The testimonial block arrives before anyone has asked a customer
A generated business website does not come with an empty testimonial section — it comes with three finished quotes, a row of client logos and a rating, written in the same minute as the rest of the page. The quotes are fluent, specific enough to sound personal, and placed next to the contact button, exactly where proof is supposed to sit.
This is the tool doing its job. A testimonial is a pattern it has seen thousands of times, so it reproduces the structure competently: a name, a company, a short sentence about a partnership that changed something. What it cannot do is ask your last five customers what they actually thought, so instead of an answer it supplies a plausible one.
Nothing on the page marks the difference. There is no footnote, no asterisk, no label reading “example”. The section looks finished, and finished-looking is precisely how a placeholder survives to launch day.
A visitor cannot tell, and that is the problem
To someone comparing three suppliers in ten minutes, two well-written quotes with ordinary names are indistinguishable from two real ones — so the fabricated version is never caught by the person it is meant to convince. It is caught by people who cannot be convinced: a customer who recognises their own company under a sentence they never wrote, a competitor who opens the same page, a journalist checking whether the referenced project exists.
Then there is the quieter failure, where the quote is close to true but sharpened. A customer who said the new site made things easier for the team becomes a customer whose enquiries doubled. The customer reads it, remembers saying something much more modest, and the next conversation opens with an apology instead of a new project.
Distorted proof is not a grey area either. Presenting invented or adjusted reviews as the experience of real customers falls under consumer protection rules that treat such practices as unfair across the EU, and a business showing reviews is expected to be able to say whether and how it checks that they come from people who actually bought. Review platforms remove what they detect, and the removal is public.
Proof is collected, not generated
There is no prompt that produces a quote from a customer who has not given one; the only method that works is asking while the work is still fresh in memory. The right moment is the days after handover, when the customer has seen the result but has not yet moved on to the next problem.
The routine itself is small. Ask one question — what is easier now that the project is done — and offer two ways to answer: a short written statement, or a review on the platform where your future customers already look. Then ask explicitly for permission to publish the answer with the person’s name, role and company. Permission is the step most companies skip, and it is the step that turns a polite email into usable proof.
Quote people verbatim. Fix a typo if you must, but do not upgrade the adjective, and never let a tool “make it sound better” — a stronger version of somebody else’s sentence is a claim they never made. Keep one list of approved proof: who said it, when, in which language, and where it may appear. That list takes an afternoon to start and saves an argument later.
A testimonial page nobody opens is decoration
The same sentence does very different work depending on where it sits: on a dedicated references page it is read by nobody, beside the service it describes it is read at the moment the visitor is deciding. Put one specific quote next to the contact form of the service it describes, another where the process or pricing questions are answered, and a third where a hesitant visitor scrolls back up.
Specificity is what makes the placement pay off. A quote with a name, a company, a situation and a result answers a doubt; a quote with only a first name and a five-star row raises one. Freshness does the same job: proof without a date drifts, and a visitor cannot tell a project from last month from one from five years ago. Logos follow the same rule — only real clients, only with permission, only clients who are still clients.
What AI does well once the material exists
With a list of real, approved quotes, the tedious half of the work becomes AI work, and it is done faster than anyone in the company would do it alone. It drafts the request message, adapts the wording for a customer who bought a small job and one who bought a multi-year project, and follows up with those who never answered.
It also reads what you already have. Feedback emails, survey answers and handover notes contain the customers’ own phrases, which beat any adjective a marketer would write — pulling those out, grouping them by service and proposing which sentence fits which page is exactly the kind of sorting a model does in seconds. It translates a Czech quote for the English version of the site while keeping the name and company intact, keeps the formatting consistent across pages, and audits the live site against the approvals list, flagging any quote that has no entry behind it.
One question before the page goes live
Open every quote, name and logo on the site and ask one question: where is the message in which this person agreed to it? If you can open it, the section is proof. If the honest answer is that the generator wrote it, the section only looks like proof, and everyone who matters eventually finds out.
Then run three checks. Does every claim inside a quote exist in the original wording, without a sharpened verb or a doubled number? Does the logo bar contain only current clients who said yes to being shown? Is anything on the page presented as a verified review when no verification actually happens?
Proof is the one part of a website that cannot be generated, only gathered. Gather it while the project is fresh, place it where the decision is made, and let AI do everything around it — the site will carry evidence nobody has to apologise for.
Related service: Web design and content strategy