--- title: "How API companies win developers when AI agents pick the API" description: "By being the API that AI assistants and coding agents choose and can integrate on their own, so usage starts in code and grows into enterprise contracts." canonical: "https://underneath.agency/resources/api-companies-ai-search" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # How do API companies win developer demand when AI agents choose the integration? By being the API that AI assistants name and coding agents can actually integrate without help, because for API companies the first integration decides the revenue that follows. Agents now write much of the integration code, read most of the documentation and increasingly call the APIs themselves. An API that is hard for a machine to find, read or test risks being replaced by a rival, or by custom code, before a developer ever compares the two. ## The short version 1. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-03-20-gartner-predicts-more-than-30-percent-of-the-increase-in-demand-for-apis-will-come-from-ai-and-tools-using-llms-by-2026) predicted that more than 30% of the increase in demand for APIs would come from AI and tools built on large language models by 2026. 2. In [Postman’s 2025 survey](https://www.postman.com/state-of-api/2025/) of over 5,700 developers, architects and executives, 89% of developers used AI, but only 24% designed APIs with AI agents in mind; 70% knew of the Model Context Protocol (MCP), and only 10% used it regularly. 3. When [Amplifying](https://amplifying.ai/research/claude-code-picks/report) asked Claude Code open-ended questions such as “add user authentication,” it chose Stripe for 91% of payment picks, and for email it chose Resend 62.7% of the time against 6.9% for SendGrid. 4. Machines already read payments documentation more than people: on fintech and payments docs sites hosted by [Mintlify](https://mintlify.com/data.md/), agents made 70% of requests in August 2026, up from 20.5% in February. 5. APIs are a revenue line, not a side project: 65% of organizations in the Postman survey earned revenue from their API programs, and a quarter of those earned more than half their total revenue from them. ## Which developers and teams choose an API, and what does an account grow into? Developers pick the API, usually on a free or sandbox account, and the company pays as usage grows. Developers judge a technology by its API first. In the [2025 Stack Overflow Developer Survey](https://survey.stackoverflow.co/2025/work), developers ranked an easy-to-use API as the top reason to endorse a technology and a robust, complete API second, ahead of reputation and cost. The 2024 survey found 75% of developers were more likely to endorse a technology if it provided good API access. Their endorsement matters: 48% of developers said they endorsed or influenced the purchase of new technology in their organization in the past year. The commercial model rewards that first choice for years. [Twilio](https://www.twilio.com/en-us/press/releases/Q4-full-year-2025-earnings) reported more than 402,000 active customer accounts at the end of 2025, counting any account with at least $5 of revenue in the month, and full-year revenue of $5.07 billion. Its dollar-based net expansion rate was 108% for the year, up from 104%: existing accounts spent more than they did a year before. Many accounts start small, and the money comes from the ones that grow. For API platform vendors, the companies that sell API management, gateways and tooling, the buyer is a platform or architecture team, and cloud defaults weigh heavily. In the Postman survey, AWS API Gateway was used by 47% of respondents and Azure’s by 26%, and 31% of organizations ran more than one gateway. Postman found 82% of organizations had adopted some level of an API-first approach, with 25% fully API-first. The stakes for the customer are rising too. Among organizations earning API revenue, 74% made at least 10% of their total revenue from APIs, Postman found. Choosing the wrong provider is a long-lived mistake, which is why developers test before they commit. ## At which points do assistants and agents already touch API selection? In three places: answering developers’ questions, writing integration code, and calling APIs directly. **Answering questions.** Developers ask assistants which API to use and how to use it. Postman found 41% of respondents used AI to generate API documentation, and its platform saw 7.53 million calls made to AI APIs in 12 months, up 40% year over year. **Writing the integration.** Coding agents increasingly choose and install the provider themselves. In Amplifying’s test of 2,430 open-ended prompts, Claude Code recorded what it installed, configured and committed, not just what it suggested. For how AI shortlists work in software buying more broadly, see [our article on B2B SaaS](https://underneath.agency/resources/b2b-saas-revenue-from-ai-search). **Reading the documentation.** Mintlify, a documentation platform, reports that agents made 61.87% of requests across the docs sites it hosts in August 2026. Fintech and payments was the most agent-read industry it tracks, at 70%. These are Mintlify’s own measurements of its own customers. The same agent readership shapes how [developer tool companies win users](https://underneath.agency/resources/developer-tools-ai-search) beyond APIs. **Calling the API.** Gartner’s prediction and Postman’s findings point the same way: agents are becoming API consumers in their own right. Postman’s report puts it bluntly: “Agents are already calling your APIs, with or without MCP.” In the same survey, 51% of developers cited unauthorized agent access as a top security risk. ## Which questions lead developers and platform teams to an API? Category, alternatives, comparison, pricing and compliance questions, plus requests that leave the choice to an agent. The agent requests are observed. Amplifying used real prompts that name no provider, such as “add user authentication” and “i need a database, what should i use,” and recorded what the agent chose. We wrote the developer and platform-team questions below as examples of how API choices get framed; they are not observed data: - Category: “Best SMS API for sending two-factor codes in India and Brazil.” - Alternatives: “Cheaper alternatives to Twilio for transactional SMS.” - Comparison: “Stripe vs Adyen for marketplace payouts in Europe.” - Pricing: “Which email API has the most generous free tier for a startup?” - Compliance: “Which payments APIs are PCI DSS Level 1 and support 3-D Secure?” - API platform: “API gateway that runs in our own Kubernetes cluster with OAuth and rate limiting.” The first five are usually asked by a developer and settled in code. The last is asked by a platform team and settled in a procurement process, where we infer an AI answer helps decide which vendors get an evaluation slot. ## How does a coding agent’s choice turn into API usage and contracts? Through integration: the agent picks the API, the code ships, usage grows, and the account becomes a contract. **Pick and test.** The faster an agent can get working keys, the more likely the integration completes in one session. Stripe’s [documentation for AI agents](https://docs.stripe.com/building-with-llms) tells agents to run a command “to get working API keys without signing up,” then install Stripe’s MCP server and agent skills to build the integration. That is documented by Stripe for its own product; it does not say how agents choose between providers. **Ship.** Once an API is in production code, replacing it means rewriting and retesting the integration. We infer that this switching cost is why the first pick is so valuable for API companies: the agent’s choice tends to persist. **Grow.** Usage-based pricing means revenue follows the customer’s growth. Twilio’s 108% net expansion shows how much of an API company’s growth comes from accounts already integrated. **Contract.** As volume rises, the account moves to committed pricing, security reviews and procurement. By then the comparison has usually already happened, often in an assistant or an agent, long before sales was involved. ## What decides whether an AI agent picks your API? No platform documents it; studies point to an established position in the stack and documentation agents can use. **Documented by the platform.** Anthropic describes MCP as “a universal, open standard for connecting AI systems with data sources,” which [it introduced in 2024](https://www.anthropic.com/news/model-context-protocol). No assistant or coding agent vendor publishes how it chooses between competing APIs. **Observed in studies.** In Amplifying’s Claude Code study, no other payment processor was recommended as the primary pick, though Paddle, LemonSqueezy and PayPal appeared as second choices. Email was more open: Resend won 62.7% of primary picks, while SendGrid was offered as an alternative 55 times but chosen as the primary pick only 6.9% of the time. Custom code built by the agent took 21.6% of email picks. The project’s existing stack mattered more than how the request was worded. A [2026 study of 37,927 agent journeys](https://arxiv.org/abs/2609.34951) by ora research, the company behind the agent-readiness score it tested, found that businesses whose sites agents could read were clearly recommended in 20 against 11 percent of runs, a 1.9 times gap. Its readiness score included whether pricing and API documentation were reachable. The study covered businesses in general, not API companies alone; our summary is in [our article on agent-readable websites](https://underneath.agency/resources/do-ai-agents-recommend-readable-websites). **Trust factors specific to APIs.** We infer that agents, like developers, favor APIs with complete reference documentation, a published machine-readable specification, working code samples in common languages, sandbox access, clear rate limits and pricing, and a public status page. Postman found 55% of teams struggled with inconsistent documentation and 34% could not find existing APIs, even inside their own organizations. If teams cannot find an API, a reasonable expectation is that an agent will struggle too. ## What does an API company lose when the agent integrates a rival? The integration, and every dollar of usage that would have followed it. Being well known is not enough. SendGrid, one of the best-known email APIs, appeared as an alternative 55 times in Amplifying’s email answers, yet the code the agent wrote used Resend or custom code in most cases. An alternative pick is a mention; a primary pick is an integration. Some categories also look closed to newcomers in agent answers. For payments, Stripe took 91% of primary picks and no rival took any, though the remaining picks were mock interfaces rather than other providers. We infer that challengers in such categories need to win on specific stacks, regions or use cases where the default does not fit, and to make that fit easy for agents and assistants to see. The loss is invisible in most dashboards: an agent that wires a competitor’s SDK into a codebase leaves no visit and no lead in your analytics, a blind spot covered in [why analytics miss AI visibility](https://underneath.agency/resources/why-analytics-miss-ai-visibility). ## How does GEO work for an API company or API platform? It makes your API easy to find, understand, test and trust for assistants and agents; it cannot guarantee a pick. For an API business, whether it sells the endpoints or the platform that manages them, generative engine optimization (GEO) typically involves six pieces of work: 1. **Agent-readable documentation.** Serve reference docs and guides as clean text or Markdown, publish an llms.txt and an OpenAPI specification, and keep every endpoint, error code and limit current. In our study of 5,902 top websites, only [11.5% published a valid llms.txt](https://underneath.agency/research/llms-txt-adoption-study). 2. **A path agents can complete.** Offer sandbox keys, a command-line setup and, where it fits, an MCP server, so an agent can go from question to working call in one session. 3. **Quickstarts for the stacks that matter.** Since context drives agent picks, publish working examples for the frameworks and languages your customers use. 4. **Independent technical coverage.** Earn mentions in tutorials, comparison write-ups, community answers and integration marketplaces, where assistants that search the web look for evidence. Our guide to [building authority for AI search](https://underneath.agency/resources/how-brands-build-authority-for-ai-search) goes further on earning that coverage. 5. **Clear pricing and compliance facts.** State per-unit prices, free-tier limits, regions and certifications on public pages, so assistants quote them correctly; when [our pricing study](https://underneath.agency/research/ai-pricing-accuracy-study) checked four assistants on 45 software products, only 61.9% of the plan prices they quoted fully matched the official page. 6. **Measurement across assistants and agents.** Track how often you are the primary pick, an alternative or only a mention, by stack and region, over repeated runs, and compare with new API keys, first calls and expansion revenue. For API platform vendors selling to architecture teams, the same work applies to comparison, deployment-model and compliance pages, since those are the questions that decide who gets evaluated. ## What don’t we know yet about how agents and assistants choose APIs? Whether better documentation or an MCP server raises an API’s pick rate is not yet measured. The agent-choice data comes mainly from one research group testing JavaScript and Python projects with one or two agents. Mintlify measures its own customers; ora research tested its own readiness score. Postman and Gartner describe demand and practice, not how agents choose. No published study yet links an API company’s visibility in AI answers or agent picks to new keys, usage or contract value. Treat those links as hypotheses to test against your own data. ## How can an API company see whether agents are choosing it or a rival? Run the requests developers give coding agents for your category, and record which API ends up in the code. The audit should put your category’s real developer questions and “add this feature” requests to the main assistants and agents, in the stacks and regions you sell to, and compare the results with new keys, first calls and expansion revenue. It shows where you are the default, where you are only an alternative and where an agent writes custom code instead. To scope that audit against your developer funnel and enterprise contract pipeline, [ask us for a review of your API’s visibility to agents and assistants](https://underneath.agency/contact). The [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization) page describes the technical side of that engagement, including agent-readable documentation, quickstarts for key stacks and tracking of primary picks. ## Frequently asked questions ### Do AI coding agents choose which API a developer uses? Often, when the developer leaves the choice open. In Amplifying’s tests, agents installed and configured a specific provider in response to requests that named none, and in some categories one provider won almost every time. ### Should API companies build an MCP server? It is worth testing. MCP gives agents a standard way to call your API, but Postman found only 10% of developers used it regularly in 2025, and no study yet shows that an MCP server raises how often an API is chosen. ### Is being listed as an alternative in AI answers valuable? Less than it looks. In the email category Amplifying tested, a well-known provider appeared as an alternative 55 times but was the integrated choice in only 6.9% of picks. Track primary picks separately from mentions. ### How should an API platform vendor think about AI search differently? API management is bought by architecture and platform teams through evaluations, so the questions that matter are comparisons, deployment options and compliance. Make those facts public, consistent and easy to quote. ## Sources - Gartner (2024), [Gartner Predicts More Than 30% of the Increase in Demand for APIs Will Come From AI and Tools Using LLMs by 2026](https://www.gartner.com/en/newsroom/press-releases/2024-03-20-gartner-predicts-more-than-30-percent-of-the-increase-in-demand-for-apis-will-come-from-ai-and-tools-using-llms-by-2026) - Postman (2025), [2025 State of the API Report](https://www.postman.com/state-of-api/2025/) - Nordic APIs (2025), [A Deep Dive Into the State of the API 2025](https://nordicapis.com/a-deep-dive-into-the-state-of-the-api-2025/) - Amplifying (2026), [What Claude Code Actually Chooses](https://amplifying.ai/research/claude-code-picks/report) - Mintlify (2026), [Data: agents vs human traffic](https://mintlify.com/data.md/) - Stack Overflow (2025), [2025 Developer Survey: Work](https://survey.stackoverflow.co/2025/work) - Twilio (2026), [Twilio Announces Fourth Quarter and Full Year 2025 Results](https://www.twilio.com/en-us/press/releases/Q4-full-year-2025-earnings) - Stripe (2026), [Agents and AI on Stripe](https://docs.stripe.com/building-with-llms) - Anthropic (2024), [Introducing the Model Context Protocol](https://www.anthropic.com/news/model-context-protocol) - Finder, Elovic, Shalev and Yosef (2026), [AX is the New AEO](https://arxiv.org/abs/2609.34951) - Underneath (2026), [llms.txt adoption](https://underneath.agency/research/llms-txt-adoption-study) and [software pricing accuracy](https://underneath.agency/research/ai-pricing-accuracy-study) --- This is the Markdown twin of https://underneath.agency/resources/api-companies-ai-search. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.