What an AI website builder can and cannot do

An AI website builder takes a description and returns a draft: sections, layout, copy, sometimes images. The useful way to think about it is as a fast, tireless junior drafter with no knowledge of your business beyond what you typed. It is very good at turning a brief into a structure, and it is unreliable at anything that needs to be true about the real world. This page sets out where that line falls and what the person publishing must do.

No particular product is assumed here. The points apply to any tool that generates a site or its content from a prompt.

The model: brief in, structure out

Everything a builder produces is derived from three inputs: your brief, patterns it has seen in general, and whatever you upload. Anything it says about your business that is not in the first or third is a guess presented as a statement. That gives a simple rule for sorting jobs.

The reliability column is a working judgment from how these tools are commonly used, not a measured benchmark. Test your own tool against the same table before trusting it.

Where drafts are strong

• Structure. A tool is good at deciding that a local service site needs a hero, a short description of services, proof, a contact route and a footer, and at ordering them sensibly. You get in minutes what would take an hour of wireframing.
• Layout and responsive behaviour. Grids, spacing and mobile stacking are well-understood patterns, and errors are visible to the eye.
• Alternatives. Asking for three versions of a headline or section is cheap, and comparing options is easier than staring at a blank page.
• Boilerplate. Standard pages, FAQ shells and form markup save repetitive work.

Where drafts fail

Invented specifics

Google's guidance on generative AI says that generative models do not retrieve facts but predict a likely sequence of words based on their training data, so outputs may contain inaccuracies, and it calls manual fact-checking and review of all AI-generated content before publishing critical. On a business site, the typical inventions are:

• opening hours, addresses, phone numbers and service areas
• founding dates, "family-owned" claims and how long a business has operated
• licences, certifications, memberships and awards
• testimonials and star ratings
• statistics ("trusted by thousands") and guarantees
• process claims that sound reasonable but are not how you work

None of these look like errors on the page. They read as confident, normal copy, which is what makes them dangerous. A false credential or guarantee can also create real legal and reputational exposure.

Generic copy

If the brief is vague, the copy will be interchangeable with a thousand others: "tailored solutions", "your trusted partner", "quality you can rely on". A quick test is to swap your business name for a competitor's. If the text still reads correctly, it says nothing about you. The cure is to supply specifics, which is the subject of writing a prompt that produces a better first draft.

Imagery

Generated or stock images can contain odd anatomy, unreadable text, or products and places that do not exist. An image that implies a real staff member, a real shop or a real result is a claim, even if no sentence says it. Alt text produced automatically is often a literal description that misses the purpose of the image. WCAG 1.1.1 asks for a text alternative that serves the equivalent purpose, which is a different test from describing what is visible.

Accessibility gaps

A page that looks fine can fail people using screen readers, keyboards, magnification or low vision. Typical gaps in generated sites include:

• headings chosen for their size instead of their structure (WCAG 1.3.1, Info and Relationships, asks that structure conveyed through presentation can be determined programmatically); skipped heading levels are not named as a failure in WCAG itself, but the W3C's writing tips advise tagging headings properly because assistive technology users navigate by them
• text and background combinations below the contrast minimums (WCAG 1.4.3 requires at least 4.5:1, or 3:1 for large-scale text)
• form fields without labels or instructions (WCAG 3.3.2)
• links that read "click here" or "read more", which the W3C's writing tips advise against, and which relates to WCAG 2.4.4 on link purpose
• no visible keyboard focus (WCAG 2.4.7), or a focus order that does not preserve meaning (2.4.3)
• motion that starts automatically, lasts more than five seconds and runs alongside other content, with no way to pause, stop or hide it (WCAG 2.2.2)

Rights and permissions

General information only, not legal advice. Whether and how copyright protects AI-generated material differs by country and is not settled everywhere. Canada's own 2023 to 2024 consultation on copyright and generative AI produced a "what we heard" report from Innovation, Science and Economic Development Canada. It says existing Canadian copyright case law suggests authorship must be attributed to a human who exercises skill and judgment, that participants generally supported keeping human authorship central to protection, and, in a footnote, that Canadian courts had not yet commented specifically on AI authorship and ownership. That report summarises views and does not change the law, so read it as background and ask a lawyer about your case. The tool's own terms govern what you may do with its output. Separately:

• Do not publish generated text or images that imitate a real brand, a named person or a specific artist.
• Keep records of the material you supplied (your photos, your words), because that is the part you can most clearly claim.
• A form that collects personal information can bring privacy obligations. In Canada the federal privacy law PIPEDA may apply, depending on the organisation and what it collects; the Privacy Commissioner's site offers a privacy guide for businesses and a tool for finding which privacy law applies. Email marketing has its own law, Canada's anti-spam legislation (CASL), which the federal government describes as focused on commercial electronic messages. Write the privacy notice for what the site really does, not from a generic template.

Search

Google's guidance says generative AI can be useful when researching a topic and to add structure to original content. It also says that using such tools to generate many pages without adding value for users may violate its spam policy on scaled content abuse. The spam policy defines that abuse as many pages generated mainly to manipulate rankings rather than help users, "no matter how it's created", so the test is the purpose and the value of the pages, not the tool. The guidance also says to review metadata, including title elements, meta descriptions, structured data and image alt text, since these can appear in search results, and suggests that sharing how content was created can give readers context. A site with a handful of pages you have checked is in a different position from a site with thousands generated to fill keywords. See titles, descriptions and snippets for the metadata review.

What a person must still do

1. Supply the truth. Write the facts in the brief, and mark everything you did not supply as "leave a placeholder".
2. Decide the claim. For every sentence that asserts something, ask whether you can prove it on request.
3. Edit for voice. Read it aloud. Cut anything that could be about any business.
4. Check the technical foundation. View the page without JavaScript, on a phone, on a slow connection. A generated site can be heavy; see Core Web Vitals in plain language for what to measure and static HTML first for why complete HTML matters.
5. Test the real journeys. Submit the contact form, follow every link, and check the thank-you state.
6. Own the result. Someone must be named as responsible for what the site says.

A mechanical check for invented numbers

Reading carefully catches most invented facts, but a tired reader skips numbers. A cheap mechanical pass lists every figure in the draft that does not appear in the brief. Save the brief and the draft as plain text files and run these in a terminal (comm -23 prints the lines found only in the first, sorted list):

grep -oE '[0-9]+([,.:][0-9]+)*' brief.txt | sort -u > brief-nums.txt
grep -oE '[0-9]+([,.:][0-9]+)*' draft.txt | sort -u | comm -23 - brief-nums.txt

On the fictional bakery below, with a brief that mentions "3 days notice" and a draft that adds a founding year, opening hours and a customer count, the second command prints three lines: 1987, 5,000 and 7. Each is a claim the owner never supplied. The method has limits. It checks digits only, so "award-winning" and "family-owned" pass straight through, and a figure that happens to match something else in the brief (a "3" that means three stars rather than three days) also passes. Treat it as a first filter and do the human read afterwards. The same idea works for names: list the capitalised words in the draft and check each against the brief.

Reviewing a draft for a fictional bakery

A fictional draft for "Harbour Bakery" at example.com includes the line "Family-owned since 1987, award-winning sourdough, open daily 7 to 7." The owner supplied only the bakery's name, its neighbourhood and the product list. A review pass finds:

Count what the review found. The draft's single line makes four factual or quasi-factual assertions (a founding year, ownership type, an award, opening hours), and the owner supplied none of them, so the supplied-to-asserted ratio on that line is 0 of 4. As a rule of thumb (an editorial judgment, not a measured threshold), if a page has twenty such sentences and fewer than half trace to the brief, write the facts down first and regenerate instead of repairing sentence by sentence.

Pre-publish checklist

• Every name, address, opening hour and contact method has been checked against a real source.
• No credential, award, testimonial, statistic or guarantee remains unless it can be proved.
• Each claim sentence has a person who can stand behind it.
• Copy has been edited for voice and specifics; interchangeable lines are gone.
• Every image is a real, licensed or generated image you are comfortable having associated with the business, with alt text that states its purpose or an empty alt if decorative.
• Heading order, contrast, labels, keyboard focus and motion controls have been tested by hand.
• Title, description and structured data are accurate and unique per page.
• The privacy notice, forms and email sign-ups match what the site actually does.
• The site works without JavaScript for its core content and loads acceptably on a phone.
• Someone has recorded who approved it and when.

Keep this

• A builder drafts structure well and facts badly.
• The brief decides how specific the output can be.
• Anything checkable must be checked by a person.
• Accessibility, rights and privacy are your responsibility, not the tool's.