Send one more question before you publish it. The chatbot didn’t make your customer dishonest — it stripped the specifics out of what they meant to say, and a testimonial with no specifics doesn’t persuade anybody. The repair takes one reply, and their answer is the testimonial you wanted.
The short version:
- Fluent generic praise doesn’t convert, because there’s nothing in it a stranger can weigh.
- A chatbot fills a blank box, and “could you write us a testimonial?” is a blank box.
- Short fields with hard character caps make a chatbot pointless, since nobody opens one to write fifteen words.
- Confirm any number, timeframe, or superlative you didn’t watch the customer type, before it goes on a page.
- Skip the AI detectors. Accusing a real customer of faking praise they wrote costs far more than publishing a bland quote.
- For the one already in your inbox, one specific question fixes it, and the reply is what you publish.
Why a customer reaches for a chatbot
A customer who agrees to give you a testimonial has agreed to a favour, not a task. What arrives in their inbox decides which one it turns into. “Could you write us a testimonial?” with no further instruction is a blank page, and a blank page is a writing assignment: no length, no subject, no way to tell when it’s finished. Busy people who genuinely want to help do what anyone does with an unwanted writing assignment. They delegate it.
That’s the whole mechanism, and it’s worth being clear about who set it running. Your customer didn’t decide to send you machine-written praise. They decided to clear an awkward job off their desk in the two minutes they had for it. The reason most testimonial requests stall is the reason this one got outsourced: the ask was harder than the favour.
One case deserves the opposite reaction. A customer writing in their second language may be using a chatbot as a translator and a proofreader rather than as an author. The opinion, the specifics, and the experience are all theirs; only the grammar is borrowed. That’s not the problem described here, and treating it as one will cost you a good testimonial and possibly a good relationship.
What a machine-smoothed testimonial stops being
Here’s the part that actually costs you money.
A testimonial doesn’t persuade because it’s positive. Your own marketing is already positive, and it’s free. A testimonial persuades because it contains things that could have been otherwise: a starting situation that was genuinely bad, a doubt the customer named, a number they measured, a moment where the thing worked. A stranger can weigh those. They can imagine the version where it went the other way.
Fluent generic praise can’t be weighed, because there’s no version of it that would have come out differently. “Working with them was a seamless experience from start to finish, and I would recommend them to anyone” is compatible with every supplier who has ever existed. It isn’t false. It just carries no information, which means it does no work, which means the slot it occupies on your page is now empty — and that slot was supposed to be earning you customers.
The same failure shows up in testimonials nobody used a chatbot for, which is why real testimonials so often read as fake. Sameness reads as manufacture whatever produced it.
Confirm the numbers before they go on a page
There’s one thing worth checking every time, and it’s the reason a polished draft deserves a second look rather than a straight publish.
Your customer tells the model that you saved them time. Fluent marketing prose doesn’t say “saved us time”, so the model writes “saved us over twenty hours a week”. Nobody lied. Your customer didn’t choose that number, didn’t check it, and will sign it off without noticing, because it reads like the sort of thing that belongs in a testimonial.
Now you have a specific, quantified claim on your pricing page, in quotation marks, with a real person’s name under it — and no way to back it up. The first prospect who asks about it in a sales call is a conversation you don’t want to have, and the customer whose name is attached would be surprised to learn they said it.
The habit is small: every number, timeframe, and superlative in a testimonial you didn’t watch the customer type is a claim to confirm before publishing, not after. A one-line reply asking “is twenty hours about right, or should we say something else?” settles it, and customers appreciate being asked.
The tells, and why detectors are not one of them
The reliable signals are about content, not style. A chatbot can’t invent your customer’s specifics, so it writes around them: praise with no starting situation, no objection that had to be overcome, no proper nouns, no number anyone measured. Sentences arrive in threes. Every paragraph is about the same length. The piece describes working with you in general rather than one thing that happened.
Detectors are a different matter, and they’re not worth acting on. They return a probability, they’re wrong often enough in both directions, and the cost of being wrong is asymmetric in the worst way: a false positive means telling a customer who sat down and wrote you something kind that you think they faked it. There’s no version of that conversation that ends well.
The tells above are useful because they point at what’s missing and give you something to ask for. A detector score points at your customer and gives you nothing.
Short, capped, specific: a form no chatbot can help with
The structural fix runs upstream of all of this, and it isn’t persuasion. Nobody opens a chat window to write fifteen words about what nearly stopped them signing up. The blank box is what invites drafting, so stop shipping one.
Replace the single open request with five to seven narrow questions, each with a hard character cap, so the questions do the work of producing specifics instead of leaving a stranger to guess what you wanted. Ask what the situation was before. Ask what nearly stopped them. Ask for one number if there is one, and let them skip it if there isn’t. Each answer is short enough to type from memory and specific enough that a model has nothing to add, because the only source for the answer is the customer’s own experience.
Caps matter more than they look. A field that accepts thirty words gets thirty words of memory; a textarea with no limit gets an essay, and an essay gets delegated. Most testimonial software lets you set that limit per field, and the word budget each display slot can hold is the number to set it to. In Testymo the cap applies as the answer is submitted rather than trimming afterwards, so the stored words stay the customer’s own.
There’s a side effect worth naming: collect this way and what arrives is already the right length and already contains the parts you’d otherwise have to chase.
What to do with the one already in your inbox
Don’t delete it, and don’t send it back. Somebody did you a favour, and the fastest route to a usable testimonial is one narrow question only they can answer.
Thanks for this, it's genuinely kind of you to take the time.
One thing would make it far more useful to someone deciding
whether to work with us: what was the situation before we
started? A number, a deadline, or the thing that was going
wrong is perfect. One line is plenty, and I'll use your
words as they come.Their reply is the testimonial. Publish that sentence with their name and role beside it, and let the fluent paragraph go. A dozen specific words beat two polished sentences that could have been about anyone, and you now have something a reader can check.
If the reply contains a number, confirm it before it goes on a page. If it never comes, publish nothing rather than the chatbot version, and ask the next customer with a better form.
Frequently asked questions
Should I label a testimonial my customer wrote with AI?
There’s no convention that says you must, and a label mostly invites a question the reader wasn’t asking. Spend the effort on the substance instead: get one specific detail from the customer, publish that, and the authorship question stops mattering because the content is now theirs.
Can I just ask them to write it again without AI?
You can, but it usually returns a worse version of the same problem, because the blank box that caused it is still there. Ask one specific question instead. A question only your customer can answer gets an answer only your customer could have written.
What if there’s a number in it I can’t verify?
Ask them to confirm it in writing before it goes anywhere. It’s the single most likely part of a drafted testimonial to be invented, and an unbacked figure on your pricing page is a problem the first time a prospect takes it seriously.
Does using AI to fix a customer’s typos count?
Tidying spelling and grammar isn’t the issue, and it’s a normal courtesy for anyone writing in a second language. The line is the same as for any edit: trimming and correcting are fine, and anything that strengthens a claim, adds a detail, or changes what was said goes back to the customer for approval first.
Is a polished testimonial always worth less than a rough one?
Not always, but roughness is evidence a person wrote it, and it’s cheap to keep. If a customer sends you something clean and specific, publish it as it came. The problem was never polish on its own — it’s polish that arrived instead of the details.