The backlog of website experiments nobody has time to run is where growth goes to die. Two ex-Lindy engineers turned that backlog into a background process - and early customers are seeing double-digit lifts inside a month.
Ask any small growth team what they wish they had time for, and you will get the same answer: the tests. The landing page they meant to rewrite. The headline they suspected was weak. The checkout step that quietly loses people. The ideas are cheap and the hands are not. Somewhere between the marketer, the lone developer, and the redesign that keeps getting pushed, most website optimization simply never ships. Sherpa, a San Francisco company in Y Combinator's Spring 2026 batch, is built around that specific graveyard - and its pitch is blunt: make the website improve itself.
The product installs with one line of code. From there it watches how real visitors move through a site, spots where they hesitate or leave, and drafts on-brand variants of pages and flows. It runs those variants as A/B tests on live traffic, measures which ones win, and ships the winners. No designer opening Figma, no developer cutting a ticket, no analyst exporting a spreadsheet. The loop that normally takes weeks of coordination runs on its own, with a human approving changes before they go live.
Conversion rate optimization - CRO, in the trade - is one of those disciplines everyone agrees is high-leverage and almost nobody staffs properly. Doing it well needs a designer to build variants, a developer to wire them up, and an analyst to read the results. Most companies have none of those to spare for a test that might move a number by a few points. So the work waits. Sherpa's read is that the bottleneck was never a shortage of good ideas. It was a shortage of hands to run them.
There is a second, quieter cost. A test that takes three weeks to ship is not just slow - it is expensive to start, which means teams only run the tests they are already fairly sure about. The result is a portfolio of safe, low-variance experiments and a long list of riskier ideas that never get a chance. When the cost of running a test falls toward zero, the calculus changes. You can afford to be wrong more often, because being wrong is cheap and being right compounds. Sherpa's argument is that most of the value in CRO was always locked up in the tests teams never bothered to run.
Sherpa breaks the manual CRO pipeline into an automated loop. The interesting design choice is the last step: rather than fully hand the keys to an agent, Sherpa keeps a person in the approval seat. The agent does the labor - observing, generating, testing - and the human decides what actually goes public.
That structure maps onto three product surfaces the company describes: Experiments, which tests improvements on live traffic; Pages, which generates fresh landing pages around specific campaigns and traffic sources; and Adapt, which personalizes what a visitor sees based on where they came from and how they behave.
The approval step is doing more work than it looks. It is the difference between a tool a growth lead can actually adopt and one that legal, brand, and the founder will quietly veto. Nobody wants an autonomous agent silently rewriting their pricing page or changing a promise in their hero copy. By keeping publication behind a human click, Sherpa lets a team move fast on the parts that are safe to automate - variant generation, traffic splitting, measurement - while holding the line on the one decision people care about, which is what the public actually sees.
Practically, that means a two-person startup can run the kind of experimentation program that usually belongs to a company ten times its size. The marketer describes what they want to improve, Sherpa proposes and tests the changes, and the team reviews a queue of winners instead of a backlog of ideas. What used to be a quarterly redesign becomes a weekly habit.
The numbers Sherpa points to come from its own early users, and they are worth reading as directional rather than audited. The pattern across them is less about any single moonshot test and more about velocity - many small experiments, run continuously, compounding.
The headline case is Lindy, where the founders previously worked. Sherpa says Lindy shipped 40 experiments in 30 days and raised total site conversion by 30% - with no designer, developer, or engineer assigned to it. Emergent reported a 15% lift in visit-to-paid, InsForge a 22% bump in signups, and both did it inside the first month.
The part that separates Sherpa from a tidier A/B testing tool is its second audience. As people increasingly shop and research through AI assistants, a website has to persuade two kinds of visitors: humans, and the agents acting on their behalf. Sherpa says it tests agent-facing versions of a site so that assistants recommend a company more often. It is an early, unproven idea - but it is a coherent extension of the same loop, pointed at a new reader.
Sherpa is two people, and their resumes are the pitch. CEO Ethan Kinnan was Lindy's first forward-deployed engineer and has taught more than 10,000 people to build custom AI agents, live. He says he has automated every job he has had with AI since the GPT-3 API shipped. CTO Norbu Sonam built agent infrastructure at Lindy used across millions of tasks, ran more than 50 growth experiments there, and tripled the company's paywall conversion - and before that shipped image generation in Microsoft Copilot to tens of millions of users.
Read together, the founding story is straightforward: the experimentation machine they ran inside Lindy turned out to be something every company wanted and almost none could staff. Sherpa is that internal process, packaged as a product other teams can install.
Sherpa lands in a crowded neighborhood. The incumbents - Optimizely, VWO, AB Tasty, and a long tail of landing-page builders - mostly give teams a console to configure and read tests themselves. That still assumes you have the people to do the work. The other default competitor is simply hiring: designers, developers, and growth analysts to run tests by hand. Sherpa's wager is that an agent doing the full loop, continuously, beats both a dashboard you rarely open and a headcount most startups cannot justify.
The business is standard B2B SaaS aimed at growth, marketing, and e-commerce teams, with pricing not yet public. Sherpa raised a seed round and joined YC's Spring 2026 batch (labeled P26, and originally X26 in the founders' earliest posts). Whether the "self-improving software" ambition extends past websites is the open question - for now, the wedge is the one job teams keep postponing.
The risks are the honest ones for a company this young. Two people is a thin team to support a product that touches customers' live sites, and the reported lifts come from a handful of early users rather than an independent study. Autonomous copy generation carries brand risk that the approval step mitigates but does not erase, and the agent-facing optimization idea is a bet on a web that does not fully exist yet. Sherpa is not claiming otherwise. Its case is narrower and more testable: that the loop from insight to shipped experiment can run without the people teams cannot afford, and that running it continuously beats running it rarely.
If that holds, the interesting part is less the specific numbers and more the shape of the thing. A website that quietly gets better each week, without a release cycle or a redesign, is a different kind of asset than one that ships once and slowly ages. Sherpa is starting with the page most teams neglect because it is the easiest place to prove the loop works. Where it points that loop next is the question worth watching.
Note: Sherpa has not published a YouTube channel or product-demo video at the time of writing. A live demo can be booked through the link above.