By Smári Ásmundsson · · 4 min read
You may be paying to rewrite words against a doubt your buyer never had
Hand your product page to AI and you get a better-written page aimed at the same wrong objection. The rewrite worth having starts in your reviews and your support inbox, where the real objection is already written down.
A retailer sends a deduction, or a broker says your sell sheet misses the point. For a direct-to-consumer brand the same week is a product page that stopped selling. The reflex is the same in both. You hand the page or the sell sheet to an AI and ask for a better one. You get back something cleaner and tighter. It is aimed at the same objection the old version was aimed at, and that objection was a guess.
The usual guesses are free shipping, photos of how the product is made, more reviews. Each is reasonable. None starts from what your own buyers and non-buyers have already said.
That is where I would start. You already hold a written record of what people doubted before they bought and what surprised them after. For a direct-to-consumer brand it is reviews and the support inbox. For a brand that sells through stores and distributors it is broker notes, retailer feedback, chargeback or deduction reasons, and distributor emails. The method is identical. The doubt is already there, in their words, in volume.
Running work through AI as a second check on what you already believe is a sound use. Here it is not asked to invent a pitch. It reads what customers actually said and tells you which doubt shows up most.
A usable definition: a real objection is a doubt that appears, unprompted, in many separate messages, in different words. A doubt you imagined in a meeting is a guess. A doubt that shows up across dozens of messages is data.
In the older days you would download every review and email into a spreadsheet, read each one, tag it with a theme, and count the tags. A day and a half later you have a ranked list.
Today, give this prompt to Claude and it reads everything and has your answer in thirty seconds. That is the reading only. Getting the files together takes you, or whoever handles your software, about twenty minutes the first time (illustrative): download the reviews, save the emails as files, and upload them to the chat. Many review apps have a download button, and an email can be saved as a file. Claude is an AI assistant you type to, like a chat window on your computer. You open it at claude.ai in a browser. The prompt is the instructions you paste in.
One privacy step before you upload anything. Strip customer names and addresses from the files, or turn on the setting that stops the company from using your files to teach its AI (look for it under privacy or data controls in your account).
I run a food brand. I have uploaded my customer messages: reviews, support emails, and any broker notes, retailer feedback, deduction reasons or distributor emails I had. Read all of them.
First, tell me how many messages you actually read.
Second, find every doubt, worry, complaint or surprise a customer or buyer expressed about the product. Group similar ones together, even when the wording differs.
Third, rank the groups by how many separate messages mention them. Show the top five. Under each, quote three real messages word for word, so I can see the ranking is real.
Fourth, tell me which group looks like the reason people hesitate before buying, as opposed to something they discovered after buying. Say how sure you are and why.
Fifth, take my current product page or sell sheet, which I will paste below, and rewrite it to answer the top pre-purchase doubt in the first screen a reader sees. Keep my facts as they are. Do not add claims I have not given you. Mark anywhere you needed information I did not supply.
My page or sell sheet:
[paste your page text here]
The only blank is your page text, and you already have it.
The First and Third steps of that prompt are there because an AI assistant can miscount a large pile, and its totals can drift. Asking for the number of messages it read, and for three real quotes under each group, is what makes the ranking trustworthy.
Here is an illustrative example, a made-up brand. Say Claude reports that 62 of 410 messages for a pasta sauce company raise the same worry: the jar looks small for the price. Nobody wrote about shipping. Nobody asked for production photos. The old page opens with the founder story. The rewrite opens with servings per jar and cost per serving. Free shipping would have been a guess. This is a reading of what people said.
The stakes are the next round of copy work. Rewrite against the wrong objection and you may pay for new words and see the same result. Rewrite against the top one and you at least know the page answers something customers raised.
One limit. Claude reads what people wrote. Buyers who left without saying anything are not in your files. So treat the top objection as the first thing to test, and let your sales numbers tell you whether it was the right one.
CPG AI Assessment
Fix the CPG workflow that keeps landing back on your desk.
I map the recurring reporting or operating workflow first, then build only when the first useful fix is clear.