Being cited by the AI is only the first level. There is a higher rung: becoming a service the AI actively uses when it has to answer a concrete request, such as checking prices, availability, or finding a specific product. Companies that expose this information in the right format don’t wait to be cited — they get integrated directly into operational answers, while competitors remain passive sources. It doesn’t require building anything complex: there are three levels of integration, and the first is within reach of any company.
So far I’ve talked to you about how to be cited in AI answers. But there is a higher level: being integrated into the answers. AI agents don’t just generate text — they call APIs, query databases, and run external tools in real time. If your business exposes an endpoint or a structured feed, the AI can use your data directly in operational answers, without going through your website.
You are no longer a passive source the AI can cite or ignore. You become an active service the AI calls.
This is the distinction that changes the level of the game. And understanding it now — before it becomes the standard — is the advantage that counts.
What’s happening in the AI research labs
One of the most comprehensive papers on the evolution of tool use in language models is recent. Haoyuan Xu et al. (2026) in The Evolution of Tool Use in LLM Agents write explicitly: “Tool use enables large language models (LLMs) to access external information, invoke software systems, and affect the real world beyond text generation.”
That sentence deserves attention. It doesn’t say the AI describes the real world. It says it affects it. An agent that books a flight, updates a stock level, sends an order confirmation — it doesn’t generate text about those actions. It performs them. And to perform them it calls external APIs.
The technical starting point is this: large language models on their own have no access to real-time data, they can’t calculate, they can’t read an up-to-date database. Tool use solves all three of these limits. When an AI agent answers a query with fresh data, current prices, live availability — it’s calling an API in the middle of generating the answer.
Minaee et al. (2025) in a survey on LLMs confirm this from a broader perspective: “More generally, an LLM can access any number of external tools (e.g. an API to a service) to augment its capabilities.” (Minaee et al., 2025)
“Any number of external tools” is the part that matters to you as a business. This isn’t a closed integration between a model and a proprietary database. It’s an open ecosystem where every service that exposes an accessible interface can become part of the AI answer.
How the mechanism works: from input to action
When a user asks an AI agent with tool use enabled a question, the flow isn’t linear like in a traditional chatbot. The model analyzes the intent, decides whether it needs external data, formulates a call to the appropriate tool, receives the result, and incorporates it into the final answer — all in a few seconds, invisible to the user.
Concrete example: “How much does a night in a 4-star hotel in Florence cost for the weekend of May 15th?” An agent without tool use makes up an estimate based on training data (which can be months or years out of date). An agent with tool use calls the Booking API or an aggregator’s API, receives real prices, updated to the second, and answers with verifiable data.
This radically changes what “showing up in AI answers” means. Until yesterday the goal was to be cited as a credible source within the generated text. Tomorrow the goal is to be the service the agent calls when it has to answer transactional or operational queries.
The difference isn’t just technical. It’s economic. An inclusion via tool use brings the user straight to your service — with your data, in your ecosystem — instead of merely mentioning you in a paragraph.
Integration via site crawling is probabilistic: the AI interprets your content, and it can get it wrong, misunderstand, or fail to find it.
The current landscape: who’s already integrated and who isn’t
Today tool use is still concentrated in a few dominant verticals. Google Flights for flights. Booking and Expedia for hotels. Wolfram Alpha for mathematical calculations. OpenAI GPT Actions for custom integrations. Anthropic MCP (Model Context Protocol) for enterprise use cases. Google Gemini Extensions for Google’s territory.
But the pattern is crystal clear: these verticals weren’t chosen because they’re “special”. They were integrated first because they had well-documented APIs, structured data, and the technical ability to meet the specifications of the AI frameworks.
The travel sector was already advanced on the API side long before LLM tool use arrived. The result is that today that category is almost fully integrated into AI agents, while other sectors — consulting, professional services, training, B2B manufacturing — are still at zero.
From this follows a deduction worth spelling out: the sectors that build their own APIs now have a window of advantage over the competitors who wait until it becomes mandatory. It isn’t permanent, but it’s real. Gao et al. acknowledge this from the risk side as well: “The development of LLM systems with access to external tools and decision-making capabilities is both exciting and concerning.” (Gao et al., 2025) — for businesses that don’t adapt, the “concerning” part is becoming invisible in operational queries.
Implement schema.org on your site if you haven’t already: Service, Offer, and FAQPage are the priority markups for anyone selling services.
Can your business become a tool?
The honest answer is: it depends on the type of data you hold and on how relevant that data is in real time.
If you sell physical products, an API with catalog, prices, and availability makes your offerings accessible to agents with live data. If you offer professional services, an endpoint with service types, price ranges, and availability windows lets the AI include you in its recommendations without interpreting your site. If you produce proprietary data — market reports, industry prices, benchmarks — a structured feed is enough to turn that data into a queryable resource.
Not every business has this immediate need. If your positioning is on informational queries — content, brand awareness, thought leadership — the path of chain-of-thought and AI reasoning remains the priority one. But if you operate on transactional or comparative queries (“which service to choose”, “how much does it cost”, “who has availability”), tool use is the level where the game is played.
The fundamental technical distinction is this: integration via API is deterministic. The data is structured, precise, up to date. Integration via site crawling is probabilistic: the AI interprets your content, and it can get it wrong, misunderstand, or fail to find it. With an API there is no interpretation. There is an answer.
Three levels of accessibility: where you are now
You don’t have to start from scratch with an enterprise API. There are three levels of progression, and each one has a growing impact on your accessibility to AI agents.
Level 1 — Schema.org and structured data. Before any API, schema.org markups (Product, Service, Offer, FAQ, LocalBusiness) are the fastest way to make your core information machine-readable. AI crawlers process them as a “pre-parsed summary” of your content. It doesn’t make you a callable tool, but it makes you far more readable — and that already counts now, in informational queries. The risks of hallucination drop when key data is structured in an unambiguous format.
Level 2 — Static JSON feed. A JSON file published on your site with services, prices, availability, and FAQs, updated periodically, is a step beyond schema markup. It isn’t real-time, but it’s machine-readable, downloadable, and processable by ingestion tools. For businesses with relatively stable data, it’s an effective solution at low technical cost.
Level 3 — Callable API endpoint. The level where you become a service that AI agents can query in real time. It requires technical development, but not necessarily a complex infrastructure. A REST endpoint documented with OpenAPI/Swagger, simple authentication, and a structured JSON response is enough for integration with the main AI agent frameworks. The multi-step planning of AI agents depends precisely on the quality and reliability of the tools it can call: the more stable and documented your endpoint is, the more likely it is to be included in execution plans.
What to do this week
- Take an inventory of your key data: which information changes frequently and is relevant to someone who has to make a decision? Prices, availability, catalog, delivery times — these are immediate candidates for structured exposure.
- Check what current AI crawling returns: use Perplexity or ChatGPT to run queries about your sector and see whether your data (prices, services) shows up correctly. If the AI gets it wrong or finds nothing, the signal is clear.
- Implement schema.org on your site if you haven’t already: Service, Offer, and FAQPage are the priority markups for anyone selling services. Product and AggregateOffer for anyone selling products. It’s level 1 — the starting point, not the finish line.
- Evaluate the AI agent integration frameworks: OpenAI’s GPT Actions and Anthropic MCP are today the main channels where businesses can register their own APIs for use by agents. Read the technical specifications and assess compatibility with your stack.
- Monitor competitors: if one of your direct competitors already exposes an API or shows up in AI agent answers on transactional queries in your sector, you have an answer to the question “should I do it?”. The multi-turn conversations of AI agents on complex queries tend to favor callable sources over merely citable ones.
The ultimate goal isn’t to have an API because it’s technically interesting. It’s to become the service the AI chooses to call when your ideal user makes the right query. That call is a customer who arrives directly to you, with the context already prepared, without going through the SERP.
Start with level 1 this week. Evaluate level 3 within the year.
