Your Buyer’s AI Is Pitching YOUR Product (Badly) | Adrian Rosenkranz, CRO @ Webflow
And Why The Efficiency Trap Is Breaking Your Number
The best AI tool your sales team built this quarter is the reason you’re going to miss your number.
It’s slick. It works. It saves four hours a week. It’s also a sales director doing rev ops work instead of being on customer calls, which is the only thing they were actually hired to do. Adrian Rosenkranz, the CRO at Webflow, calls this the efficiency trap. After spending an hour with him on the podcast last week, I’m convinced it’s the single biggest threat to GTM productivity right now, and almost nobody is naming it.
Adrian ran global sales at Tableau before Webflow, scaling a multi-billion-dollar org across enterprise, SMB, and self-serve. He’s now sixteen months into the CRO seat at Webflow, supporting 300,000 customers and watching two simultaneous rewires of go-to-market play out in real time. One outside the building, one inside it.
The outside view: 20% of traffic across Webflow’s customer base is now AI bots. Sixteen months ago, that number was less than 1%. Forty-three percent of that bot traffic is agents pulling live answers for a human buyer who will never click your link. AEO has quietly replaced SEO as the discovery layer for B2B software. Most revenue leaders haven’t updated a single dashboard to reflect it.
The inside view: every revenue team is now buried under well-meaning AI hacks. Sales managers building coaching tools. Reps wiring up call-prep agents. Marketing building its own version of the same workflow rev ops is also building. Each project feels like a win. Almost none of them move the actual outcome.
Adrian’s diagnostic for whether your AI investments are working is the most useful question I’ve heard a CRO ask in years, and I’ve already stolen it.
Six takeaways from a conversation that genuinely changed how I’m thinking about the next twelve months.
1. The Efficiency Trap Is a Leadership Failure
Adrian told me a story I keep replaying.
A sales director on his team walked him through an AI coaching tool he’d built. Cool architecture. Real engineering. The kind of side project that makes a manager look forward-thinking in a 1:1. Adrian listened, nodded, and asked one question.
“Are you on more customer calls?”
The director paused. “What?”
“Yeah. This is nice. Are you on more calls?”
That’s the cleanest articulation of the efficiency trap I’ve heard. Anyone who gets really good at responding to emails ends up getting more emails. Anyone who gets really good at building AI workflows ends up building more AI workflows. Productivity at the input has almost nothing to do with productivity at the outcome.
Stuart Butterfield (Slack Founder) has a phrase I’ve stolen for moments like this: hyper-realistic work-like activity. The actions that look like work, feel like work, and produce nothing real. They’re seductive because the dopamine hit of finishing a task is real. The business value is the part that’s missing.
The instinct in modern GTM is to give every rep their own AI stack. Build your own call-prep agent. Build your own follow-up automation. Build your own coaching loop. The thinking is that decentralization equals empowerment. The reality is that you’ve taken your highest-paid customer-facing employees and turned them into part-time rev ops engineers. Their tools usually work. They also take hours to build, hours to maintain, and ship a fraction of what your applied AI team would deliver in the same week.
Adrian’s reorg is the cleanest answer I’ve seen. He collapsed marketing ops, sales ops, post-sales ops, and rev ops into one GTM engineering team. Their job is the connective tissue: the workflows that cross departments, the agents that share context, the eval frameworks that measure whether the system is working. Customer-facing reps are not on this team. They’re with customers.
This is Goodhart’s Law in modern dress. When a measure becomes a target, it stops being a good measure. Track customer minutes. Everything else is theatre.
A few months ago I wrote a Substack arguing for centralized AI deployment in GTM. It became my most popular post by 3X. People felt it. They were watching their best people drift away from customers and into config files, and they didn’t have clean language for why it was bad. Adrian gave it to me.
Every AI investment in GTM should be measured by one question. Are reps spending more time with humans? If the answer is no, the strategy is the problem.
2. The Narrative Funnel Beats Feature Enablement
Your product team is shipping faster than your enablement team can possibly keep up. Reps can’t memorize what shipped last week. Customers ask about features that didn’t exist last quarter. The default response is more enablement: more decks, more videos, more SKO sessions, more Slack channels nobody reads.
Adrian’s argument is that the entire frame is wrong. Reps learn the narrative first. Features fit inside it.
He uses an analogy I’m going to ruin in dozens of 1:1s. Imagine your manager surprise-Slacks the team on a Tuesday: “Drop everything, we’re doing a cold blitz right now.” Cue eye roll. Same manager, same activity, but on a standing Monday at 9am: “It’s Monday, we’re doing the blitz.” Suddenly it’s the run of the business. Same activity. Different frame. Different meaning.
A great narrative is the Monday cold blitz of go-to-market. When you have a clear story about how the world is changing and what your customer should do about it, every new feature becomes a slot in that story instead of a separate thing to memorize. Webflow’s narrative is that web traffic is fundamentally shifting from human-only to human-plus-agent. Every feature they ship plugs into that story. When Markdown for Agents launched, reps didn’t have to learn a new pitch. They just connected the dots.
Andy Raskin has been arguing for years that strategic narrative is the highest-leverage piece of GTM content a company can produce. Most companies have positioning. Very few have narrative. The difference becomes obvious the moment your PMs start shipping faster than your enablement team can keep up.
There’s a second move Adrian makes that’s worth stealing. He categorizes every new feature into one of two buckets. Reason-to-call features get a proactive motion: an agent scans his Slack channels daily, matches new features to the ICP, and surfaces which open deals should hear about it. Ambient features just get added to the knowledge base and surfaced in Slack via a bot when a rep asks.
Most teams treat every release as a reason to call. That’s why your reps are tired and your buyers are tuning out.
If your narrative is strong, features get easier. If your narrative is weak, every release is a fire drill.
3. 20% of Your Traffic Is Now Non-Human
The numbers will hit you.
Sixteen months ago, less than 1% of traffic across Webflow’s 300,000-customer base came from AI bots. Today it’s nearly 20%. And that 20% isn’t a homogenous blob of model-training crawlers. Forty-three percent of it is live agents fetching answers in real time for a human buyer.
When ChatGPT changed how it surfaced citations last August, moving from inline links to aggregated source lists, Webflow saw a 30 to 40% drop in referrals overnight. The traffic didn’t disappear. It moved. Buyers started getting their answers inside the LLM and typing webflow.com directly into their browser, skipping the click entirely.
This is the single biggest shift in B2B discovery since SEO became a discipline. SEO competes for a click. AEO (agent experience optimization) competes for a citation, which is to say, for whether the LLM mentions you at all when a buyer asks a question.
The good news: SEO and AEO overlap. Authority, freshness, and structured content help both. The new work is mostly architectural. Schema markup so agents can interpret the page (per the schema.org standard). Markdown versions of every page so agents can extract clean content fast. Metadata that tells an LLM what your page actually is.
Adrian uses a sharp visual to explain the gap. Look at any beautifully designed Webflow site, like Lando Norris’s. Animations, gradients, micro-interactions, the works. Then right-click and inspect. That stripped-down code is the only thing the agent sees. The animations are for humans. The markdown is for everyone else.
If you’re a CRO, the takeaway is brutal in its simplicity. You need a dashboard that tracks your share of LLM citations the way you track your share of search results. You need someone whose job is AEO. And you need to start thinking about your website as two simultaneous experiences: the one your buyer sees and the one their agent reads on their behalf.
The 6x conversion lift Webflow sees on referred traffic versus typed-in URL traffic is the lagging signal. The leading signal is whether you show up in the LLM at all.
4. Shape, or Be Shaped
Adrian was in France last month, sitting across from a customer, when he realized something he couldn’t unsee.
The customer was asking the LLM questions in real time. While Adrian was talking. About what Adrian was saying.
“Is this pricing fair?” “What are alternatives?” “What questions should I be asking him?”
He told me about it half-laughing, half-horrified. The exact thing CROs have been worried about in the abstract for two years is now happening in conference rooms in Paris. Your customer has a second analyst in the room and they can’t see it. You can’t either.
Adrian’s framing of this is the line I’ve been turning over in my head since we recorded.
“Are we going to step up and shape it, or are we going to be shaped by it?”
There is no third option. Either your messaging, your positioning, your competitive narrative, your pricing rationale, all of it, is structured cleanly enough that an LLM can fetch it and represent you accurately, or the LLM will improvise on your behalf using whatever it finds on G2, Reddit, and your competitor’s blog. It will not check with you first.
This is the part of the AEO conversation that goes well past schema markup. AEO is a content strategy. A messaging discipline. A competitive intelligence function. The companies that will win are the ones who decide right now that every objection, every alternative, every pricing question, every fairness comparison gets a clear, sourceable answer that an LLM can find and use.
If your strategy for AI buyer discovery is “we’ll just keep doing good marketing,” your LLM is improvising. And it’s not on your team.
5. Knowledge Bases Are No Longer Documents
Every revenue org has a knowledge base. Most of them are graveyards.
The onboarding deck from 2023. The win-loss report from Q2. The ICP doc your previous CMO commissioned three exec teams ago. They’re sitting in Notion or Drive or Guru, technically findable, functionally dead.
Adrian has rebuilt this entire concept and it’s the most copyable idea in our episode. His ICP doc isn’t a doc. It’s a markdown file that updates itself every week via a cron job. It pulls from closed-won deal data, customer call transcripts, Slack channels where the team is talking about deals, and external best practices he’s curated. Every Monday it produces a fresh version, and you can see how the ICP has drifted week over week.
The actual file contains anti-ICP patterns, ICP patterns, customer quotes that validate segmentation, jobs-to-be-done frameworks, and competitor data. Then every other AI agent in his stack consumes it as ground truth. The just-in-time call-script agent reads it. The deal-coaching agent reads it. The market-fit analysis agent reads it.
Why this matters: the value of any AI workflow is capped by the quality of the context it can access. A static ICP doc is technical debt. A self-updating one is leverage that compounds every week.
I’ve been building toward the same pattern. My VP of rev ops built something we call the Gardener. It scrapes Slack channels, Notion docs, and call recordings to keep a master business-context file alive across our org. Anything any agent in our stack does, it does with that file as the ground truth.
The future of GTM enablement isn’t going to be courses and decks. It’s going to be living context files that the entire team, human and agent, reads from. If your knowledge base doesn’t update itself, you’re already operating on stale information.
6. Context Is the New Moat. Walled Gardens Are Dead.
Legacy revenue tools sold you a story for the better part of a decade. “Our intelligence is the moat. Our data is the lock-in. The longer you use us, the more we know.”
That story is over.
Adrian said something in our conversation that perfectly captures where we are:
“I work really hard sometimes, it’s pretty funny how hard I work to make sure my Gong calls are not in Gong.”
I cut Gong off our stack two years ago for the same reason. Their walled-garden approach to call data was incompatible with how I wanted to use the intelligence. We moved to Momentum, which lets me push call context anywhere I want, run my own prompts on top of it, and feed it into whatever agent or workflow needs it. Same data, dramatically different leverage.
Adrian’s prediction: every tool that holds your context hostage is going to lose. The new moat is the opposite. It’s how much context the platform can feed into agents acting on your behalf. The vendors that thrive over the next three years will sound like, “here’s all our data, here’s all our APIs, here’s a clean way to plug us into your stack.” The ones that resist this are signaling weakness, not strength.
There’s a corollary for build-versus-buy worth pulling out. Adrian’s framework: buy systems you can build with AI on top of. Don’t rebuild governance, roles, or permissions. Don’t rebuild any infrastructure with a 1,000-day shelf life. Do build short-lived, throwaway tooling for things that need to exist for a quarter. The longevity of the workflow is the test.
If a vendor can’t tell you how an agent reads from and writes to their system, they’re not a long-term partner. They’re a short-term dependency.
What Actually Matters
Two things are happening at once and they’re related.
Outside your building, AI agents are inserting themselves between you and your buyers. Twenty percent of traffic. Forty-three percent of that, live queries. A buyer in a conference room in Paris quietly fact-checking the CRO across the table. The discovery layer for B2B software has fundamentally moved, and most revenue leaders haven’t updated a single dashboard to reflect it.
Inside your building, every revenue team is drowning in well-meaning AI hacks while the actual unlock sits one org redesign away. Centralize the engineering. Build living context files. Make sure every reorg ends with reps spending more time with customers, not less.
The CROs I’m watching pull ahead in 2026 have one thing in common. They’re using their own context, their own thinking, their own taste, to multiply their team. They’re modeling the way themselves. They’re getting their hands dirty.
If you take one thing from this conversation, take Adrian’s question. Walk into your next 1:1 with whoever’s been telling you about their AI projects and ask it.
“Are you on more customer calls?”
Listen to what comes back.
If this conversation hit, do me a favor and rate the podcast. Five stars on Apple or Spotify is the easiest way to make sure more conversations like this one happen. Adrian was generous with the playbooks. The least we can do is get more people listening.
Thess are some great insights for any GTM leader.
I’ve read recently, the narrative-first framing is the right approach and something most teams are missing entirely.
The Revenue Flywheel adds one layer to this: once the narrative lands, the job isn’t just to keep the rep talking. It’s to keep the customer engaged in solving their own problem. The best flywheel conversations I’ve seen aren’t feature pitches or even narrative pitches. They’re diagnosis sessions. The rep’s job is to help the buyer articulate the problem more clearly than they could before the call. When AI takes over the feature explanation, that actually frees up the human for the only thing AI can’t replicate. Genuine problem solving with another human.
The question I’d add to Adrian’s framework: is your rep spending more time helping customers understand their problem or explaining your product? Those are very different conversations, and only one compounds.
Love this conversation, thanks for sharing!
Hey Kyle!
Feel there's a fuzzy and sometimes tricky-to-know-in-advance line between the AI efficiency trap and creating actual AI efficiencies.
For example, finding context engineering is to a large extent just building systems to measure and manage the quality of the systems that measure and manage the quality of the work (except some days it can feel like turtles all the way down)
How do you navigate this within the context of the efficiency trap?
Especially curious about your personal AI stuff where you might not have an engineer to delegate the build work, bug fixes, and ongoing context management and maintenance too?