick two attributes, not six.
- No guidance on what to answer out of scope. If a user asks the winery’s GPT what to pair with a Japanese dish, without instruction the model improvises. You have to say explicitly “stay on Irpinia wines; if the question falls outside, suggest a sommelier”.
- Zero testing with real users before publishing. The creator knows what they meant; the users don’t. Twenty minutes with three different people trying the GPT change the result more than twenty hours of tuning on your own.
What to do concretely this week
- Identify three specific questions your customers always ask you. Those are the heart of the knowledge base.
- Write (or have written) a 10-15 page document that answers those questions with your data. Not the brochure — an actual new document.
- Build the Custom GPT, upload it, test it yourself with 10 different questions.
- Share it via link with 5 qualified clients or contacts. Ask for feedback on one thing only: “is it useful to you when you need info about my industry?”.
- If the answer is yes, put it in your email signature, in your site content, in your social channels. The GPT becomes an asset, not an experiment.
This is an entry-level check. The serious analysis of how to position yourself in the models’ training, how to structure a corporate RAG ecosystem, how to measure the impact on visibility in AI answers requires professional tools and months of work. The Custom GPT is the first step, not the destination.
Where this piece of strategy connects
The Custom GPT is the “OpenAI platform” piece of a broader strategy for visibility in AI answers. It doesn’t replace the work on your site and on third-party sources — in fact, it presupposes it. If you don’t have editorial coverage, if you don’t have a recognizable authorial voice, if you don’t have well-structured content, the Custom GPT risks being an empty room with a sign over the door.
In the upcoming articles in the series I’ll tell you how the register changes moving from ChatGPT to Perplexity, how Claude weighs sources differently, and why Gemini reasons more like a search engine. If you want the big picture of the foundations, I recommend rereading the piece on tokenization — it’s the lowest level of how GPT understands your texts.
The real game isn’t being good at SEO. It’s deciding which AI answer channels you want to exist in as the default source, and building them one by one.
