From Cash Back to Outcomes: Bryan Leach on Building Ibotta, Going Public, and Betting on Denver

Full analysis

Bryan Leach, founder and CEO of Ibotta, spent an hour on Signal & Noise making one big claim that matters to anyone who buys or builds with AI: promotions should be priced on verified incremental sales, not impressions or clicks. Along the way he dropped a number worth chewing on for a different reason. His engineers are shipping software 45 to 50 percent faster with Claude Code than they were six months ago. That second claim is the one this council cares about, because it's a data point on the question every AI buyer is trying to answer right now: how much does an AI coding assistant actually change your output?

How hard is this to undo? For the reader, nothing here forces a decision. This is a founder telling his story on a podcast. But two of his claims are testable and both have short shelf lives: the 45 to 50 percent coding lift, and the prediction that retail media networks get repriced once incrementality is disclosed.

The Skeptic

A 45 to 50 percent productivity lift from an AI coding tool is a management estimate with no methodology behind it, and Leach knows better than anyone that his engineers want to look busy and modern. Velocity is not shipped features. It's not fewer bugs. It's not revenue. Every CEO who bought Claude Code seats has an incentive to report a big number to justify the spend and to sound AI-forward on a podcast. And notice the framing: "the velocity with which we are able to build things." Build, not ship, not maintain. Nobody asked the question that actually matters: did Ibotta ship more product to customers in the last six months, or just generate more code that someone now has to review?

The Researcher

The one hard third-party number in this episode is the Circana meta-study: 50 lift studies, Ibotta campaigns driving seven to eight times the incremental sales lift of an average digital CPG ad. Circana is a paid measurement vendor with commercial ties across the retailer ecosystem, and the "average digital ad" is undefined. Which formats? Which time windows? A 7-to-8x gap that wide usually means the benchmark is soft, not that the winner is a miracle. On the coding claim, Leach's 45 to 50 percent is directionally consistent with what other engineering leaders report from AI assistants, but self-reported velocity gains cluster suspiciously around "about half," which is what people say when they mean "a lot" and haven't measured it.

The Builder

Strip the story down and there's a real playbook here. Ibotta uses Claude Code for software, AI for real-time offer targeting, and AI agents for internal legal Q&A, all on top of Databricks and clean rooms. That legal Q&A agent is the quietly useful one. Most companies have a lawyer answering the same twenty routine questions all week, and that's exactly the kind of bounded, high-volume, low-stakes task where an agent earns its keep on Tuesday morning. The offer-targeting piece, personalizing discount depth by each shopper's price sensitivity, maps directly onto bid factors you already run for audience, geo, and device. Same multiplicative logic, applied to coupon depth instead of a bid. Nothing exotic. The interesting part is that a receipt-scanning app now runs a modern AI stack without a research team.

The Enterprise Buyer

Leach is selling a purchase model, not a technology: pay only for sales that wouldn't have happened otherwise, measured by a randomized holdout against loyalty IDs. For a CPG marketing buyer that's genuinely attractive, because it kills the oldest complaint about coupons, that you're paying to discount people who'd have bought anyway. But the whole thing rests on trusting Ibotta's holdout methodology, and the measurement runs inside retailer data through clean rooms Ibotta helps operate. A CTO signing that contract is trusting the vendor to grade its own homework. The out-clause to demand: an independent audit of the incrementality math, or the right to run your own holdout. Without it you're buying a number the seller computed.

Where the personas split

The real disagreement is about whether Leach's two big numbers survive an audit. The Enterprise Buyer wants the incrementality model because it fixes a genuine pain in CPG spend. The Skeptic and Researcher both point at the same soft spot: every impressive figure in this episode comes from a party with a reason to inflate it. Circana gets paid by the ecosystem it's measuring. Leach's coding lift is unaudited. The retail media repricing prediction is Ibotta's sales pitch stated as a market forecast.

The second split: the Builder sees a company quietly running a competent AI stack, while the Skeptic sees a founder converting ordinary tool adoption into a podcast talking point. Both are right. Ibotta really does use these tools. That doesn't make 45 to 50 percent a real measure of anything.

What this actually hinges on

For the AI buyer, the useful takeaway isn't Ibotta's coupon model. It's the coding number, because it's a live data point in the question you're all asking: does an AI coding assistant move real output, or just perceived velocity? Leach's answer is a loud, unaudited "yes, by half." Treat it the way you'd treat any vendor-adjacent testimonial. If you're evaluating Claude Code or a rival for your own team, don't buy the 50 percent. Measure your own team's shipped features and code-review load before and after, for a quarter, and see what's left after the enthusiasm fades.

Prediction: No major retail media network (Walmart Connect, Amazon Ads, Kroger Precision Marketing, or Albertsons Media Collective) will replace CPM or CPC pricing with a cost-per-incremental-dollar model as its standard buying unit by the Q4 2026 earnings season in February 2027.

Confidence: High. RMNs make more money hiding incrementality than disclosing it.

Why: Leach predicts retail media networks face "repricing pressure" once true incrementality is exposed, and he's got an obvious reason to say so, because Ibotta sells the incrementality product that would replace them. The mechanism that kills his prediction is money: CPM and CPC pricing charges the advertiser for impressions and clicks that would have converted anyway, and a cost-per-incremental-dollar model charges only for sales the ad genuinely caused, which is a much smaller, much less flattering number. No retailer voluntarily switches to the pricing model that shrinks its own revenue and admits most of its ad load was padding. The opposite outcome, a major RMN adopting incremental pricing as standard, would require Walmart or Amazon to reprice their fastest-growing profit line downward on purpose, and nothing in their incentives points that way.

Revisit by 2027-02-28: We're right if none of Walmart Connect, Amazon Ads, Kroger Precision Marketing, or Albertsons Media Collective has made cost-per-incremental-dollar its standard buying unit by the Q4 2026 earnings season. We're wrong if any one of them has.

That doesn't mean incrementality reporting goes nowhere. Albertsons publishing white papers is a real signal that buyers are pushing. But publishing a study is free. Repricing the revenue line is not, and that's the part that won't move.