Startup Spotlight: Riveter turns Abby Grills' manual research into a live data API

Why it matters

Riveter is trying to turn web research from a recurring manual chore into a reusable data workflow. Its early spreadsheet-selling process gave founders Abby Grills and Cody Watters real customer tests, while the ChatGPT and Codex integrations put that workflow closer to where users make requests. The business depends on making the results reliable enough for repeat use, not merely impressive in a demo.

Abby Grills spent weekends checking customer spreadsheets by hand before she tried to automate the work. That early routine became the product thesis for Riveter: let a user describe a dataset, then send software to find, extract and organize the information across live websites. On September 15th, 2026, Grills said in Riveter's announcement that Riveter had become available in ChatGPT and Codex, putting that workflow inside tools users already use to ask questions and build software.

Grills' path to the product ran through payroll and business identity, not the web-scraping industry. She joined Gusto as an early product manager and later led mid-market payroll and products including Time Tracking and state payroll-tax registration. At Middesk, she worked on address-risk products. Her Y Combinator profile describes a decade in product and a string of zero-to-one launches. Those jobs gave her a close view of the gap Riveter targets: useful facts are often public, but collecting them in a consistent form takes work that few teams want to own.

The spreadsheet came before the software

Riveter's early sales process was a manual service wrapped around a product still being built. Grills wrote that the founders sold spreadsheets of data, asked prospective customers for sample datasets, and ran tests for them. Some took tens of hours. She spent weekends inspecting the results and working with CTO Cody Watters to adapt the system.

That work exposed the hard part of the problem. A web-data tool has to find the right sources, cope with different website formats, and return output reliable enough to use. The founders could have filled gaps manually for each customer. Instead, Grills said, they used those tests to learn what the application needed to handle. She says Riveter won work with Snapchat, Antler and Sunco by selling data it could find and structure when alternatives fell short. The commercial lesson was unusually direct: customers paid for the finished information, and each paid job became a test of the software.

Watters had spent more than five years at Gusto before moving to Retool, where he worked on the IDE and related product areas. He and Grills met at Gusto. In a 2025 post announcing his move to Riveter, Grills said they had known each other for eight years; Watters described joining after two and a half years at Retool. The pairing fits the product's two demands: Grills' experience defining and shipping workflow products, and Watters' engineering work on software used to build software.

The initial insight was shared. In Riveter's YC launch description, the founders said they had watched engineering, data, operations and growth teams struggle to turn online information into structured datasets. They pointed to manual research, contractors and months spent building scrapers as the alternatives. That framing still describes Riveter's pitch, though the product has broadened from a strategy-research copilot into a general web-data API and dataset-building system.

Search is only the first step

Riveter's central distinction is the distance it tries to cover between a question and usable rows in a table. A conventional search tool returns links. A scraper extracts information from a chosen page or site. Riveter says its agents can search, navigate pages and extract requested fields, then return the results in a structured format. Riveter also says agents can read PDFs and images, while users can enrich data they already have or build a new list from a prompt.

That breadth matters for the use cases Riveter highlights: competitor prices, product specifications, lead lists, business-risk research and public filings. The public web rarely presents those facts in one clean database. Details may sit on different pages, documents or image-based menus, and the set of companies or products a customer cares about may not match a vendor's predefined categories. Riveter is betting that an agent can do enough source discovery and interpretation to cover that gap without requiring a custom scraper for every job.

The workflow covers more than one model call returning a clever answer. Riveter's API documentation separates dataset building from enrichment: one operation finds records, while a later step can fill in requested attributes. Developers can use saved enrichment configurations for repeatability, or generate an extraction plan from a natural-language prompt. Production workflows need fields, rules and outputs that behave consistently across thousands of rows; a one-off research answer can tolerate some variation.

Riveter's docs also describe scheduled monitors that rerun enrichments daily, weekly or monthly and can send results to a webhook. That turns collection into an ongoing operation: a team could check competitor prices repeatedly or refresh records as websites change, rather than run a one-time scrape and leave the dataset to age. A workflow like that shifts the burden from maintaining site-specific code to controlling what the agent should find and how it should return it. Riveter still has to earn confidence that it can find the right source and produce dependable results.

The integration with ChatGPT and Codex is a distribution move as much as a product extension. In her September announcement, Grills described users asking for a dataset in plain language, with Riveter searching the web and returning structured results in the conversation. She listed examples from collecting NBA player information to tracking a basketball score every minute. Those examples demonstrate the range of tasks she wants users to imagine; they are not reported customer performance results.

Riveter also publishes a self-serve pricing page, with a free tier and a $249 monthly plan listed at the time of reporting, plus pay-as-you-go credits and custom enterprise pricing. That gives smaller teams a route to try the product without a negotiated deal. The credit model also makes the cost of repeated research visible, though buyers still have to weigh the cost of agent runs against the time saved and the quality of the resulting data.

Customer names do not measure deployment scale

Riveter's YC profile and launch post name OpenAI and Snapchat among the companies using the product. Grills has also cited Antler and Sunco as early customers, while Riveter's site features testimonials from users at Zeffy, Roundabout Technologies, Fermat and Multiply. These are useful signs that the use case has reached teams outside the founders' own experiments. They do not establish how broadly Riveter is deployed within those organizations or how much data it processes.

The customer mix hints at a wide potential market. A sales team may want a list of narrowly defined prospects; an e-commerce operator may need competitor pricing; a risk team may want public information to support business checks. Each use case starts with facts scattered across the web, but they impose different expectations for accuracy, freshness and auditability. Riveter's challenge is to serve those workflows without turning every new customer request into bespoke services work - the very work its founders initially did by hand.

The surrounding market is moving toward fuller-stack tools for AI agents to retrieve web information. On September 22nd, 2026, web-crawling company Firecrawl announced a $75 million Series B alongside Alexandria, a product that combines its indexes, official data providers, custom connectors and the live web. The announcement shows investors backing companies that want to sit between AI applications and the information they need. Riveter's approach is narrower in disclosed scope: it focuses on turning web research into structured datasets and recurring enrichments. The overlap is real, but the products are not interchangeable simply because both touch web data.

Riveter's more direct competitive pressure also comes from conventional data vendors, scraping platforms and spreadsheet-based enrichment tools. Its pitch pushes against each in a different way: live search instead of relying only on a fixed index, agents instead of a hand-built scraper, and a complete dataset workflow instead of a one-off chatbot response. The claim that this can work at scale is the commercial hinge. A polished demo can show that an agent finds an answer; an operating data product has to keep finding correct answers across changing sites, unusual formats and repeat runs.

A founder's workflow made into infrastructure

Grills' hands-on launch approach makes Riveter's origin concrete. She sold the output before every part of the application was ready, used real customer data to expose the gaps, then worked with Watters to move the repetitive work into software. That sequence resembles the problem itself: first someone researches, checks, corrects and structures the information; only then can the process be automated with some confidence.

The product's move into ChatGPT and Codex follows that logic. Riveter is asking users to start where the request already occurs, then hand off the sourcing and row-by-row extraction to a specialized service. For Grills, the bet is that businesses will pay to replace repeated manual research and fragile one-off pipelines with a reusable system. The work she once did over a weekend is now the product's promise. Riveter's next test is whether that promise holds across the less tidy datasets customers bring after the demo.