Two years ago, half of the products I track sent their emails through SendGrid. But for the ones that launched this year, it droped to 1/8. Resend went the other way, from 41% to 82%.
It doesn't surprise me, though. What surprises me the most is: when I sampled recent GitHub pull requests titled "Firebase to Supabase", "Vercel to Cloudflare" and so on, half of them were opened by a coding agent. A year and a half ago it was almost zero.
Some context first. For a while now I've been cataloguing what shipped products actually run on. It's at 7,302 products, and every "uses X" links to where it was found: a script on the live site, a DNS record, a dependency in a public repo, or the maker saying it themselves. It's all on howmakersbuild.com if you want to check my work. Most of this post comes out of that data, plus a GitHub crawl.
What new products pick
The trap with "most popular tool" lists is that old tools win by default. They've had more years to collect users. So I split products by when they launched (on Product Hunt) and compared the share of each tool among products launched in 2026 with the share among products launched in 2024 or earlier.
| Decision | Launched ≤2024 | Launched 2026 |
|---|---|---|
| Transactional email | SendGrid 51%, Resend 41% | SendGrid 13%, Resend 82% |
| Database | Supabase 33%, Firebase 19%, MongoDB 17% | Supabase 64%, Firebase 9%, MongoDB 5% |
| Product analytics | PostHog 52%, Mixpanel 30% | PostHog 88%, Mixpanel 9% |
| Web analytics | Plausible 42%, Umami 6% | Plausible 22%, Umami 25% |
| Hosting | Cloudflare 8%, Railway 2% | Cloudflare 21%, Railway 8% |
| Payments | Stripe 74%, Polar 2% | Stripe 65%, Polar 11% |
Note: Shares add up to more than 100% in places because plenty of products use two of something.
A few notes, since half of these are search terms people actually type:
- SendGrid alternatives. If you're looking, the new-product crowd has already voted. 82% of 2026 launches that send email use Resend. Postmark dropped from 25% to 10%, which I didn't expect. (email API breakdown)
- Supabase vs Firebase. Not close anymore for new projects. Firebase halved, MongoDB lost nearly three quarters of its share. (database decision)
- PostHog vs Mixpanel. PostHog is now the default in 9 of 10 new products that track product analytics. (product analytics)
- Stripe is still the default for payments, but merchant-of-record services (Polar, Dodo) are eating the edges. Solo makers apparently don't want to deal with sales tax in every country they sell to. Fair.
I left out mobile on purpose. React Native "drops to 0%", but that's because apps built with Expo get counted as Expo, and Expo is React Native. Data is fun like that.
I expected AI to pick differently. Mostly they don't.
About 1,460 products in the dataset were built with an AI coding tool or app builder (Claude Code, Cursor, Codex, Lovable, Bolt and friends), either because the maker said so or because the repo has the tool's config in it.
Compare them with everyone else and the gap looks huge: Resend 69% vs 40%, Supabase 66% vs 39%, PostHog 82% vs 60%.
Then I remembered that 850 of those AI-built products launched this year. Of course they look like 2026 products. So I compared only products launched in 2026, AI-built vs not:
| 2026 launches | Built with AI tools | Others |
|---|---|---|
| Resend | 86% | 79% |
| Supabase | 71% | 59% |
| PostHog | 92% | 86% |
| Polar | 12% | 11% |
| Stripe | 68% | 64% |
Most of the gap disappears. Supabase keeps a 12 point lead (78% among apps built with Lovable, Bolt and the like, which makes sense, since some of those builders wire Supabase in for you). Everything else is within a few points.
So "AI is picking your stack" is only half true. Everyone launching in 2026 picks roughly the same stack, whether or not an agent wrote the code. My guess is the same tutorials, the same templates, and the same AI answers feed both groups. Nobody gets their defaults from nowhere. Ryan Carniato wrote a good piece about this a couple of months ago, Nobody Argued For Your Stack, on how conclusions now spread without the arguments behind them. I mostly agree with him. This is the data side of it.
Where AI clearly changes things: leaving
Picking a stack was never the hard part. Getting off one was. That's where agents show up.
I counted GitHub pull requests titled " to " over the last two years. A few of the biggest flows:
| Migration | PRs | The other way |
|---|---|---|
| Firebase → Supabase | 297 | 113 |
| Vercel → Cloudflare | 242 | 83 |
| Next.js → Astro | 178 | 48 |
| Prisma → Drizzle | 140 | 27 |
| Next.js → TanStack Start | 92 | 9 |
Across all the flows I track, there were 253 of these PRs in the first half of 2025 and 2,838 in the first half of 2026. That's 11x in a year. GitHub got bigger in that time, but not 11x bigger.
Then I pulled the newest PRs for each flow (1,628 of them, leaving out migrations between AI tools and model APIs, since those are a different story) and checked who opened them. I counted a PR as agent-made if it came from an agent's branch (claude/…, codex/…, cursor/…), an agent's bot account, or had a "Generated with" or "Co-Authored-By" line naming one.
- First half of 2025: 11%
- Second half of 2025: 41%
- First half of 2026: 48%
- Since July: 44%
That's a floor, not an estimate. Anyone who squashes commits or renames the branch is invisible to me. About 7 in 10 of the agent PRs were Claude, then Copilot, Cursor, Codex, Devin and Jules.
Some flows are almost entirely agent work now. In this year's sample: Express → Fastify 14 of 20, Netlify → Cloudflare 13 of 20, React → Astro 11 of 20.
What I think happens next
These are guesses, labelled as guesses.
- Defaults get even stickier for new projects. When a template or an agent reaches for Supabase, Resend and PostHog, most people won't argue. Expect the 2027 numbers for the top pick in each category to be higher, not lower.
- Lock-in stops being a moat. If moving costs an afternoon, "it's a pain to migrate" stops protecting anyone. Tools will have to keep you on price and reliability. The question shifts from "how fast can I set this up" to "what does this cost me at 10k users".
- Being readable by agents becomes distribution. 51% of the active product repos I track already have an AGENTS.md or CLAUDE.md. For repos created this year it's 72%. A tool whose docs an agent can follow without guessing will get picked, and migrated to, more often.
- Model APIs stay the least settled layer. Claude → Gemini: 108 PRs. Gemini → Claude: 104. People switch models like they switch tabs.
Caveats
- Shares are of products I can see evidence for, not market share. Hosted services show up on websites; backend libraries mostly in code. How the evidence is collected.
- The "≤2024" group is mostly 2024 launches (1,313 of 1,824).
- Migration counts are PR titles across all of GitHub, not just these 7,302 products, and only PRs titled that way round.
- The agent sample leans recent (newest 20 per flow), and early 2025 is thin: 37 PRs.
- About 450 PRs failed to fetch, mostly rate limits.
All of this is browsable: every decision, every migration flow, and what agent instruction files look like in practice.
Did an agent pick your stack, or did you? And if you've moved off something this year, did you write the migration yourself?
