Key takeaways
- Search is still the default way developers look for answers (82.7% in Stack Overflow's 2026 survey), but asking an AI agent is close behind at 69.9%. Discovery is now a two-channel game.
- Coding agents are where AI use lives: 65.9% of respondents use coding assistants or agents, and 73% of those users use them daily. Claude Code (65.5%) and GitHub Copilot (58.7%) lead the agents people name.
- Only 12.2% of developers say they search Stack Overflow to verify AI output. Most run the code locally (76.5%) or compare it to their codebase (63.9%).
- ChatGPT software citations skew toward docs, vendor sites, GitHub and review sites. Stack Overflow does not show up in the top domains, and social posts barely register.
- If you ship a dev tool, the work is the same boring stuff done well: crawlable docs, a live GitHub presence, complete review profiles, and a way to see what agents actually say about you.
How developers actually look for answers in 2026
Stack Overflow's 2026 Developer Survey (30,903 responses from 169 countries) asked a plain question: where do you go when you need an answer? The numbers are worth reading slowly.
| Where developers look | Respondents | Share |
|---|---|---|
| Online search tools for general guidance and discourse | 13,774 | 82.7% |
| Ask an AI agent | 11,643 | 69.9% |
| Go straight to a maintained repository and search it | 7,288 | 43.7% |
| Message a trusted human directly | 4,803 | 28.8% |
| Post in a group chat or private discussion | 2,462 | 14.8% |
Two things jump out. First, nobody has abandoned search. Four out of five developers still start with it. Second, 70% asking an AI agent is a huge number for a behavior that was marginal three years ago. These are not mutually exclusive channels, of course. People use both, often in the same hour.
The repository number matters more than it looks. Nearly 44% go directly to a repo and search inside it. For tool discovery, that means your README, your examples folder and your issue tracker are part of your marketing surface whether you like it or not.

Inside the company, humans still win
When the question is about internal context rather than public tools, the picture flips. Coworkers or teammates are the top source (71.9%), followed by code repositories and comments (62.7%) and internal docs (60.1%). Official docs come in at just 30.5% in that context.
The pain points are familiar: information is incomplete (63.2%), important context lives in people's heads (60.9%), and it is outdated (56.5%). None of that is new. What is new is that 49.7% of respondents already use AI to search, summarize or interpret internal knowledge, and another 36.1% expect to do more of it.
From chatbots to coding agents
The 2025 survey was the first to ask about agents; 31% said they used them. Stack Overflow's April 2026 pulse survey put it at 59%, and Claude Code's usage rose from 41% to 55% in that window. The full 2026 survey shows where things landed.
| AI tool category | Use it | Daily users among them |
|---|---|---|
| AI coding assistants or coding agents | 65.9% | 73.0% |
| General-purpose AI chat tools | 62.5% | 65.6% |
| AI agents or automated workflows | 26.2% | 63.1% |
| Internal AI tools built by my company | 17.8% | 47.7% |
| Don't use AI tools | 17.2% | n/a |
So most developers who use AI use it every day, and the biggest category is not the chat window, it is the agent that lives in the terminal or editor. That is a different discovery context from a browser tab. An agent that is mid-task picks a library, a CLI or an SDK and moves on. There is no results page to optimize for.
Which agents developers name
The agent leaderboard from the survey, top ten:
| Agent | Respondents | Share |
|---|---|---|
| Claude Code | 8,030 | 65.5% |
| GitHub Copilot | 7,196 | 58.7% |
| OpenAI Codex | 3,620 | 29.5% |
| Cursor | 3,127 | 25.5% |
| Google Antigravity | 1,956 | 16.0% |
| Gemini Code Assist | 1,789 | 14.6% |
| Open Code | 1,662 | 13.6% |
| JetBrains AI | 1,569 | 12.8% |
| Windsurf | 796 | 6.5% |
| Warp | 641 | 5.2% |
Percentages add up past 100 because people use several. Claude Code leading Copilot is the headline. Quartz summed it up well: an early start gave Microsoft nothing permanent, and neither company holds a safe lead. Compare that with the 2025 survey, where ChatGPT led at 81.7% and Copilot followed at 67.9%. A year is a long time here.
If you want to try the two most-named options, GitHub Copilot and Claude are both in our catalog.

Organizations are not standardizing as fast as the individual numbers suggest. In the survey's workplace question, 32.8% say they use a few approved tools or models, 18.6% mostly one, and 17.3% let teams choose their own. Another 12.5% are still trying several before deciding. For a vendor, that means procurement-level decisions about agents are unsettled, and bottom-up developer choice still carries weight.
What developers ask agents to do
The survey breaks AI use down by task. The strongest current use is generating code in a familiar area (69.3%), then debugging, troubleshooting or refactoring (63.8%), then answering straightforward technical questions (59.4%). Generating code in an unfamiliar area is at 56.0%, which is the one most relevant to discovery: when someone works outside their comfort zone, the agent picks the tools.
A few other rows worth noting:
- Understanding or onboarding to a new codebase: 49.7% use AI now, 35.1% expect to use more
- Architecture or technical design exploration: 47.4% now
- Answering complex or ambiguous technical questions: 50.0% now
- Deploying or troubleshooting production systems: only 19.9% now, and 36.0% say they do not plan to
That last one is a useful reality check. Agents are everywhere in the editor and still unwelcome near production.
Trust, but verify (locally)
Here is the part I find most telling for anyone who cares about where tool recommendations come from. When developers get AI-generated output, what do they do next?
| Verification step | Share |
|---|---|
| Run it locally | 76.5% |
| Compare against codebase context | 63.9% |
| Inspect tests and security implications | 53.0% |
| Read official docs | 38.0% |
| Search Stack Overflow | 12.2% |
| Use as-is | 9.6% |
| Ask a coworker | 6.9% |
Only 12.2% go to Stack Overflow to check the AI's work. That is a striking number from the site's own survey, and to its credit they published it. The verification layer has moved from a Q&A site to the developer's own machine, plus official docs for 38%. If your docs are the thing people check against, they had better be right.
Also, 9.6% use output as-is. Small, but across a big population that is a lot of agent-chosen dependencies nobody reviewed.
What AI search cites when developers ask about software
Chat agents that browse the web still cite sources, and those sources shape which tools get recommended. Promptwatch Data analyzed ChatGPT citations for organic software prompts (comparing tools, finding alternatives, asking how a product works) from Sep 28 to Oct 4, 2026.
| Domain | Share of software citations |
|---|---|
| github.com | 1.2% |
| learn.microsoft.com | 0.98% |
| g2.com | 0.73% |
| docs.aws.amazon.com | 0.41% |
| docs.cloud.google.com | 0.35% |
| gartner.com | 0.34% |
| capterra.com | 0.32% |
| intercom.com | 0.28% |
| zapier.com | 0.24% |
Grouped by source type across the 23 domains charted (7.89% of all software citations), docs and help centers earned 2.64%, vendor sites 2.01%, review sites and analysts 1.57%, GitHub 1.2% and app stores 0.47%. Vendor sites plus docs total 4.65%, about three times the share of review sites and analysts.
Two absences stand out. Stack Overflow does not appear in that top-domain list, and Reddit is not among the top software domains in the chart either. The long tail is wide, so the shares are small, but the ordering is clear: documentation and code hosting beat opinion and Q&A when ChatGPT answers software questions.
Social is a Google story, not a ChatGPT story
Social citations differ sharply by engine. In Promptwatch's November 2026 social citation report (Oct 11 to Nov 9), social platforms made up 13.16% of Google AI Overviews citations, 1.94% of Perplexity's and 0.64% of ChatGPT's. Within AI Overviews, YouTube alone was 6.71%, more than the other six platforms combined. Reddit was 0.40% of ChatGPT citations, back as its top social source after the sharp drop covered in Promptwatch's Reddit citations report, where reddit.com fell from about 4% to 0.5% on August 14, 2026.
Note the caveat: the October comparison covers seven days and November covers 30, and the Reddit rebound is a share of all ChatGPT citations, not just software prompts. Still, the direction is plain. For developer tools, a YouTube walkthrough can matter in Google's AI answers, while a Reddit thread or an X post is unlikely to be what ChatGPT quotes.
The shift in who does the choosing
Stack Overflow's own framing is that every individual becomes almost a manager of agents. I think that is half right. Developers still decide, but their shortlist is increasingly assembled by something that read your docs, your repo and a few review pages in a few seconds.
The satisfaction data adds an odd texture. 45% of respondents describe themselves as complacent at work and 33% as not happy. The State of Devs 2026 survey (5,463 respondents, July to September) found nearly half had experienced job insecurity and around two thirds felt less motivated and more cynical recently. Tired, nervous developers lean harder on tools that save time. That is a tailwind for agents, and it also means patience for bad documentation is at a low.

Also worth a look is the rise of solo operators: those working in an organization of one jumped from 4% to 10%. Solo builders choose tools fast, rarely have procurement, and ask agents a lot. They are a discovery audience of their own.
What this means if you build a dev tool
The data does not call for a new playbook. It calls for doing the old one carefully, because both humans and agents now read the same materials.
1. Make your docs readable by machines
Docs and help centers are the single largest source type in ChatGPT's software citations. Keep them server-rendered, indexable and current. Check that AI crawlers are not blocked by robots.txt or your CDN. Promptwatch's AI crawler URL inspector and Agentic Readiness Check are free ways to test this, and the LLMs.txt generator is there if you want to publish one. Be honest about what that file does: it helps agents, and markdown files make up just 0.05% of AI search citations according to Promptwatch Data.
2. Treat GitHub as a distribution channel
GitHub is the top cited domain for software prompts at 1.2%, and 43.7% of developers search within repositories directly. A clear README, working examples, tagged releases and responsive issues all feed that.
3. Keep review profiles complete
G2, Capterra and Gartner together earn a visible slice of citations. They are not glamorous, and they are the cheapest fix on this list.
4. Verify, then run local
Because 76.5% of developers run generated code locally before trusting it, quick starts matter. If your install takes forty minutes and a config file, the agent's suggestion dies at the first error. Copy-paste quickstarts that work are discovery assets.
5. Measure what agents say about you
You cannot manage what you cannot see. Tools like Promptwatch track how ChatGPT, Gemini, Perplexity, Claude and others mention and cite a brand, and they also cover coding agents such as Codex, Claude Code and OpenCode, which is the part dev-tool companies tend to forget. IP-verified AI crawler logs show whether bots can actually read your docs.

If you want to compare options before committing, the GEO software directory at bestgeosoftware.com lists the wider field. A few other monitors in our catalog:
Profound

| Need | What to look at | Free starting point |
|---|---|---|
| Are AI crawlers reading my docs? | Crawler logs, robots.txt checks | AI crawler URL inspector |
| What does ChatGPT say about my tool? | Prompt tracking and sentiment | Free AI visibility report |
| Which pages get cited? | Citation reports, page tracking | Promptwatch Data reports |
| Do coding agents recommend me? | Coding agent visibility | None, needs a platform |
Where the survey data runs out
A few honest limits. The survey measures what developers say they do, not what agents do on their behalf. Percentages on the agent table come from the subset who answered the question, and people can pick several tools. Citation shares from Promptwatch are small per-domain numbers drawn from one week of software prompts, so treat the ordering as more reliable than the decimals. And nobody has solid public data yet on which libraries coding agents pick when no human is looking. That is the gap I would watch in 2027.
Daily.dev's ecosystem survey adds one more wrinkle: some developers (23%) do not plan to try new AI tools next year. Discovery by agent does not reach everyone, and search plus word of mouth will keep carrying plenty of the load.
Bottom line
Dev tool discovery in 2026 runs on search, agents and repositories at once. Four out of five developers search, seven out of ten ask an agent, and most verify by running code rather than reading a forum. Vendors that publish accurate, crawlable docs and a healthy GitHub presence are cited most. Everything else is secondary.
For more categories of software to explore, browse the software directory at surferstack.com.

