Peter Curran, Chief Product Officer at Coveo, had a Q&A with Greg Kihlström for The Agile Brand Guide’s Expert Mode series.
For about fifteen years, the brand site was where discovery happened. Budget followed that assumption. Redesigns, content programs, conversion work — all of it premised on the customer eventually showing up. Now a growing share of buyers start with an assistant that answers before they land anywhere, and the worry inside a lot of marketing organizations is that we’ve been funding a destination that’s quietly becoming a data source for somebody else’s answer.
Peter Curran, Chief Product Officer at Coveo, declined to pick a side when we put that question to him. His argument is that both are true at once, for different buyers, and that the better question is what the site is now supposed to contain. Worth saying plainly: Coveo sells the search and relevance layer underneath these experiences, so he has a commercial stake here. Which is why the sharpest parts of his answers are the ones where he says what not to fund.
The Site Didn’t Lose Its Job. It Got a Different One.
Curran’s case against picking a side starts with how unevenly this stuff actually arrives. He pointed to the William Gibson line about the future already being here but not evenly distributed — the town that still has a corner drugstore instead of a Starbucks isn’t living in 2006, it just wasn’t touched at the same rate. Some buyers still get everything from a human sales rep. Others use the site. The assistant is a third option, and it’s moving the funnel again.
“For many years, websites were built because that’s where the funnel moved… But AI is adding a third option and is moving the funnel again. Many customers now treat the website as one piece of a larger puzzle. As a result, sites are evolving from a conversion destination or content kiosk and into a statement of the brand’s system of truth and system of action, a kind of ontological expression to help Answer Engines with faithful brand representation.”
His example is worth sitting with, because it’s the kind of question a marketing site almost never handles today: “Which hydraulic pump fits the machine we own, replaces this part number and can arrive within a week?” Answering that requires product data, entitlement data, compatibility rules and live availability, connected well enough that a machine can traverse them. For a CMO, that reframes the budget conversation. The question stops being how much to spend defending the site as a destination, and becomes whether the brand’s own data can answer a question that specific — because an assistant certainly can’t answer it if the brand can’t. That work sits closer to answer engine optimization and product information management than to anything on a typical brand-refresh roadmap, and it’s rarely owned by marketing today.
The Intent Data You Already Paid For
We asked which signals teams already have sitting in their systems and aren’t acting on. Curran named three, and none of them require a purchase. The first is the search trail — years of queries, refinements, clicks, dead ends.
“If someone searches for a product, changes the query twice and then leaves, they’ve told you where your experience failed. That should influence ranking, recommendations and even what content or products you create.”
The second is support behavior: case histories, chatbot transcripts and knowledge-base searches, all of it locked inside service systems and used only to close tickets. If customers keep asking how to fix a particular product, that’s merchandising information as much as support information — it should change what gets recommended, what gets flagged before purchase, maybe what gets sold at all. The third is context already in CRM, ERP and catalog systems: what the customer owns, what they’re entitled to, what’s compatible, what’s in stock. “None of these signals are especially exotic,” as he put it. The problem is ownership. They live in separate systems, get used for separate moments, and no single leader is accountable for the path across marketing, commerce and service. A practical first move for a CMO is smaller than a platform decision — pull last quarter’s zero-result searches and the top support cases, and see whether the two lists describe the same unsolved problem.
Let AI Handle Ambiguity. Let Rules Enforce Certainty.
We gave him the skeptic’s case directly, using his own company’s numbers. Coveo’s reported fiscal 2026 results put generative AI at roughly 13% of total SaaS annual contract value, with net expansion near 100% — real adoption, but still a minority of contracted spend at the vendor best positioned to benefit. He didn’t argue with the framing. He said he’d be worried if companies were swapping out proven systems just to move that number.
“My rule is straightforward: if an answer must be exactly right every time and can be expressed deterministically, keep it rules-based. Pricing, inventory, product eligibility, warranties, regulated disclosures and contractual terms should come from authoritative systems with explicit business logic. AI can help interpret the customer’s question or retrieve the right information, but it should not invent the policy or make an irreversible decision.”
That line is easier to draw before a pilot than after one, and drawing it is a marketing leadership job rather than an engineering one. Two more refusals came with it. Curran wouldn’t put AI in front of a customer who needs empathy or judgment — the frustrated buyer with an unusual problem, the consequential exception, the high-value purchase where somebody has to own the outcome. And he wouldn’t deploy a generative experience on top of content that’s fragmented or out of date, which he described as giving yourself “a more fluent way to be wrong.” For anyone about to approve a conversational front end for a knowledge base nobody has audited in two years, that’s the sentence to bring to the meeting.
Engagement Is the Metric to Stop Trusting
Asked what he’d defend as proof that AI is improving the customer experience, Curran started with the boring version: pick the job the customer is trying to complete, baseline it, then check whether the outcome improved. Did they find the right product or resolve the issue, and did it take fewer searches and fewer escalations? Did that show up in conversion, service cost or repeat visits? Then he named the metric he’d challenge, which is the one most AI pilots report.
“More interaction is often presented as evidence that customers find it valuable, but if someone spends five minutes going back and forth with an assistant to get an answer that should have taken thirty seconds, that is not success. It is friction dressed up as engagement.”
Engagement is the easiest number to produce and the least trustworthy, and adoption isn’t much better — it tells you people encountered the feature, not that it helped. Two things make the measurement honest. One is a control group: a comparable set of customers on the existing experience, so improvement is demonstrated rather than assumed. The other is segmenting by intent, because AI tends to help enormously on messy research questions and barely at all on a simple navigational query. Average the two together and the result looks like a modest lift that explains nothing. Any CMO reviewing an AI pilot this quarter can ask both questions in about a minute.
The first thing to do next quarter costs nothing. Build an intent map from what you already have — the most common searches, the failed ones, support cases, chatbot conversations, the questions sales reps keep fielding. Identify the ten or twenty things customers are most often trying to accomplish, then document whether each one can actually be completed today, where the information lives and how you’d know it worked. Curran’s point is that the exercise usually exposes the real problem: the answer exists but can’t be found, or two systems give conflicting answers, or nobody owns the job end to end.
And the inverse, which is the harder budget conversation. What he’d stop funding is “interface-first transformation” — another chatbot, another redesign, another content push pointed at a thin FAQ database while the underlying information stays fragmented and customer intent stays invisible. A more conversational front end doesn’t compensate for product data nobody maintains. So the question a marketing leader should be asking isn’t whether the website still matters. It’s whether the brand can answer its customers’ twenty most common questions accurately, wherever those questions get asked. As Curran puts it: “The interface should be the final decision, not the first investment.”