The easiest way to make a website sound modern is to say, "we should add AI to it."
Sometimes that is right.
Often it is early.
Not wrong. Just early in the same way buying a very good coffee machine is early if the kitchen has no sockets, no mugs, and a sink that only works when it feels included.
AI can make a website more useful. It can help visitors find the right product, answer questions from approved content, qualify an enquiry, draft the next step, or move someone towards a booking or basket.
But it does not float above the site doing clever things by itself.
It depends on what the site already knows. It depends on whether the answers are current. It depends on whether the next step is clear. It depends on what should happen when the answer is uncertain.
That is why, before we recommend adding AI to a website, we check the boring parts first.
Very often, they are not boring at all. They are the difference between an AI feature that earns its keep and one that sits on the page looking faintly expensive.
The first question is not "can we add AI?"
The first question is: what job would it do?
That job needs to be more specific than "help customers" or "improve the website".
Useful website AI usually does one of a few things:
- answers repeated questions from approved content
- guides a visitor through product or service choices
- qualifies an enquiry before it reaches the team
- prepares a first reply for review
- helps a visitor complete a booking, order, or request
- points people to the right page when navigation is not enough
Those are real jobs.
"A chatbot on every page" is not a job. It is a piece of furniture.
The distinction matters because the site may already be failing at the job without AI. If the service pages are vague, the proof is hidden, the product data is thin, or the contact form asks the wrong questions, AI will not fix that by being present.
It may just provide a more conversational way to expose the same gaps.
We check whether the site has source material worth trusting
An AI assistant can only be as useful as the material it is allowed to use.
For a normal services site, that might mean:
- service descriptions
- pricing guidance
- good-fit and poor-fit criteria
- frequently asked questions
- case studies
- process notes
- contact and response-time rules
For an ecommerce site, it might mean:
- product titles and descriptions
- variants, sizes, materials, compatibility, and stock rules
- delivery and returns policies
- pricing and availability
- reviews or buying guidance
- basket rules and checkout steps
For a booking site, it might mean:
- availability rules
- service lengths
- locations
- deposits
- cancellation policies
- follow-up instructions
If that material is current, clear, and available in text, we have something to work with.
If it is spread across an old PDF, three product spreadsheets, a service page written in 2019, and the head of the person who normally answers the phone, the first job is not AI.
The first job is getting the truth into a shape the website can use.
That is not admin decoration. It is part of the product.
We have had our own AI tools surface old project context cheerfully enough to know that "connected to your content" is not the same as "safe to trust".
Check the visitor task before adding AI
There is a simple rule here.
If a human visitor cannot understand the page, an AI assistant probably should not be asked to rescue it.
The page still needs clear headings, plain service language, useful proof, strong next steps, and forms that collect the right information.
This matters more now, not less.
Google's AI search features still rely on normal search foundations: pages need to be crawlable, important content needs to be available as text, internal links need to make content findable, and structured data should match what the visitor can actually see. ChatGPT's shopping work points in the same direction for ecommerce: current product details, price, availability, and a clean route back to the retailer-owned buying journey.
The practical lesson is not "write for AI instead of people".
It is the opposite.
Make the site clear enough that people and AI systems can both understand what is true, what matters, and what happens next.
If the page has hidden answers, vague claims, decorative copy, and a contact form acting as a bin for every possible enquiry, AI has very little solid ground.
We check where the visitor gets stuck
AI is most useful when it sits at a real point of friction.
Not theoretical friction. Actual friction.
The visitor cannot tell which product fits.
The same pricing question appears in every enquiry.
People submit forms without the information needed to quote.
Visitors bounce between service pages because the difference is unclear.
Customers ask support questions that are already answered somewhere, just not somewhere they can find.
That is where AI can help.
For example, on an art-gallery ecommerce site, a useful assistant is not just a polite box that says it can help. It needs to understand the product catalogue, the options that change the order, the materials available, and the rules for putting a configured item into the basket.
The customer does not care that this is clever. They care that they can ask a real buying question and move closer to buying the right thing.
That is the level of job we are looking for.
If the assistant can only answer generic questions that should have been one paragraph on the page, the site probably needs better content before it needs AI.