At Profound's Zero Click New York 2026 event, Ramp presented an experiment that most SEO teams would never think to run: they hid money where only a machine could find it. A $3,100 sign-up incentive, embedded in machine-readable code, invisible to any human clicking through the site. Twelve days passed with no result. Then Claude started citing the exact number, and the case became one of the cleanest live demonstrations to date of what ai agent marketing actually looks like when it works.
The $3,100 line of code nobody was supposed to see
Ramp is a corporate card and spend management company, and the mechanics of what they built are, per Profound's recap of the event, published June 15, 2026, straightforward once you see them. Somewhere in Ramp's site or docs, they embedded a $3,100 sign-up incentive in machine-readable code. A normal visitor scrolling the page would never see it. There was no banner, no popup, no on-page copy announcing the offer. The only thing that could find it was software parsing the page the way an AI agent parses a page: reading structure and data, not rendering a layout for a person to look at.
That distinction is the entire experiment. Most GEO advice through 2025 and into 2026 has been a variation on the same theme: write clearly, structure your headings, answer the question directly, and you'll get picked up by both Google and the AI engines, because good content for humans turns out to be good content for machines too. Ramp tested a different premise. What if you built something with no human-facing version at all? Not content that happens to also work for an AI crawler, but data whose only intended reader was an AI crawler, sitting in a spot a person would never think to check.
Twelve days of nothing, then Claude found it
The first phase of the experiment is the part most case studies skip past, because it isn't flattering. For twelve days, nothing happened. No AI engine picked up the embedded incentive. No citation, no traffic bump, no signal that the bet was working. If Ramp had pulled the plug on day ten, this would be a footnote about a tactic that didn't pan out.
On day twelve, Claude started citing the $3,100 figure directly in its responses. Not a paraphrase, not a rounded estimate — the exact number Ramp had buried in code. From there, per the recap, Ramp's agent citations climbed from 40 to 370 within three weeks of Claude's initial pickup, close to a tenfold increase. Traffic that Profound attributes specifically to Claude rose 180% over the same experiment period. One engine finding one piece of machine-readable data changed Ramp's citation volume by an order of magnitude and moved real traffic behind it.
The lag matters as much as the payoff. Twelve days is long enough that a team watching a dashboard for quick validation would likely have written the tactic off. It's also short enough, against the scale of a normal content or SEO initiative, that the eventual 10x citation jump reads as fast once it started. Both things are true at once, and neither is a coincidence you can plan around precisely — you can build the machine-readable data, but you don't control when, or whether, a given engine's retrieval and training pipeline surfaces it.
This is ai agent marketing, not SEO with better manners
The instinct to file this under GEO best practices — write clean structured data, use schema, keep things crawlable — undersells what Ramp actually did. Structured data that a human could also read if they went looking is still, fundamentally, human-facing content wearing machine-friendly formatting. What Ramp built had no human-facing version. It existed for one audience: an autonomous agent parsing the page on someone's behalf. That's the frontier this case study points to, and it's a genuinely different discipline from generative engine optimization as most teams currently practice it — closer to designing an interface for a non-human user than writing content for a human one.
Gamma, the AI-native presentation tool, presented a version of the same shift at the same event. Per the recap, AI search became Gamma's top referral source outside word-of-mouth within a single year, and the company now generates 10 million monthly impressions on third-party high-authority sites. Their strategy centers on documentation written specifically for AI agents to parse — not blog posts optimized to also please a crawler, but reference material built with an agent as the primary reader. Coca-Cola, CVS Health, U.S. Bank, and Delta also presented strategies at the event; the recap doesn't break out specific performance numbers for those four, so there's nothing to cite there beyond the fact that enterprise brands well outside tech are now treating this as a category worth a stage at an industry conference.
There's also a quieter implication in the shape of Ramp's bounty. A dollar figure an agent can read and repeat isn't just a citation hook — it's a step toward agentic commerce, where the agent doesn't just describe an offer but eventually acts on it on a user's behalf. Ramp built for a reader that consumes data and can, in principle, transact on it. That's a different design target than a landing page meant to persuade a person scrolling on a phone.
We've made a version of this argument before in the context of why AI agents can't find your price — that agentic tools increasingly fail not because a company's pricing is hidden from people, but because it's structured in a way no agent can parse cleanly. Ramp's experiment is the inverse proof of the same point. If an absence of machine-readable data can make an agent unable to answer a question about you, a deliberately placed piece of machine-readable data can make an agent choose to cite you, specifically, with the exact number you gave it. That's not a crawlability checklist item. It's closer to running a small classified ad in a channel with exactly one reader, and that reader happens to write your company into millions of downstream conversations. Teams in AI/ML and adjacent technical categories, where the buyer is increasingly likely to route a first query through an assistant rather than a search bar, have the most obvious reason to test this now.
Position beats frequency: the ai agent marketing lesson in G2 and LinkedIn's numbers
Getting cited isn't the finish line. Two other data points from the same recap sharpen what actually matters once citations start happening. G2 saw a 44% increase in citations after adding context summaries to its product pages — a smaller, more targeted move than Ramp's, and one that still moved the number meaningfully. But the more important finding from G2, per the recap, is that citation position matters more than citation frequency for revenue signals. Being mentioned isn't the same as being mentioned first, or being the source an engine's answer leans on. The recap states that 8 of 10 B2B buyers purchase from their day-one answer-engine short list — which means the citations that count are the ones that land you on that list from the first query, not the ones that accumulate slowly over dozens of mentions.
Supporting data from the same Profound recap (G2 and LinkedIn), Zero Click New York 2026
LinkedIn's numbers, also from the recap, explain why position matters so much for a B2B audience specifically. 94% of B2B buyers now consult an answer engine before talking to a salesperson, and answer-engine research has compressed B2B evaluation timelines by 40%. LinkedIn itself is described in the recap as the most-cited domain for professional queries across leading AI platforms, with its citations up roughly 2x year-over-year. Put together with G2's position-over-frequency finding and our own research on how AI engines decide what to cite, the picture for enterprise B2B teams is specific: your buyer is very likely forming a shortlist before you know they exist, that shortlist gets built inside an answer engine, and the difference between being on it and being one of many citations buried further down is the difference that actually shows up in revenue.
One case study, not a playbook
Here's the honest version of what Ramp's experiment proves and what it doesn't. It proves a mechanism is real: machine-readable data, placed with no human-facing equivalent, can get picked up by an AI engine and cited verbatim, and that pickup can move real traffic. It does not prove the mechanism is repeatable on command, that Claude will find the next company's version of this within any predictable window, or that the $3,100 figure itself — as opposed to the specificity of a concrete dollar amount — was what made the content citation-worthy. It's one company, one experiment, recapped by Profound, a vendor whose entire business is AI visibility measurement and who has an obvious interest in stories like this one circulating. None of that makes the story false. It does mean the twelve-day lag and the 40-to-370 jump are a single data point, not a formula with error bars.
It's also worth being clear about what wasn't reported: baseline traffic before the experiment, whether Ramp tested this on more than one page or format, and whether other engines besides Claude ever picked up the same data. The recap frames this as a Claude story specifically, not a cross-platform one, and that's a meaningful limitation for anyone trying to generalize it to ChatGPT citations or other engines' retrieval behavior. Our own work with technical, product-led companies — Matroid's no-code computer vision platform among them — points the same direction without needing a single dramatic bounty: structured, machine-legible data placed near content that already has authority tends to outperform generic on-page cleanup, even when the lift is smaller and slower than Ramp's twelve-day spike.
What to actually do about it
Most teams reading this aren't Ramp, and shouldn't try to replicate the experiment at Ramp's scale. A company-wide overhaul chasing a single dramatic citation spike is the wrong lesson to take from a case study this size. The right lesson is smaller and lower-risk: pick one thing you already know matters to buyers — a specific price, a specific spec, a specific number you'd want an agent to quote back verbatim — and give it a machine-readable home next to content that already ranks or already gets cited, rather than rebuilding the site around the idea.
This won't produce a guaranteed Ramp-sized result. Nothing in the recap supports promising that. What it will do is give you a real, low-cost test of the underlying mechanism — content and data built for agents specifically, sitting in a format and location a human reader never needs — without betting the site on it. If third-party citation strategy is already part of your GEO program, this is the next layer down: not where you're mentioned, but what you've built specifically for the thing doing the mentioning.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.