Your website has a new visitor, and it doesn't browse like you do
At Opticon New York, Optimizely set out a future built for three audiences at once: marketers, customers and AI agents. It's a compelling pitch. But Optimizely's own work on interfaces, experimentation and governance points to a harder truth: the tools are arriving faster than the operating models needed to run them. That gap is the thing worth paying attention to.
New York in early September is keynote season, and this year's pitch from the Opticon keynote stage had a tidy, almost architectural logic to it. Our long-standing partner Optimizely didn't turn up with a single headline feature. It turned up with a claim about who a website is now for.
For years the answer was simple: people. Then it was people and search engines. Now, Optimizely argued, there are three distinct audiences to design for, and if a digital experience isn't built with all three in mind, it's already behind.
Three Audiences, One Platform: MX, CX, and AX
- The framework is easy to say and rather harder to build.
- There's the Marketer Experience (MX): giving marketing teams more capacity to get through the work.
- There's the Customer Experience (CX): the familiar business of making sites more relevant, conversational and adaptive for human visitors.
- And now there's the Agent Experience (AX): making sure AI agents can find, understand and act on what's actually on a site.
None of the three is new on its own. What's new is how much Optimizely has now built underneath each one.
On the marketer side, Optimizely’s Virtual Teammates were the headline act, digital colleagues pitched as capable of taking on ongoing work rather than sitting idle until someone prompts them: product marketing, SEO and GEO, analytics, chief-of-staff tasks.
The distinction Optimizely is drawing matters more than the branding suggests. A standalone AI agent in e-commerce, on the company's own account, behaves like a tool you open, use and close, no persistent identity, no memory once the session ends. A Virtual Teammate is provisioned more like a person: it turns up in the team roster and the audit log, holds scoped permissions, and keeps working while nobody's watching.
On the customer side, the announcements pushed towards individual experience at scale: pages built and managed for a single account or persona, agents running the experimentation cycle end to end, and front-end agents composing experiences that shift mid-session to suit whoever's looking at them.
And on the agent side, the genuinely new pillar, Optimizely introduced tools to see what AI agents are actually doing on a site, support for WebMCP so browser-based agents can complete actions reliably, and a way to test whether content changes improve how often a business gets surfaced or cited by AI in the first place.
The third audience: AI agents that read, not browse
The most consequential idea from the keynote isn't a feature at all. It's the observation that AI agents are now a distinct website audience, with their own habits and their own ways of getting things wrong.
Businesses have spent two decades designing for people and, behind them, search engines. There's now a third party retrieving information and, increasingly, completing tasks on a buyer's behalf, and if a site's content isn't structured for that party to find and represent accurately, a business can end up invisible in AI-driven discovery, or worse, misrepresented when someone asks an LLM for advice.
That's not an abstract worry in B2B. Forrester puts the share of business buyers now using AI somewhere in the purchasing process at 94%. Long buying journeys, several stakeholders and complicated products are exactly the conditions where an accurate answer carries outsized weight, and exactly the conditions where a wrong or missing one does real damage before a human ever gets involved.
The first move here isn't a platform overhaul. It's an audit: what can today's LLMs and AI agents actually pull from your site, how does your company show up in their answers, and where are the content gaps creating blind spots?
Why the interface is the actual battleground
If AI agents are the new audience, the marketer's view of those agents is where the vision is won or lost. Optimizely's own product team has been unusually honest about a design trap the whole industry is walking straight into: ambient AI, the idea that the best AI is the AI you never notice.
It works nicely for consumer software, a Spotify recommendation, an autocompleted email. It works badly in the enterprise, where an invisible agent that gets something wrong leaves the user with no way to see what happened, let alone fix it.
The underlying problem has a name borrowed from cognitive science: a gulf of evaluation (can a user tell what the system is doing?) and a gulf of execution (can they tell how to make it do something different?). Optimizely's design team found that handing power users more knobs and APIs doesn't close either gulf for the other 80% of users, who came for the capability rather than the configuration. What closed the gap, in their own product, was surfacing agents where people already work and giving them a live view of what the agent is doing while it works, rather than a report afterwards.
It's a fair test to apply to anything with "agent" in the name at the moment, Optimizely's own tools included: can a marketer see, shape and trust the output, or are they simply handed a clean answer and expected to take it on faith?
From AI execution to AI judgement in experimentation
Trust in the interface is only half of the problem. The other half is knowing what to aim all that new capacity at, and perhaps the sharpest idea running through the Opticon event is what happens once AI makes doing things nearly free. Optimizely's research team calls it an execution-clarity paradox: the cost of running an experiment, drafting a variation or launching a test is collapsing towards zero, while the cost of knowing what's worth testing, and what a result actually means, keeps rising.
The evidence backs it up. Teams embedding agents across the whole experimentation lifecycle, not just for one-off drafting, are running noticeably more experiments while also seeing better win rates, according to Optimizely's own analysis of tens of thousands of Opal interactions across several hundred companies.
But the same research is candid about where most teams get stuck: without a defined process, without the tools talking to each other, and without proper governance, more AI usage just produces more inconsistent output, faster. Scarcity hasn't disappeared from experimentation programmes; it has moved from execution to judgement — defining the right metric, setting the guardrails, and knowing when to override the system rather than let it chase the nearest local optimum.
And as AI moves from individual tools to teams of agents and autonomous workflows, another challenge emerges: AI orchestration. The more AI capabilities an organisation introduces, the more important it becomes to connect them to the systems, data and workflows they depend on. Otherwise, AI remains a collection of clever capabilities rather than a coordinated way of working.
For teams running experimentation centres of excellence, the job doesn't shrink. It moves. Less time policing every test, more time deciding what's worth learning in the first place.
What agentic AI means beyond the CMS in commerce
The same argument shows up well beyond Opticon's stage, in how B2B organisations run commerce as much as content. Whether it's a Virtual Teammate drafting a campaign or an AI agent placing a reorder, the pattern holds: an autonomous system is only as trustworthy as the consistency of the decisions sitting underneath it. A pricing rule, a piece of product data or a piece of content that differs depending on which door it's approached through doesn't become safer once a machine is the one walking through it. It becomes riskier, and at machine speed.
That's the thread connecting Optimizely's AX pillar to the wider shift under way across B2B digital: agentic commerce and autonomous systems don't fix fragmentation; they industrialise whatever they're built on top of. The organisations getting real value out of AI agents, in marketing, in commerce, in service, tend to be the ones that did the less glamorous work first — one source of pricing truth, one governed product model, one place where content and context actually live.
Where this leaves digital leaders
Optimizely didn't simply add a feature at Opticon 2026. It named a new constituency for the modern website and started building towards it in earnest. The vision; Virtual Teammates doing real work, experiences composed individually at scale, a business legible to the agents now doing the shopping on a buyer's behalf is genuinely significant.
But the company's own research is the best argument for treating that vision with a degree of discipline rather than pure enthusiasm. Capability is arriving on schedule; the operating models to govern it are not. The organisations that close that gap — clear governance, one version of the truth, interfaces that keep people in the loop rather than sidelined by it — are the ones for whom all three audiences, marketer, customer and agent, end up genuinely well served.
Columbus was on the ground at Opticon New York 2026. If you'd like a second opinion on where your Optimizely estate has untapped value, or where your content currently stands with the AI agents now reading it, that's exactly the conversation we're set up to have.
