From the field
What Do You Sell When AI Can Build the Website?
I used to make good money building websites. Years later, after my online business declined, I was back on LinkedIn applying for jobs—and watching AI produce work I used to charge for. Here’s what that experience taught me…
Alexander Balanoff
13 min read
I remember sitting in a small Airbnb in north London, with LinkedIn’s jobs section open on my laptop.
I was applying for positions offering an annual salary comparable to what my online business had brought in during two or three good months.
That comparison was unfair. Business revenue comes with expenses and risks that a salary figure doesn’t capture. I understood the distinction. It still stung.
Over roughly eight years, my online publishing business had generated more than US$1 million in cumulative revenue, before costs and tax. Much of it came from advertising, with referral income added later.
Now I was trying to persuade someone to interview me.
It was difficult to consider the next step without comparing it with the best months of the previous eight years. An employer, meanwhile, needed evidence that I could do the job they were hiring for.
“I used to earn more than this” is a weak opening line for a job application.
I had also made good money building websites. That was work I understood. Watching AI produce pages from a reference image and a prompt made the obvious fallback feel considerably less comfortable.
If someone could generate the website themselves, what would they be paying me for? Rebuilding my own site gave me a practical way to investigate that question.
AI wasn’t the reason I was back on LinkedIn
My publishing business relied heavily on Google for visitors and AdSense for revenue. I saw search performance weaken in 2023 and associated the decline with Google’s updates. I’ve written separately about building the business and watching it decline.
That experience exposed my dependence on platforms I did not control. It does not establish that AI caused the decline.
The career question came afterwards. I needed to translate years of running websites into work someone wanted to buy now. Past success gave me experience to draw on. It did not settle the conversation.
The craft still matters. The first version is easier to get.
A website once brought together several jobs clients struggled to do themselves: interpreting a brief, designing the layout, preparing images, writing code, configuring a content management system and getting it online.
Typography, spacing and navigation made a business look credible and its information easier to use. I enjoyed getting those details right.
Templates and website builders widened access. AI has made another part of production more accessible: describe the page, supply a visual reference, generate a version and ask for changes.
In Anthropic’s account of how its teams use Claude Code, product designers describe turning screenshots into interactive prototypes. It is a vendor’s account of its own work, rather than an independent comparison of finished websites. The distinction matters, but the capability is relevant to the services I used to sell.
Some clients can produce an adequate simple site themselves. My proposal has to explain what additional work their particular project requires.
My homepage became the test case
I supplied a Pinterest reference, my portrait and a brief for a website promoting ecommerce, SEO and growth services. I requested two visual concepts and chose one.
Then I started making decisions. I rejected a lime-green revision and returned to orange and black. I reordered services, removed the selected-work section and gave testimonials more attention. The main call to action needed to open an enquiry form.
It also had to become a WordPress theme while retaining my About page, Now page, blog and existing posts. Mobile layouts and article sidebars needed further work.
Those choices had different purposes. Colour was a preference. Keeping existing content was a requirement. Making it possible to contact me was the reason for building the page.
Before asking for another version, write down what the current one fails to help a visitor do. If you cannot name the problem, you may just be collecting options. Decide what the visitor needs to understand, what they should do next and how you will check it. That gives your experience a job.
Alexander Balanoff
The page looked finished. My inbox stayed empty.
The contact form exposed the gap between a page that looked ready and a process that was working.
Enquiries were being saved in Flamingo, the message-storage plugin, but the notification emails were not arriving in my inbox. A visitor could submit a message. The stored record existed. The handover to me was failing.
After the sending setup and IP authorisation were addressed, I received all three test emails.
That gave me a specific result: three test messages received. It demonstrated delivery under the tested conditions. I do not have a measured sales uplift to attribute to that change.
The useful discovery was where the failure occurred. Rewriting the headline or changing the button colour would have left it there. Someone needed to trace the enquiry beyond the page.
AI can help with that investigation, too. GitHub’s guidance on reviewing generated code covers functionality, security, project fit and maintainability. Those checks belong in the delivery, including when AI produces the first version.
Sell a defined piece of work someone can assess
Consider an online shop asking for a redesign. This is a hypothetical example. Shoppers might be leaving because delivery charges appear late, product questions remain unanswered or the advertising attracts unsuitable visitors. Each explanation suggests different work.
I would start by agreeing a small investigation: review one buying journey, relevant customer questions and the available sales data. The deliverable would identify observed problems, plausible explanations and the next change worth testing.
For a service business, an enquiry-journey review could cover one landing page, one form, its record storage, notifications and the handover to whoever replies. The client would receive documented findings, priorities and a scope for any implementation.
That is an offer a buyer can question and compare. It also gives me boundaries: how much I will inspect, what access I need and what the fee includes. A promise to “grow the business” leaves too much undefined.
Price the uncertainty before you price the page. A clean brochure site and a fragile checkout can both arrive as “we need a website”. They need different investigation and responsibility. Separate diagnosis from delivery, explain the boundaries and agree what successful completion looks like. You can quote the work more honestly when you understand it.
Alexander Balanoff
What I would tell my design students now
I used to teach a full-time multimedia course at Natcoll in Auckland. Students asked where to apply for jobs. Today I would spend more time helping them explain why someone should hire them: the problem they investigated, their own contribution, a difficult decision and the result.
There is still work to prepare for. The US Bureau of Labor Statistics projects 5% employment growth for web developers and digital designers from 2025 to 2035. It also says AI may soften growth through productivity gains and by allowing people in other occupations to perform basic development tasks. That US forecast cannot establish a UK freelance rate or an individual’s prospects.
I would treat web design after AI as a reason to strengthen the work around production: investigation, judgement, testing, maintenance and a service someone understands.
10 skills I would develop—and something to show for each
Start with the three most relevant to one service you can already partly deliver. Build evidence as you learn.
1. Turn a vague request into a testable problem
Read support enquiries and speak to customers where possible. Produce a one-page brief separating observations, assumptions and unanswered questions. “Visitors cannot find delivery costs” gives you something to investigate. “The website needs more impact” needs unpacking.
2. Understand what a worthwhile customer is worth
Learn contribution margin: what remains from a sale after its variable costs. Include product, fulfilment, payment and acquisition costs where relevant. Build a simple calculation using the client’s figures. For a service business, follow enquiries through to qualified opportunities and paying customers.
3. Make your AI workflow repeatable
Take a small project from brief to delivery using one tool. Keep requirements, changes and checks. Use version control to inspect and reverse edits. Record correction time as well as generation time. Finish with a project you can explain and maintain.
4. Keep practising design judgement
Compare layouts through hierarchy, readability, interaction and accessibility. Explain how your choice helps someone complete a task. Include mobile, error and empty states in your portfolio. Those reveal decisions a polished desktop screenshot cannot show.
5. Write copy that resolves a buying question
Identify questions that stop someone proceeding: suitability, price, delivery, evidence or what happens next. Rewrite one page around them. Ask people to explain the offer back to you. Record what they understood and where they hesitated before claiming better conversion.
6. Measure something closer to money
Record a baseline and verify relevant analytics events, with appropriate consent. Distinguish submissions from qualified enquiries and orders from profitable sales. Note changes in traffic, offers and advertising. Separate what changed from what you can reasonably attribute to your work.
7. Protect discoverability when a website changes
Learn crawling, indexing, redirects, canonical URLs and internal linking. Inventory existing URLs before a rebuild, including pages with useful traffic or links. Produce a migration plan showing what stays, moves or disappears. Check it against the released site.
8. Test the complete customer journey
Check mobile layouts, keyboard navigation, form errors, submission and follow-up. Record expected and actual results. Include what happens after an enquiry arrives. Bring in specialist help for work you cannot assess reliably. The worksheet below gives you a starting point.
9. Build depth around a recurring problem
Choose something you can investigate repeatedly: ecommerce migrations, enquiry handling or complex product information. Learn the systems and customer questions around it. Document your role, constraints, decisions and evidence. Label demonstration projects clearly so readers can distinguish them from client results.
10. Package work a client can say yes to
Specify the problem, deliverables, boundaries, fee and completion criteria. Agree separately whether an investigation includes implementation. Publish a worked example and share it with relevant contacts. Practise explaining the offer in a paragraph a buyer can understand without a sales call.
Use this worksheet before commissioning another redesign
On your own site, or with the owner’s permission, trace one clearly labelled test enquiry. Record what happens at each handover. This is a focused functional check, rather than a complete accessibility or security audit.
| Check | What to do | Evidence to record |
|---|---|---|
| 1. Find the next action | Open the landing page on a phone. Identify the offer and how to enquire. | Page URL, device and any point where the next step is unclear. |
| 2. Submit a valid enquiry | Use a unique test label and complete the form with contact details you control. | Submission time, test label and the confirmation shown to the visitor. |
| 3. Follow the message | Find the matching record where storage is configured. Check the intended recipient’s inbox and any promised acknowledgement. | Which records and emails arrived, where they arrived and their timestamps. |
| 4. Try an invalid submission | Leave a required field blank. Check that the error can be understood and corrected using the keyboard. | Error wording, focus behaviour and whether entered information is retained. |
| 5. Complete the handover | Confirm who responds and send a reply to your test contact. | Reply received, responsible person and when the journey will be tested again. |
Label each recommendation accurately: an observed failure needs correcting; an untested idea needs investigation; a visual preference needs agreement. This helps a client decide what to fund first. You can use the same record to check the work after delivery.
Where I would start in the next 30 days
In week one, I would choose a recurring problem and speak to three people who encounter it. In week two, I would investigate one example with permission and agree a small scope with clear completion checks.
In week three, I would implement and test the change. In week four, I would write up the evidence and offer a defined paid pilot to similar businesses. A first case study might establish that a broken process now works; increased revenue would need its own measurement.
Let the project expose the next skill you need. Trying to become an expert in everything before speaking to a customer can postpone the most useful part of learning.
The next client has a reason to hire the person I am now
Back in that Airbnb, I was comparing opportunities with the best months of a business that had already changed. Understanding the difference between revenue and salary did very little for my pride.
I would still feel the loss. But I would spend less of the application explaining my old income and more of it showing current work: a problem investigated, a decision made and a result checked.
Three emails arriving in my inbox sounds modest beside a million-dollar revenue figure. It is also a result I can explain, with a clear connection to why the website exists.
I would keep building websites. I would make the next proposal about something the client needs working, and show why they can trust me to deliver it.
The difficult questions about web design after AI
Can I still charge a substantial fee if AI makes the work faster?
Yes, when the agreed scope, quality and responsibility justify it. A fixed fee can cover investigation, delivery, testing and support. Explain what the client is buying and how AI fits the process. If billing hourly, bill the hours actually worked. Faster production does not justify concealing the process or claiming results you have not measured.
If AI can suggest strategy too, what advantage do I have?
Calling yourself a strategist provides little protection. Your advantage needs evidence: understanding the actual customer problem, working within the business’s constraints and implementing and assessing decisions. AI can assist throughout. If your whole service is a generic plan produced from a short brief, I would question how defensible it is.
Do I still need to learn code if AI can write it?
Learn enough to assess the work you accept responsibility for. A simple site requires understanding structure, styling, forms and basic debugging. More complex software needs deeper skills or a developer who has them. If you cannot explain, test or recover from a change, reduce the scope or bring in help before shipping it.
What if a client arrives with an AI-built website and wants me to finish it?
Assess it before quoting. Check access, rights to supplied assets, dependencies, maintainability, backups and important user journeys. Identify what can be retained and what needs repair or replacement. Offer that assessment as a defined service. AI’s involvement alone does not establish that the site needs rebuilding or that it is nearly finished.
How do I prove value with too little traffic for an A/B test?
Match the evidence to the claim. Demonstrating that an enquiry reaches its recipient or that observed participants understand an offer is useful. It does not establish a reliable conversion uplift. Keep measuring qualified enquiries and sales, note other changes and state the limits of the sample. Functional improvements and commercial outcomes are different findings.
Should I take a job, keep freelancing or rebrand as an AI consultant?
Start with your obligations, available work and current competence. A job can provide stability while you develop a service. Freelancing needs time for sales as well as delivery. An AI label helps when it describes a problem you can solve. Choose a workable next step without making it a verdict on your entire career.
Your website is built. Is the enquiry journey working?
If your website looks ready but you are unsure what happens between a visit and a useful enquiry, bring me the URL and the problem. We can agree a focused review of the page, form and handover, then scope the work the findings support.
My work connects ecommerce, SEO and growth consulting with how the website actually operates. Scope and fee are agreed before work begins.