• AI visibility

Guide · AI search

How do AI coding tools win developers through AI search?

By being the tool an AI assistant names when a developer asks what to use, backed by the public evidence developers check before they trust it: documentation, benchmarks, security facts and community threads. Developers already work inside AI tools all day, so the recommendation often happens in the same window as the code. Most revenue then comes from seats, as one developer’s trial turns into a team and then a company-wide license.

By Underneath Team Guide Updated

The short version

  1. Use is near-universal but trust is falling. In the 2025 Stack Overflow survey (opens in a new tab), 84% of respondents use or plan to use AI tools, yet more developers distrust their accuracy (46%) than trust it (33%).
  2. The market is concentrated. Among developers who use or build AI agents, 82% had used ChatGPT and 68% GitHub Copilot as ready-made assistants; Copilot (opens in a new tab) passed 20 million all-time users, and Cursor (opens in a new tab) crossed $1 billion in annualized revenue.
  3. Revenue grows by seats: Cursor’s plans (opens in a new tab) run from $20 a month for an individual to $40 per user per month for teams, and Gartner expects (opens in a new tab) 75% of enterprise software engineers to use AI code assistants by 2028.
  4. AI answers already send developers to products. ChatGPT accounted for around 10% of new signups at Vercel (opens in a new tab), a developer platform, up from 1% six months earlier.
  5. A coding tool that ChatGPT knows by name still may not be suggested. In Sharma’s test (opens in a new tab) of 112 Product Hunt startups, using questions such as “What are the best AI coding assistants launched in 2025?”, ChatGPT recognized 99.4% by name but surfaced only 3.32% in discovery questions.

Who pays for AI coding assistants, and how big can an account get?

Developers choose them first; engineering leaders and procurement buy the seats once a team depends on them.

The buying motion starts bottom-up. An individual developer tries a tool on a free or personal plan, such as GitHub Copilot’s (opens in a new tab) Free, Pro ($10 a month), Pro+ ($39) or Max ($100) tiers, or Cursor’s $20 individual plan. If it sticks, the team asks for a shared plan, and the company eventually signs for every engineer, adding single sign-on, data controls and central billing.

That is why a single developer is worth far more than one subscription. Cursor charges teams $40 per user per month, so a 200-engineer organization is a different deal from one developer on a personal plan. Microsoft reported that Copilot is used by 90% of the Fortune 100 and that its growth among enterprise customers rose about 75% from the previous quarter. Anthropic (opens in a new tab) said Claude Code was generating over $500 million in run-rate revenue within months of its full launch in May 2025. AI voice companies see a similar pattern, where a developer starts small on an API and spending grows with usage, as how voice companies turn AI answers into revenue shows.

The enterprise wave is still forming. Gartner’s survey of 598 respondents found 63% of organizations piloting, deploying or already using AI code assistants, and it forecasts adoption among enterprise engineers rising from less than 10% in early 2023 to 75% by 2028. Each organization that standardizes on one tool decides where hundreds or thousands of seats go.

Where do AI assistants sit in how developers find tools?

Everywhere: developers ask AI for answers daily, then check documentation, GitHub and community threads before they commit.

Developers are heavy users of the assistants that now recommend tools. Stack Overflow found 51% of professional developers use AI tools daily, and that developers who mostly use AI in their workflow frequently use it to search for answers or learn new concepts. JetBrains’ 2025 survey (opens in a new tab) found 85% of developers regularly use AI tools for coding and 62% rely on at least one AI coding assistant, agent or code editor.

They do not stop at the answer. The Stack Overflow results summary (opens in a new tab) shows developers rely on a portfolio of community resources: Stack Overflow (84%), GitHub (67%) and YouTube (61%). People learning to code still use technical documentation more than any other resource (68%). And 75% said they would still ask another person for help when they do not trust AI’s answers.

Some vendors can already see AI answers in their signups. Vercel itself traced around 10% of its new signups to ChatGPT, up from 4.8% the previous month. Developers are also asking more than one assistant, and ChatGPT’s lead is narrowing: its share of generative AI website visits went from about 76% in June 2025 to roughly 53% by May 2026 in Similarweb’s 2026 data (opens in a new tab), as Gemini and Claude grew.

Distribution matters as much as discovery. GitHub’s Octoverse 2025 (opens in a new tab) reports that nearly 80% of new developers on GitHub use Copilot within their first week. A tool built into the platform where developers already work starts with an advantage that a challenger must earn through recommendations. The same contest plays out for DevOps tools engineers ask AI about.

Which questions do developers ask about AI coding tools?

Technical ones: which tool fits a stack, how it scores, its cost per seat, and what happens to the code.

These coding-tool prompts are our own examples of what developers and engineering managers ask; they were not observed in real logs:

  • Comparison: “Cursor vs GitHub Copilot vs Claude Code for a large TypeScript monorepo.”
  • Workflow fit: “Best AI code review tool that comments on GitHub pull requests.”
  • Evidence: “Which coding agent has the highest SWE-bench Verified score, and how was it run?”
  • Security: “Which AI coding assistants don’t train on our code and support self-hosting?”
  • Cost: “What would AI coding assistants cost per seat for 200 engineers with single sign-on?”
  • Migration: “Is it worth switching from Copilot to Cursor for a Python data team?”

Answers to these questions hinge on hard facts that assistants can get wrong. Coding assistants rename plans and change usage credits often, and in our pricing study, only 61.9% of the plan prices four assistants quoted for 45 software products were fully faithful to the vendor’s pricing page. Benchmark claims age quickly too, as each model release reshuffles the public leaderboards (opens in a new tab).

How does one developer’s trial grow into a company-wide seat license?

Through a trial that spreads: one developer tries the tool, the team adopts it, and procurement buys seats.

Named. An answer to a comparison or workflow question puts your tool on a developer’s list. With free tiers common, the cost of trying is a few minutes. AI video tools run a similar path from free plan to upgrade.

Individual trial. The developer installs the tool, often from an editor marketplace, and tests it on real work. This is where the product either earns trust or is dropped; Stack Overflow found 66% of developers frustrated by AI solutions that are “almost right, but not quite.”

Team plan. Developers who keep using a tool ask for it at work. Seat prices such as Cursor’s $40 per user per month turn a personal choice into a budget line.

Enterprise license. Security, legal and procurement review the vendor. Data handling is a deciding question. GitHub, for example, documents that it does not use Copilot Business or Enterprise data to train its models, while individual subscribers’ interaction data may be used unless they opt out.

AI search can enter at more than one step. A developer may first hear of a tool in an answer, and an engineering manager may later ask an assistant to compare the three tools the team is already using. We infer that the second question matters as much as the first, because it is where hundreds of seats are decided.

Why does an assistant suggest Copilot, Cursor or a newcomer?

Platforms document little; studies point to independent coverage and developer communities; developers add hard evidence and security.

Documented by the platform. ChatGPT, Claude, Gemini and Copilot say nothing public about how they pick which code assistant to suggest. GitHub’s own plans page describes Copilot as “the competitive advantage developers ask for by name,” a vendor claim, not an observed result.

Observed in studies. In Sharma’s Product Hunt study, startups whose own sites scored higher on GEO-style content checks were no more likely to be surfaced, while the number of other sites linking to them and community discussion predicted visibility in Perplexity. Across US software ranking questions in the work of Chen and colleagues (opens in a new tab), AI search drew 72.7% of its sources from independent “earned” sites, against 45.4% for Google, so outside coverage counts for more in AI answers than a coding tool’s own site. In our Reddit study, Google’s AI Overviews cited Reddit in 35.4% of B2B software answers, and specialist communities supplied 58.7% of the Reddit threads shown in software searches. In our freshness study, four assistants cited pages first published about half as long ago as Google’s top 10 for the same questions (a ratio of 0.50), which matters in a category that ships new models every few weeks.

Trust factors specific to coding tools. Developers are skeptical by habit, and the evidence gives them reasons:

  • Measured productivity. In a randomized trial by METR (opens in a new tab) with 16 experienced open-source developers working on 246 real issues, developers took 19% longer with AI tools, though they expected a 24% speedup and afterward believed AI had sped them up by 20%. Results from one setting in early 2025, but developers know this study.
  • Security of generated code. Veracode (opens in a new tab) tested over 100 models and found 45% of code samples failed security tests. Vendors that sell code security face their own version of this, set out in our article on application security vendors.
  • Overall sentiment. Google’s 2025 DORA report (opens in a new tab) found 90% of respondents use AI at work, while 30% report little or no trust in the code it generates.
  • Transparent benchmarks and data policies. We infer that public, reproducible benchmark results, clear training and retention policies and honest limits give both developers and assistants something checkable to cite.

What does a coding assistant lose if AI answers skip it?

The trial, and with it the team and enterprise seats that grow from it.

The market leaders already have distribution that does not depend on answers: Copilot inside GitHub, ChatGPT as the general assistant most developers use. Among developers who use or build AI agents, 82% had used ChatGPT and 68% Copilot as ready-made assistants. A challenger that the assistants do not name has fewer ways to reach the developer who is about to try something new.

Seat economics raise the stakes. Once a team has standardized on a tool, configured it and built habits around it, expansion revenue flows to that vendor, and switching means a new security review. That is our inference from how these products are sold; no study has yet measured AI-driven switching between coding tools. For the general mechanics of missing new products, see our article on why ChatGPT misses new products.

What does GEO look like for a code assistant or coding agent?

It gives assistants and skeptical developers evidence they can check; no coding tool can be promised a recommendation.

For a coding assistant, generative engine optimization (GEO) breaks into seven jobs:

  1. Documentation that machines can read. Public, current docs and quick-starts, available as clean text. Our llms.txt study found only 11.5% of top websites publish a valid llms.txt file, and our agent-readable web study found 3.2% return Markdown when an agent asks. Our article on what AI agents do when they cannot read your site explains why it matters.
  2. Benchmarks with methods. Publish results on public benchmarks with the setup, date and model version, and link the raw runs. Developers and assistants can check reproducible numbers; they discount bare claims. Research tools face a similar test on proof of citation accuracy.
  3. Community presence on the record. Real answers in GitHub issues and discussions, Stack Overflow, specialist Reddit communities and developer YouTube, from your engineers, not marketing accounts.
  4. Honest comparison and migration pages. Fair comparisons with named alternatives, including where they are better, and step-by-step migration guides. We look at whether such pages earn citations in our summary on comparison pages.
  5. Security and data pages. What you train on, what you retain, where data is processed, self-hosting options and certifications, in plain words on public pages.
  6. A pricing page an assistant can quote. Seat prices, usage credits and overage rules in one current place, with old plans retired.
  7. Repeated checks where developers ask. Rerun a fixed set of comparison, workflow, security and price questions in ChatGPT, Claude, Gemini, Perplexity, Copilot and Google, and ask each new workspace how it heard of you. The broader software picture is in our B2B SaaS article.

What is still unknown about AI answers and coding-tool adoption?

Whether an assistant’s answer decides which coding tool a company standardizes on; nobody has measured it.

The adoption figures come from developer surveys with different samples and questions, and the revenue figures are company-reported. Vercel’s signup share is a single developer-platform example, not a coding-assistant figure. The productivity and security studies measure the tools, not how assistants recommend them. No published study follows an AI recommendation from a developer’s first trial to an enterprise seat count, and we have not tested how assistants answer coding-tool comparisons ourselves. Ways to judge whether that work pays off in seats are covered in our article on business results.

How can a coding tool vendor see whether AI answers are costing it seats?

Check what assistants say when developers and engineering leaders compare you with tools they already use.

A practical first step is an audit of comparison, workflow, security and pricing questions across the main assistants, matched against where your trials, team plans and enterprise seats come from, so the gaps that cost the most seats are fixed first. To see that comparison mapped against your trials and team plans, ask us for a coding-tool answer review. What the follow-on work covers, from docs and benchmark pages to developer community presence, is laid out on our generative engine optimization service page.

Frequently asked questions

Do benchmark scores decide which coding tool an AI assistant recommends?

Not on their own, as far as anyone has shown. No assistant documents how it chooses. Benchmarks help when they are public, recent and reproducible, because they give independent sites and developers something to cite.

Should coding tool companies publish comparison pages against Copilot or Cursor?

Yes, if they are fair and specific about where each tool is better. Developers distrust one-sided pages, and assistants draw mostly on independent sources, so your own page works best alongside third-party comparisons.

Do developer communities like Reddit and Stack Overflow affect AI answers?

They appear in them. In our research, Google’s AI Overviews cited Reddit in about a third of software answers, mostly from specialist communities. How much any community shapes ChatGPT or Claude answers is not documented.

How can a developer tool see whether signups come from AI search?

Read three signals side by side: referrals from AI assistants, a question at signup or in the editor extension about where the developer heard of you, and regular checks of what assistants say for your main comparisons. Vercel’s experience shows the share can rise quickly once you start measuring it.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work