UXofAI

We understand why this happens. A chatbot is visible, and it's easy to point at in a board deck. This gives the organization the satisfying feeling of having made something tangible. If AI progress is being measured by what customers can see, a chatbot looks like progress. It usually isn't.

Having a visible AI feature is not going to give your brand real value. The most important work is the work that actually moves the needle for your business. It happens behind the scenes in the unglamorous process of rebuilding your content operations, internal search, recommendation infrastructure, and customer data plumbing. The brands who went chatbot-first often spent 18 months learning that lesson the hard way.

This is the first and most important decision in any AI roadmap. Not "should we use AI," but "where should we use it first."

If you're starting from zero, start from the invisible ring and work your way out. Each ring addresses a specific part of the ecosystem, and most of the work is less about AI than about getting the underlying material in shape so AI can do something useful with it.

Editor's noteThe full "putting this into practice" text for each ring opens in the map above when you select a ring. It's the same copy as the print page, so nothing is lost on the Quick track.

The first question should be a brand question: what does your brand sound like when it's not you?

For most of branding history, this was answered by hiring good copywriters, writing a voice guide, and trusting that the humans representing the brand had judgment. The guide lived in a PDF nobody opened, but the brand survived because those humans could read a room. They knew when to be warm and when to be brisk. AI has none of that judgment by default. It only has the judgment you give it.

A generic large language model has a voice. However, that voice is the statistical average of the internet, which is to say, it's the voice of everyone and therefore nobody. It apologizes too much, and sounds like a junior employee trying very hard not to get fired. If a brand deploys that voice and puts its logo on it, the brand now sounds like that too. Every customer interaction becomes a chance for the brand to sound exactly like every other brand doing the same thing.

Training a brand AI is mostly a content and judgment exercise, because the technology itself is a commodity.

Start by auditing what the brand has actually said in the last 24 months across all channels. Customer service transcripts, social responses, email templates, product copy, press responses. This body of work is the raw material for voice training whether or not anyone formalizes it. If it's inconsistent, and it almost always is, that inconsistency will show up in the AI. The AI will reproduce whatever it's trained on, ergo, inconsistent source material will produce inconsistent output.

Then build three kinds of guardrails:

Content guardrails, which govern what the AI knows. Restrict the model to a curated knowledge base rather than the open internet.

Behavioral guardrails, which govern how the AI responds under pressure. Pre-write the hardest scenarios. What happens when a customer is angry? What happens when someone tries to get the AI to say something embarrassing for a screenshot? Issues like these happen every day when dealing with a live AI model.

Escalation guardrails, which govern when the AI hands off to a human. The best brands know their limits and route customers to real people gracefully.

They're cranking out text, images, and translations faster than ever before, and while that's fine, it's a surface-level approach to everything AI has to offer. If you focus entirely on what your AI can say rather than what it can actually do for your customers, your brand is missing the bigger opportunity, which is to build utility.

The shift from content brand to service brand shows up most concretely in how websites are structured.

In practice, this means fewer top-level categories, shorter menus, and a much more prominent input layer.

The product page becomes a decision page. Instead of a detailed specs-and-features layout that assumes the customer already knows what they want, the page asks what they're trying to accomplish and then configures itself around that. A paint brand's product page doesn't lead with thirty color swatches, it leads with "What room are you painting, and what kind of feeling do you want in it?"

The FAQ disappears. Service-oriented sites don't need a wall of static text covering edge cases because the entire interface dynamically answers questions as they arise.

The problem is that most personalization initiatives underdeliver, and if this happened at the same rate in any other discipline, it would be a scandal. The brands that get it right see meaningful revenue gains.12 The brands that get it wrong spend a lot of money making their customers uncomfortable.

The difference between those two outcomes has almost nothing to do with technology. It comes down to trust. Personalization works when your customer believes you're using what you know about them in their interest. The moment that belief breaks, personalization starts to feel like surveillance, and it doesn't matter whether you technically complied with privacy regulations. Legal permission is not the same as customer comfort.

AI makes this harder, not easier. An AI-powered personalization engine can infer mood from session behavior, predict future purchases, and customize messaging in real time. When it works, it feels like great service. When it misses, it feels invasive in a way that's harder to recover from.

Not all personalization carries the same trust cost. Some forms of it customers welcome. Some they tolerate. Some they quietly resent. Knowing the difference is most of the work.

Get explicit about what you collect and why. If you haven't already, write down, in plain language, what customer data you actually use and what value the customer gets in return. This should be a legible, public document, not a legal disclosure. Customers are increasingly willing to share data with brands that can clearly articulate the exchange, and by the same token, they're increasingly skeptical of brands that can't.

Move the consent conversation up. The standard industry practice is to treat consent as a checkbox at the bottom of a signup flow, when brands should really treat it as an ongoing, visible part of the experience. This includes a small indicator showing what the brand is currently using to personalize, a "why am I seeing this?" link that actually explains, and a clear way to turn specific kinds of personalization off without losing the account. These may feel like small UX details, but they're actually the most important personalization features a brand can build. Let customers see what you think you know about them.

For decades, companies have felt the control of building their own website experience, but customer behavior is shifting dramatically. The first (and sometimes only) interaction a customer has with a brand increasingly happens inside general-purpose AI assistants like ChatGPT, Claude, Perplexity, and Gemini. ChatGPT alone draws over 800 million weekly users who use the tool to research purchases and look for recommendations,3 and if your brand is not prepared for this shift, your customers may never find you.

The word "agent" has been doing a lot of work in AI marketing lately, and it means at least three different things depending on who's saying it. Worth separating them, because they have very different strategic implications.

For most brands, the second and third categories will matter far more than the first. Almost none of the playbooks circulating right now are helping brands prepare for them.

This changes what matters about a brand's digital presence. A decade of web strategy has been about getting customers to your site. The next decade is partly about making sure your brand shows up well when the customer never visits at all.

This question comes up in nearly every conversation we have with marketing leaders, and the answers are evolving quickly. Here is the current landscape:

Generative Engine Optimization (GEO) shows early evidence that a few techniques work better than others. (You may also hear this work called Answer Engine Optimization, or AEO, particularly when the focus is Google's AI Overviews rather than conversational AI assistants. The tactics overlap substantially. We use GEO as the umbrella term because it covers the broader set of surfaces this playbook addresses.)

What the research shows works. Princeton research tested a range of optimization tactics and found three that consistently increased AI citation rates by roughly 30 to 40 percent:4

• Citing credible sources within your own content. Pages that reference authoritative third parties are cited more often than pages that don't.
• Adding specific statistics and data points. AI systems disproportionately cite content with concrete numbers.
• Including direct quotations from experts. Content with attributed expert voices lands in more answers than content without.

The deeper pattern: AI systems are trained to look for content that reads like a verified source rather than marketing copy. The brands that show up most often in AI responses are the ones publishing content that reads like journalism, not like advertising.

The earned media pattern. This is the finding that surprises marketing leaders most. Industry research tracking AI citations has found that AI platforms cite third-party earned media (reviews, press coverage, analyst reports, independent content) 69 to 82 percent of the time, compared to 36 to 45 percent for traditional Google search.5 Your own website is not your best AI visibility asset. The coverage other credible sources produce about you matters more. This means influencing AI visibility is partly a PR and communications investment, not just a content one.

A note on timelines. AI visibility tends to respond faster than traditional search, though it also decays faster. Realistic expectations, based on current industry data:6

AI citation patterns are less stable than search rankings. Roughly half of all content cited in AI answers is less than three months old.7 Freshness matters more than in traditional SEO, and content requires more frequent maintenance to stay cited.

One caveat. This space is moving faster than any discipline we've worked in. The techniques that work today may work differently in twelve months. Major AI platforms update their models regularly, and each update can shift citation patterns overnight. Which means we need to treat AI visibility as an iterative program with monthly monitoring, rather than a project with a defined end state.

Monitor how your brand shows up, literally. Ask ChatGPT, Perplexity, Claude, and Gemini the questions your customers are asking about your category. See what they say about your brand. See what they say about your competitors. Do this monthly, with a running record. The first time most brands do this, they find surprises.

If your reports show that engagement is up and the team is moving faster, but you can't actually isolate AI's contribution from the rest of your marketing, it'll be difficult to defend that budget next quarter.

While metrics like page views, time spent on the site, click-through rates, and conversion funnels still matter, they were not built to measure what AI actually changes. If you force your AI initiatives onto those metrics, you'll get incomplete, even misleading, results.

Build a measurement plan before the AI investment launches. The number one reason AI measurement fails is that nobody decided what success looked like before the program launched. Retrofitting metrics to a live program is how "we think it's working" becomes the answer. Every AI initiative should have measurable targets in at least two of the four dimensions, defined up front, with a baseline established before launch.

Measure AI against what would have happened without it, not against last quarter. The most common measurement mistake is comparing post-AI results to pre-AI results without accounting for everything else that changed. A 15% improvement in conversion might be because of the AI, or it might be the new homepage, or the market. Isolation matters. For any AI investment big enough to justify serious measurement, running a holdout, a staged rollout, or a proper A/B test is the only way to prove what is actually driving the results.

Make risk a standing agenda item, not a quarterly review topic. The efficiency, experience, and revenue dimensions can survive quarterly review cycles. But the risk dimension cannot. An AI system operating at scale can generate weeks of exposure before anyone notices if risk metrics aren't being tracked continuously. This sounds like over-engineering until the first incident happens, so we recommend getting into practice now.