ChatGPT no longer just recommends brands: it can directly query a company's systems while talking to a customer — prices, availability, bookings. If your competitor has connected their systems and you haven't, the AI doesn't just cite them: it uses them. Being cited is no longer enough. Becoming the supplier the AI queries directly is the next level, and the companies that do it first build an advantage that's hard to catch up with.
The user asks ChatGPT about your brand. The AI remembers. Then they ask about a competitor. The AI connects the dots. The memory feature changes the game: every interaction with a customer is a permanent opportunity.
This is the symptom my most advanced clients have been seeing over the past few months. ChatGPT no longer just answers: it remembers previous conversations, builds a profile of the individual user and, in certain cases, calls your service directly inside the chat. The leap is huge. For years the GEO goal has been “getting cited.” Now there’s a level above that: getting the AI to query your API while it’s talking to the customer.
Actions inside OpenAI’s Custom GPTs are today the technical mechanism that opens this door — the operational legacy of the old ChatGPT Plugins, which OpenAI retired in April 2024 and replaced with exactly this Custom GPT plus Actions architecture. Here I’m not going to talk about “how to program it”: I’m going to talk about what changes for your visibility in AI answers if you decide to expose an Action, and what you lose if you don’t while your competitors do.
What an AI model’s memory is (and why it concerns you)
In the world of language model research, memory isn’t a technical detail: it’s the piece that makes conversations continuous and personalized. A 2025 survey defines it very clearly.
In the era of large language models (LLMs), memory refers to the ability of an AI system to retain, recall, and use information from past interactions to improve future responses and interactions.
Translated for you: an AI model’s memory is the ability to retain, recall and use information from past interactions to improve future responses. It’s not a cosmetic feature, it’s a structural piece of how ChatGPT and the other assistants are evolving.
The operational consequence for your business is simple: if a user has already interacted with your brand inside ChatGPT — maybe by asking for information, comparing you with a competitor, reading a piece of your cited content — that trace stays. And when the model also has an active channel to your API (a plugin or an Action), the memory is no longer passive. It becomes a direct access point to your catalog.
Actions: the difference between being cited and being used
Up until 2023, the GEO game was played on citations: appearing in the sources the AI lists at the end of an answer. It’s still 90% of the work, and in my previous articles I’ve explained how to build that signal: from the tokenization of your content to author entity recognition to the backlink as a citation proxy.
Actions are a different plane (the legacy Plugins did the same job before the April 2024 retirement, so the concept isn’t new, it’s just been rebranded inside Custom GPTs). They’re not there to get you cited: they’re there to get the AI, while it answers a user, to call your API to retrieve live data. If you run an e-commerce, the AI can check availability and price. If you run a travel agency, it can search for packages. If you manage a B2B catalog, it can show the updated spec sheets.
The memory survey explains why this is consistent with the evolutionary path of the models.
Moreover, the introduction of memory enables LLMs to retain historical interactions with users and store contextual information, thereby providing more personalized, continuous, and context-aware responses in future interactions.
Translated: memory lets models retain past interactions and context, producing responses that are more personalized and continuous over time. From this it follows that an AI with memory + access to your API no longer “suggests” you to the user: it integrates you into the flow. The recommendation becomes a transaction.
If the AI doesn’t know you, the endpoint stays switched off.
The case of the livestock farm in the Polesine
Let me tell you about a concrete situation I followed over the last 6 months. A farm in Rovigo, in the Polesine area, raising Friesian and Romagnola cattle and producing short-supply-chain fresh cheeses and dairy. Direct sales, local large-scale retail, a few starred restaurants in Veneto and Emilia.
The owner had a precise problem: chefs were searching for information on “producers of fresh dairy with a certified supply chain in Veneto” inside ChatGPT and Perplexity, but they always found the same 4 or 5 big names. He, with 200 head and a limited output, never showed up.
We set up a 6-month strategy on two parallel tracks. The first track was the classic authority work: technical content on the breed, structured product sheets with an Organization schema verified on Google’s Rich Results Test, a Wikidata entry for the company, mentions in Veneto and Emilia agri-food trade outlets.
The second track — the one I’m telling you about today — was the exposure of a custom Action: an endpoint that answers three questions that are fundamental for a professional customer. Product availability today, production batch, traceability of the source animal. Three pieces of information a chef wants to verify before ordering.
After 6 months, the result I measured on a sample of 40 test queries on ChatGPT and Perplexity (an indicative test, not a rigorous study): the brand shows up in 22 out of 40 answers on the general queries “short-supply-chain dairy Veneto,” and in 3 cases — with users who had activated third-party plugins — the AI directly queried the endpoint to answer “yes, available today.” Three out of forty is little in absolute terms, but it’s a kind of result that the other 4 or 5 cited competitors don’t have.
The operational lesson I take home is that plugins and Actions don’t replace the citation work: they amplify it on a level the competitors aren’t yet covering.
When you decide to expose an Action, write the OpenAPI description as if it were editorial copy for a model: clear, with examples, no jargon.
The test you can run in 20 minutes
Before thinking about plugins and Actions, you need to know whether the AI knows who you are well enough to make exposing an endpoint worthwhile. Here are three binary checks.
- Open ChatGPT and run 10 typical queries from your industry (“best [product] in [your province],” “[category] producers with [characteristic],” “comparison between [competitor 1] and [your category]”). If your brand doesn’t appear in at least 2 or 3 answers out of 10, you’re not ready for an Action: you first need authority work.
- Check on Wikidata whether an entry exists for your company. If not, the model has few anchors to latch onto you. The Wikidata entry is free but has to be done according to the community’s rules.
- Compare yourself with the 3 to 5 competitors the AI cites in your industry: run the same 10 queries, note who appears and how many times. That gap is the real measure of the work to do before investing in an Action.
These are entry-level checks. The real analysis, the one that drives a 6-month strategy, requires professional tools and structured monitoring of the queries.
The mistakes I see most often
Over the past few months I’ve seen four recurring patterns in companies that ask me to “build a ChatGPT plugin” (outdated terminology: today we talk about Actions inside Custom GPTs, the actual Plugins were switched off in April 2024).
The first mistake is building the Action before having the citation. If the AI doesn’t know you, the endpoint stays switched off. No user explicitly asks “use the Action of [unknown brand]”: they only activate it if the brand already surfaces in the organic answers.
The second mistake is exposing too many functions. An Action that does 15 things confuses the model. Better an endpoint with 2 or 3 very clear functions: “check availability,” “calculate a quote,” “show traceability.” The model calls what it understands well more reliably.
The third mistake is forgetting to maintain the OpenAPI description. The model decides to call your API by reading the description. If it’s written in tech-speak, it won’t invoke it. The description has to be written thinking that the first “user” is the model, not the developer.
The fourth mistake is neglecting the authority signal around the Action. An endpoint connected to a brand with little authority gets used little: the model prefers to call endpoints of brands it recognizes. It’s the same principle as implicit reference weight: the more recognizable the brand, the more its channel gets used.
What to actually do in the next 90 days
I’m not telling you to build a plugin tomorrow. I’m telling you to get ready for the moment when it’ll make sense to do it.
- Map the 20 queries in your industry where you want to appear. Check today where you’re positioned on ChatGPT, Perplexity, Gemini.
- If you don’t appear, work for 4 to 6 months on the previous level: structured content, entity recognition, mentions on authoritative sources in your industry.
- If you appear but only occasionally, ask yourself which function of your service the AI might want to call. Availability? A quote? A verification? Three at most.
- When you decide to expose an Action, write the OpenAPI description as if it were editorial copy for a model: clear, with examples, no jargon.
It’s not a magic factor. It’s not enough on its own. But it’s the kind of investment that, within a coherent AI-answer visibility strategy, moves you from “cited among the others” to “integrated into the answer flow.”
Where it fits in the GEO journey
Actions are the last mile of a journey that starts much earlier. You first need an inverted pyramid in the content that the model knows how to read. You need AI-specific E-E-A-T signals. You need your brand to be recognized as an entity in the Knowledge Graph.
Actions arrive once that ground is prepared. In the next articles in this series I’ll show you how the AI platforms differ from one another: what changes between ChatGPT, Claude and Perplexity in terms of citation mechanisms, and how to calibrate the strategy when your ideal customer searches on one engine rather than another.
The thread always stays the same: being visible in AI answers. But the level changes. From cited, to used.
