“AI-powered” is on every agency homepage right now — including ours. The difference between hype and value comes down to one question: does the AI make the experience measurably better for a real customer?
The good kind of AI on a website
The features worth building share a trait: they remove friction or answer a question faster than the customer could on their own.
- Instant answers: a grounded assistant that responds from your docs, pricing, and FAQs — no more digging.
- Smart search that understands intent, not just keywords.
- Personalization that adapts copy or offers to the visitor's context.
- Onboarding copilots that guide users through setup in real time.
The hype to avoid
If a feature exists to say “we use AI” rather than to help someone, skip it. Chatbots that can't actually answer, auto-generated content that reads like sludge, and personalization no one notices all erode trust.
Grounding is everything. AI that draws on your own content is helpful; AI that guesses is a liability.
Grounding is the whole job
Almost every AI feature that fails in production fails for the same reason: the model was asked a question about your business and had nothing authoritative to read. Grounding means retrieving the relevant passage from your own content before the model answers, and giving it no license to improvise beyond that.
In practice that means a few unglamorous things. Your content has to be structured well enough to retrieve — a PDF price list from 2023 is not a source of truth. Answers should cite where they came from, so a customer can verify and your team can debug. And there has to be a graceful failure: when retrieval finds nothing, the honest answer is a handoff to a human, not a confident guess.
Decide what it is allowed to say no to
The most common mistake is scope. A support assistant that tries to answer everything will be wrong about something that matters — billing, legal terms, medical or financial advice. Draw the boundary deliberately, write it into the system prompt, and test it. An assistant that says “I can't help with that, here's who can” is more valuable than one that is right ninety percent of the time about things it should never have touched.
How to tell whether it worked
Pick the metric before you build, not after. The useful ones are specific:
- Containment: what share of conversations ended without a human, and did those customers come back angry?
- Time to answer: how long until the visitor got what they came for, compared with search or a support ticket.
- Conversion lift on the pages where it lives — measured against pages where it does not.
- Deflection cost: tickets avoided, multiplied by what a ticket actually costs you.
If none of those moves after a month of real traffic, the feature is decoration. Turn it off and spend the budget on something else.
Budget for the second year
AI features are not a launch, they are a subscription. Models get deprecated and replaced. Your pricing, programs, and policies change, and the retrieval index has to change with them. Someone has to read the transcripts and notice that a quarter of visitors are asking something your site does not answer — which, incidentally, is the most valuable content research you will ever get for free.
The teams that get lasting value treat that review as a standing job, not a post-launch chore. The teams that do not end up with an assistant quietly giving out last year's prices.
When not to build it
If your traffic is low, your content is thin, or nobody on your team can own the thing after launch, an AI feature will cost more attention than it returns. Fix the fundamentals first — a fast site, clear pages, an obvious path to contact. AI amplifies whatever is already there; it does not substitute for it.
A quick checklist
- Is it grounded in your real data?
- Does it have a measurable job (conversion, support deflection, activation)?
- Does it fail gracefully and stay on-brand?
- Would you keep it even if you couldn't mention “AI”?
If you can answer yes to all four, build it. If not, it's probably hype.