The Revenue Leadership Podcast · 29 Apr 2026 · From the week of 27 April
E67: Your Buyer's AI Is Pitching YOUR Product (Badly) | Adrian Rosenkranz, CRO @ Webflow
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Adrian Rosenkranz, CRO of Webflow, joins host Kyle Norton to discuss how AI agents are changing web traffic, sales and go-to-market operations. The conversation covers AI bot crawls making up nearly 20% of Webflow customer site traffic, how to make pages readable to agents with markdown and schema, and why LLM referrals are increasingly showing up as direct visits. Adrian then describes internal AI work at Webflow, including a merged go-to-market engineering team, an email drafting agent, a weekly ICP document and a just-in-time outbound script generator. His central argument is that teams should avoid AI efficiency gains that don't move outcomes, and should keep ownership of their context rather than lock intelligence inside vendor tools.
For founders
- Adrian argues against building a separate site for AI agents, saying SEO work also carries AEO and the advantage comes from doing one experience well.
- Webflow offers a toggle that generates a markdown version of every page for agents, and Adrian also recommends adding schema markup within the same site.
- Kyle says he is wary of tools that keep intelligence in black boxes; Adrian agrees, arguing walled-garden intelligence is no longer a sustainable moat and that tools giving agents better context will win.
- He says build-versus-buy should favour systems you can build on top of, with throwaway builds reserved for short-term needs.
- He says the hard part of AI workflows is keeping them working on day 20, 50 and 100, not the first build.
For revenue leaders
- Webflow saw about a 6x conversion improvement on LLM referrals, and a 30-40% fall in its referrals after ChatGPT changed how it showed sources, with traffic reappearing as direct visits.
- Adrian says collapsing marketing ops, RevOps, sales ops and post-sales ops into one go-to-market engineering team lets workflows cross handoffs.
- He warns that AI can make reps more productive at tasks that don't drive outcomes, and says customer-facing people should spend their time on customer calls.
- He describes a weekly ICP markdown file regenerated from calls and deals to track drift, and a call script generated when a rep names a company in Slack.
- He says a strong narrative lets new features slot in, and that customer success business review prep dropped from five to eight hours to about 15 minutes with an agent.
What was said 24, most useful first
Nearly 20% of traffic to Webflow customer sites now comes from AI bot crawlers, and about 43% of that is AI fetching a page to answer a user's live question. Listen
Adrian said AI traffic was under 1% of customer sites about 16 months ago, when it was mostly model training. Since then models can browse the web and citation behaviour has changed. He argued that teams will have to sell to both humans and agents to reach their prospective buyers.
“it's nearly 20% of traffic going to Webflow customer sites are actually AI bot crawlers.”
Listen to the episode Positioning & marketing Link to this Report a problem
Webflow sees about a 6x improvement in conversion when a prospect is referred from an LLM. Listen
Adrian said these prospects are much further down the funnel; Kyle suggested this is because they have probably had an extended conversation with the LLM first. Adrian noted that users increasingly get answers inside the LLM and then type the URL directly rather than clicking through.
“we see typically when someone does get referred, it's about a 6x improvement in conversion.”
Listen to the episode Pipeline & demand generation Link to this Report a problem
When ChatGPT moved from inline links to an aggregated sources list, Webflow's referrals fell 30-40% but the traffic reappeared as direct visits. Listen
Adrian said the change came around last August and that each model change alters how sources are shown, which changes prospect behaviour. He said Webflow's referral drop showed up as people typing webflow.com after getting their answers, so referrals alone understate AI-driven discovery.
“we basically saw a pretty significant 30 40% drop in our own referrals, but it just showed up in a different place which was people typing webflow.com as an example.”
Listen to the episode Pipeline & demand generation Link to this Report a problem
A separate agent-only site would be a disadvantage, so the goal is one experience that serves both SEO and AEO. Listen
Adrian said what you do for SEO gives you AEO for free, and that brand authority counts in both. Extra items such as schema markup and markdown versions of pages are added within the one site. He said he gets this question most weeks from CMOs and VPs of marketing.
“Based on the data that we see now that would actually be a disadvantage.”
Listen to the episode Positioning & marketing Link to this Report a problem
Adrian warns of an efficiency trap where AI makes people faster at activities that don't move revenue. Listen
He said people who get better at responding to emails just get more emails, and that AI can have someone spend 20 minutes on an email instead of five. He argued the test is whether the work drives outcomes rather than whether it is done faster. Kyle added that reps are especially prone to this because it can replace harder work like cold calling.
“so they start like spending 20 minutes on sending an email instead of just spending five minutes on writing it”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Webflow merged marketing ops, RevOps, sales ops and post-sales ops into one go-to-market engineering team. Listen
Adrian said the specialist ops teams had always handed workflows from one to another, and that AI workflows need data and context shared across those handoffs. The team is built around go-to-market engineers who work on the connective tissue.
“which I've actually collapsed everyone from marketing ops and revops sales ops post sales ops. It's one team with go to market engineers”
Listen to the episode Leadership & culture Link to this Report a problem
Webflow regenerates its ICP each week as a markdown file from calls, Slack and messaging, so the team can track drift over time. Listen
Adrian said the job combines Webflow's messaging and positioning doc, historical customer call transcripts and Slack activity, plus frameworks he borrowed from others. The output includes anti-ICP patterns, ICP patterns, customer quotes and jobs-to-be-done, and other agents use it. He can compare this week's file with last week's and look back a month.
“And so it's on a cron job. And so every week it just makes a new file and you can see the drift from the previous one.”
Listen to the episode Strategy & market Link to this Report a problem
Webflow cut customer business review prep from five to eight hours to about 15 minutes with an agent that builds a password-protected page. Listen
Adrian said a customer success manager used to look up data and build a slide deck by hand. Now an agent in Claude desktop looks up account information and call history and builds a password-protected Webflow page the customer can review. He said they can now do more of them, even monthly, because preparation, not the call, was the constraint.
“they took what took something that took normally 5 to eight hours depending and it's now 15 minutes.”
Listen to the episode Retention & customer success Link to this Report a problem
Webflow's email agent drafts customer notes from Slack requests, call history and calendar in Adrian's voice, and close to 90% go out unedited. Listen
Adrian said the agent reads a Slack channel for requests, pulls 60 days of customer call conversations, checks his calendar and Gmail, and saves a rewritten draft in Gmail. He built it in Claude Code with MCP connections to Slack, call data, Gmail and Google Calendar, plus a voice skill trained on emails he had already sent. It grades its own drafts, scoring lower when he edits them or gets no reply.
“right now it's actually gone close to 90% of the time I don't have to edit it.”
Webflow offers a toggle that automatically generates a markdown version of every page for agents. Listen
Adrian said agents ignore animations and visual design and read the underlying code, so markdown is what they consume. Webflow built the feature so customers can switch it on instead of setting up metadata or alternate files by hand.
“So the way we solved it is we just give people a feature that they can just toggle on and it automatically creates markdown for agents of every page.”
Schema markup should be on your pages, and agents make it easier to keep it updated across the site. Listen
Adrian pointed to schema.org as the general standard and said most websites have pages where schema markup is missing. He said agents make it easier to keep that markup current everywhere.
“You should have schema markup.”
Listen to the episode Positioning & marketing Link to this Report a problem
Agents answer questions on a brand's behalf without its control, so companies need to become the knowledge base LLMs draw on. Listen
Adrian said LLMs will answer nearly any question, so the long tail of buyer questions is much longer than marketing has covered. He said prospects ask LLMs whether pricing is fair and what the alternatives are, so revenue leaders must decide whether they shape those answers or are shaped by them.
“the question becomes, how do you become that knowledge base that an LLM will use to answer questions about your industry, about your product?”
Listen to the episode Positioning & marketing Link to this Report a problem
A strong narrative lets reps absorb new features without relearning their pitch. Listen
Adrian said reps should know the market narrative, what is changing, the company's point of view and why it matters to customers, so new features can be slotted into it. He gave the markdown feature as an example, presented as support for an existing story about showing up for agents.
“If you have a really strong narrative and positioning, then you should be able to slot in different features, functionality into the narrative.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Webflow built an agent that scans product release channels daily, flags features that fit the ICP and narrative, and matches them to open deals. Listen
Adrian said release volume was too high to track by hand, so the agent reads the Slack channels where PMs and PMMs sign off on go-to-market releases. It flags features that fit the ICP and messaging, then checks open deals against customer call transcripts to see which features are worth raising.
“I actually built an agent that said okay based off on our ICP based off our messaging and positioning I want you to look at these channels daily”
Listen to the episode Sales process & deals Link to this Report a problem
Adrian sorts new features into those that give a reason to call a prospect and those that strengthen a position in an existing deal. Listen
He described the first as features that address a problem a prospect has and are new enough to justify outreach. The second he called a good addition: not compelling enough to start a call, but useful to insert into a live deal. He said this split may be overly simplistic.
“is this features and functionality that give us a reason to call about a problem that they have and it's a new thing or is this features and functionality that actually are going to just like strengthen our position in deals”
Listen to the episode Sales process & deals Link to this Report a problem
Adrian asks customer-facing staff whether their AI work is keeping them off customer calls, and says AI work that ops can do adds no value. Listen
He described a sales director on his team building AI coaching tools and asked whether he was on more calls. He said that if a customer-facing role builds AI workflows that his ops team could build, it is not adding value, because customer conversations can't be handed to someone else.
“And I said great are you on calls more?”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Ongoing maintenance, not the first build, is the hard part of AI workflows. Listen
Adrian said the first version of an agent is rarely the problem, and that keeping it working on day 20, 50 and 100 is. He said he had to change his habit so that each bad output leads to an update of the underlying skill and context, rather than retrying or dropping the output.
“the day one is never the problem. Making sure things are like working on day 20, 50, 100. That's the hard part.”
Adrian names usage data and customer calls as the two most useful inputs for LLM context in any go-to-market function. Listen
He said usage data, meaning how customers use the product, and customer calls hold the most information. Marketing, sales, post-sales and support all want to use them, which he gave as the reason for bringing that work into one team.
“I always say the two most useful pieces of information and data you possibly can have to give LLM context in any go to market function is your usage data”
Adrian expects GTM teams to need people who build eval frameworks, maintain knowledge bases and translate business needs into agents. Listen
He said the skill of knowing one tool best is changing, and listed three needs: people who judge whether agents do good work, people who manage constantly evolving knowledge bases, and builders who turn business owners' needs into agents. He said these could be one person or several.
“who are the people on the team that are going to be managing the knowledge bases that are constantly evolving and learning that are used in a bunch of different other workflows?”
Listen to the episode Hiring & team building Link to this Report a problem
Adrian plans to compare an open deal with the ICP of Webflow's wins to find gaps and suggest questions to ask. Listen
He described asking the agent to analyse what has been discussed in a deal, compare it with the ICP of when Webflow wins, and say what is missing. He gave the example of a deal competing against WordPress. He framed this as where the work is heading rather than something he described as finished.
“Analyze what I've talked about in my deal. compare it to the ICP of when we win and tell me what I'm missing and what questions I should ask.”
Listen to the episode Sales process & deals Link to this Report a problem
Webflow generates a call script for each SDR as soon as they name a company they are about to call in Slack. Listen
Adrian said the agent enriches data on the company, looks at market activity, applies Webflow positioning and provocative challenger-style ideas, and returns a script. He said the goal is to get people on the phone with more customers; Kyle noted it compresses the time reps need to get ready for a call.
“right when they say that in Slack, they get a message back that says great here's your call script.”
Listen to the episode Pipeline & demand generation Link to this Report a problem
Adrian has agents grade their own output and runs a mini retrospective after frustrating sessions so the skill itself gets rewritten. Listen
He said each run of the email agent scores itself, with 100% meaning he didn't touch the draft and got a response. When a task is harder than it should be, he asks the agent to review the conversation, act on the actionable feedback and rewrite its approach. He said this has been a lifesaver.
“Time for a mini retrospective.”
Adrian's build-versus-buy rule is to buy systems you can build on with AI, and build throwaway tools only for short-term needs. Listen
He said governance, roles and permissions should be bought so teams can build AI on top of them rather than rebuild and maintain them repeatedly. One-time analyses can be lightweight artifacts, but recurring needs should become applications with a proper view. He also asks why building beats buying, weighing it against price relative to value.
“You need to buy systems that you can build with AI on top of it.”
He doesn't want tools that trap his intelligence, and Adrian argues walled-garden intelligence is no longer a sustainable moat. Listen
Kyle said he gets nervous when a platform tells reps what to do next without showing why. He prefers tools where he can inspect markdown files and context, and said he left Gong over its walled garden for Momentum, which pushes context into Salesforce and his own prompts. Adrian agreed that locking intelligence in is not a moat worth having, and said he works hard to keep his Gong calls out of Gong. He expects buyers and agents to pick tools that give agents better context.
“I don't really want to use tools that trap my intelligence.”