Try it now: ask ChatGPT for a recommendation in your industry. If your name doesn't show up, it's not bad luck — it's that you haven't made it into either of the two channels through which ChatGPT builds its answers. It isn't a single algorithm: there are two distinct paths, and missing just one is enough to disappear. Knowing how both work completely changes your approach — and you see the results within a reasonable timeframe.
Ask ChatGPT for the “best artisan bakery in Genoa” — is your name there? ChatGPT isn’t a single algorithm: it’s a chain of technical choices you can influence.
That chain decides whether your brand gets cited when a potential customer asks a question about you, about your industry, about your products. And the interesting part is that the channels ChatGPT draws information from are two, clearly distinct, and they require different strategies.
Let me explain it with an example I just finished testing: a ceramics maker from Grottaglie, in Puglia. A hyper-local industry, where digital reputation counts as much as the work done in the workshop. And where today, if you’re not in ChatGPT’s answers, you simply don’t exist for a growing share of customers.
The two channels ChatGPT uses to build its answers
When you ask ChatGPT a question, the model does two things in parallel (or in sequence, depending on the configuration).
First: it draws from its internal memory, meaning everything that was stored during training. This is the “deep” part of the model: Wikipedia, books, dumps of authoritative sites, news articles, papers. That’s where brands and facts end up once they’ve been repeated and cited enough times to settle in as knowledge.
Second: if active, it queries the web in real time through an external search engine (in ChatGPT’s case, Bing). This is the “fresh” channel: news, updates, recently published pages.
For your brand to be cited consistently, it has to exist in both channels. One alone isn’t enough.
What the research world says
In the research on the answer architecture of large language models, this two-tier structure has a technical name: RAG, Retrieval-Augmented Generation. Gao and colleagues, in 2023, synthesized over 100 studies on the topic in a survey that became a reference for the field.
Researchers have also been exploring ways to enhance language models in the pre-training stage through retrieval-augmented techniques, advanced RAG, and modular RAG.
Translated for a business owner: the research documents that models like ChatGPT no longer rely solely on the memory “pre-loaded” during training, but integrate the retrieval of information from external sources — both during training and at answer time.
From this follows a precise operational consequence for you: to show up in AI answers you have to work on two different time horizons. The “slow” horizon (being present in high-weight sources that enter the training) and the “fast” horizon (being well indexed on Bing, with updated and structured content). Ignoring one of the two means being invisible half of the time.
Ranking first on Google does not guarantee being cited by ChatGPT.
Why the training channel is the one that weighs the most (but is the slowest)
The “training memory” channel is the one that decides the answers when ChatGPT works without active browsing. And it’s also the most persistent channel: once you’re in, you stay there for months or years, until the model’s next update.
The problem is that you don’t get to decide when you get back in. ChatGPT’s training happens at intervals that OpenAI doesn’t make public, and it includes sources with very specific weight: Wikipedia, major editorial outlets, institutional websites, academic archives. A company blog, however well done, weighs little in the training if no one cites it from the outside.
In the previous articles in this series I explained why backlinks work as a proxy for citations and why the author as a recognizable entity is one of the strongest signals. In the training channel these factors count double: the model doesn’t just see your site, it sees who talks about you and who you are in the eyes of the web.
Sign up for Bing Webmaster Tools: it’s free, and without Bing indexing, ChatGPT’s browsing channel won’t see you.
Why the Bing browsing channel is the one you can influence in weeks
The second channel — real-time browsing via Bing — has much shorter reaction times. If you publish a well-structured page today and Bing indexes it within a week, ChatGPT can already cite it when it browses for a relevant query.
Here the levers are the ones a SEO knows well: clean indexing, schema markup, updated content, domain authority on Bing. With one important difference: the AI engine doesn’t “click” the first result — it reads the top 5-10 and synthesizes them. So being first isn’t enough: you have to be citable, meaning you need paragraphs that can be extracted in a self-consistent way.
The test I ran: 10 ceramics makers from Grottaglie on ChatGPT
To understand how much the two channels actually weigh in practice, I took 10 artistic ceramics makers from Grottaglie (Puglia) — a historic district, with small and mid-sized brands, many with a modest website but a strong tradition.
On ChatGPT I ran 6 different queries: from the generic (“best ceramics makers in Grottaglie”) to the specific (“where to buy handmade ceramic pumi in Puglia”), to the product-level (“who makes traditional Apulian capasoni”).
The result: 4 brands out of 10 get cited by ChatGPT when browsing is turned off (pure training channel). The same 4 were also the only ones to have a Wikipedia entry on the district mentioning them, or articles in national outlets from the last 5 years.
With browsing active the picture changes: the cited brands rise to 7 out of 10, and makers with an updated site, clean schema markup, and fresh reviews on Google appear. The 3 that stay out have a site frozen 4-5 years ago, no structured presence, and not even Bing indexes them well.
An indicative test, not a study: 10 brands and 6 queries don’t make statistics. But the pattern is clear-cut and consistent with what I see in clients from other craft industries.
The mistakes I see most often
Working only on your own site. The classic one: “let’s rebuild the site, add structured data (Schema.org), optimize the content”. Good for the Bing channel, useless for the training channel. The training draws from outside your site.
Ignoring Wikipedia and local outlets. For local brands (bakeries, wineries, artisans, professional firms) a mention in the Wikipedia entry of the district/city or an article in a regional online newspaper is worth 10 posts on your own blog.
Expecting immediate results on the training. If you’ve just started building external citations, the training won’t see you before the model’s next update cycle. In the meantime, work on the browsing channel.
Confusing SEO and GEO. Ranking first on Google does not guarantee being cited by ChatGPT. The criteria for selecting the “citable passages” are different from those of traditional ranking.
What to do concretely this week
- Open ChatGPT and run 5 real queries about your industry, with and without browsing active. Note which brands come up in the two scenarios: those are your competitors in the AI channel.
- Check whether your brand has a Wikipedia mention (even an indirect one, in the entry for the district/city/industry). If not, that’s the first lever to study with professional tools.
- Open Google’s Rich Results Test, paste the homepage URL, and check that there is at least one valid “Organization” schema. This is for the Bing channel.
- Sign up for Bing Webmaster Tools: it’s free, and without Bing indexing, ChatGPT’s browsing channel won’t see you.
- Map 3-5 authoritative industry outlets or directories where you could earn an organic mention in the next 6 months.
This is an entry-level, honest check: it tells you where you stand. The real analysis — understanding the relative weight of the two channels for your specific brand, measuring citations over time, comparing them with the 3-5 competitors the AI cites in your industry — requires professional tools and continuous monitoring.
The thread: two channels, a single visibility in AI answers
ChatGPT’s architecture isn’t a single algorithm: it’s a chain where training memory and web browsing integrate. To show up in AI answers you have to be present in both channels, with different strategies and timelines.
In the next articles in this series on ChatGPT we look at exactly how the source selection mechanism works during browsing, which sites weigh the most in the training, and how to measure your presence over time. If you haven’t done so yet, start from the article on how tokenization works and the one on the implicit weight of citations: they’re the prerequisites for understanding why some brands get “memorized” and others don’t.
