Hi, it's Peggy Burnett.
I mined twenty-three real customer sentences out of public reviews, pasted them into Claude, and asked for a sales lead. The model used none of them. It did not borrow five consecutive words from any of the twenty-three sentences real buyers had written.
Then it told me that almost nobody fails at one habit over eight weeks, and that people fail at five habits in eight days. Neither of those figures exists. Nobody measured them. The model produced them to fill a proof gap I had deliberately left open.
I ran that arm a second time with a fresh session to check whether I had drawn a bad sample. It borrowed nothing again, and it produced a second invented statistic in almost the same rhetorical shape.
I went in expecting the verbatim quotes to produce the most grounded copy in the test. The measurement put the finding somewhere else, and it is more useful than the one I was looking for.
The Sorting Is the Part You Are Paid For
Here is what I set up, so you can run it yourself and disagree with me.
I wrote one offer and froze it. An eight-week coached habit program, with the format, the price, the guarantee and the mechanism all spelled out, and every proof field left deliberately empty and marked "to confirm."
Leaving the proof blank on purpose is what makes the fabrication countable rather than a matter of impression. Anything numeric that showed up in a lead had to have been invented, because there was nothing to quote.
Then I ran the same brief five times under four input conditions, each in a separate session that could not see the others and did not know an experiment was running.
Arm | What it received alongside the offer |
|---|---|
A | Nothing. The offer block on its own. |
B | A customer persona, written the way briefs usually write one, distilled from the same twenty-three quotes. |
C | All twenty-three quotes, raw, unsorted, no instruction about them. |
C2 | A repeat of C, fresh session, to test whether the result held. |
D | The same twenty-three quotes plus an explicit instruction to mirror them and invent nothing. |
The quotes came from public reader reviews on two books in the habit-change category. Category mining rather than single-product mining, which is what you do when the offer is new and has no reviews of its own.
One note on the mining itself, because it nearly invalidated the whole test. My first pass used a tool that fetches a page and hands back a description of it. What came back read like quotation and was paraphrase.
An experiment about verbatim language cannot run on quotes something has already rewritten, so I threw that pass away and re-read every page as rendered text. If you take one operational thing from this issue before the findings, take that one.
Zero Borrowings, Twice, From Twenty-Three Real Sentences
I counted a borrowing as five or more consecutive words appearing in both a source quote and the finished lead. That convention is arbitrary, which is why I am stating it, and you can recount at a different threshold if you think five is wrong.
A, cold | B, persona | C, verbatim | C2, repeat | D, guided | |
|---|---|---|---|---|---|
Borrowings from the quotes | no quotes given | no quotes given | 0 | 0 | 8 |
Invented proof-shaped claims | 0 | 0 | 2 | 1 | 0 |
The market's loudest theme survives | no quotes given | yes | no | no | no |
Usable as written | yes | yes | no | no | no |
The two unguided runs behaved identically, down to the shape of the sentence they invented.
What it wrote | |
|---|---|
Arm C | "Almost nobody fails at one habit over eight weeks. People fail at five habits in eight days." |
Arm C2 | "Nobody fails at building one habit. People fail at building four habits disguised as one commitment." |
Neither of those comparisons exists. Two sessions ran with no contact between them and reached for the same fabricated contrast, in the same position on the page.
Handing a model raw quotes with no instruction does not produce grounded copy, and it appears to license the opposite. The two arms holding less material invented nothing at all. The arms holding the most raw material were the ones that reached for a statistic.
I have a guess about the mechanism and I want to mark it clearly as a guess. Twenty-three unlabelled quotes look like evidence, and a page that opens on evidence seems to want a figure in it. The model had no figures, so it supplied some.
The Summary Carried the Market Better Than the Quotes Did
Nine of the twenty-three quotes are about being burned by productivity books already. It is the largest cluster in the file by a wide margin, and it is the single most useful thing in there for a copywriter, because it names the disliked alternative the reader is choosing against.
Here is what each arm did with it.
Arm B held a one-sentence summary that named the trait, and opened on it in its first five words.
Arms C, C2 and D each held all nine of the underlying quotes, and none of them mentioned books, prior advice or the category at all.
The arm given the conclusion used it. The three arms given the evidence for that conclusion did not.
That result is the one I keep turning over. A summary is lossy by construction, and it still transmitted the signal, because somebody had already decided the signal mattered and said so in a sentence. Twenty-three unsorted quotes contain the same information and hand the model no indication of which of them is load-bearing.
The Mirror Rule Stops the Invention and Costs You the Copy
Arm D is where the technique finally worked as advertised. I wrote the instruction fresh: build the lead from the customers' own words, change first person to second person and split long sentences, change nothing else, invent no statistic or frequency or outcome, and make every sentence about the reader trace to a specific quote.
It produced eight borrowings and no fabrications, on the same model, the same quotes and the same offer as the arm that produced neither.
The instruction was the variable the whole time, and the quotes never were.
Arm D is also the only lead in the test I would not send, and it fails in three separate places.
The grammar breaks where the rule forbade tidying. The copy reads "you have habits you want to change" and then "stuffing up your lives" in the same sentence. A person swap is not always grammatically free, and the rule allowed no repair.
Two unrelated quotes get welded into one sentence that argues with itself. Skipping a new habit on a bad day and abandoning an established one are different problems, and the fused sentence asserts they are the same problem.
The register comes through untouched. It imports a reviewer describing her own brain as "a bit bitchy" into the second paragraph of a sales page for a mainstream coaching program, because the rule permitted no change of register.
None of those three failures is a reason to abandon mirroring. All three are the same missing step, which is a human deciding which quotes go in before anything is instructed to mirror them.
What I Would Do on a Real Project Tomorrow
The sequence the results support has three steps, and the middle one is not automatable.
Mine verbatim, and verify it is verbatim. Read rendered pages rather than trusting a tool that summarises. Keep the source next to every quote.
Sort and cut by hand. Group the quotes, decide which cluster carries the disliked alternative, and throw out everything that is merely vivid. This is the step no arm in my test performed and the step that decided every outcome in it.
Then instruct. Paste the surviving handful with the constraint attached.
Here is the instruction from Arm D, which is the reusable part. Give it eight or ten quotes you have already chosen rather than everything you found.
Below is an offer and a small set of quotes taken word for word from real customers in this market.
Build the lead out of the customers' own words. Wherever the copy describes the reader's situation, use the wording from the quotes rather than your own phrasing.
The only changes you may make to a borrowed phrase are these: change first person to second person, split one long sentence into two, and adjust grammar where the person swap breaks it. Do not improve, tidy, modernise or elevate the wording beyond that.
Invent nothing. No statistic, no frequency, no percentage, no number of people, no success rate, no time-to-result, and no claim about what usually happens to anyone. If it is not in the offer or in a quote, it does not go on the page.
Every sentence describing the reader must trace to a specific quote. Tell me which quote each one came from.
Where a quote would strengthen the page but the offer has no proof to support it, say so and leave the gap visible rather than writing around it.
Flag any quote whose register is wrong for this audience instead of importing it.Two clauses in there came directly out of the failures above. The grammar allowance exists because the strict version broke a sentence. The register flag exists because the strict version shipped a word no client would sign off on.
Where This Stops Being About Your Market
Four things would stop me presenting this to a client as settled.
The sample is thin. One offer, one category, five leads. The unguided arm is the only one I ran twice, so the other three conditions each rest on a single draw.
The offer is fictional, and that has a cost. Every proof field was blank by design, so the invention count is measured against a guaranteed gap rather than against available truth. A real offer carrying real numbers might not tempt the model the same way, and I have not tested that.
The writer was blinded and the analyst was not. I designed the conditions, wrote the persona, set the counting convention and did the counting, with no second reader. Recount them if this matters to your work.
Nothing here touches conversion. Five leads sat on a page and none of them went to a list. What each input put on the page is a different question from which page sells.
The Burnett Matrix
The mining is cheap and the instruction is cheap, so neither of them is what a client is paying for. The twenty minutes where somebody reads twenty-three quotes and decides which four matter is the part that changed every result in this test, and it is the part no rule and no model performed on its own.
More opens, clicks, and conversions,
— Peggy Burnett
