A nine-person home fragrance brand on Shopify turned on per-client learning in our engine on June 1. Thirty days later: organic sessions up 18 percent, AI-engine citations up from 2 a month to 19, the owner's content time down from six hours a week to 50 minutes, and an edit rate that fell from 70 percent to 12. This case note is the day-by-day of how.
Names and identifying details are changed per our case policy. The numbers are as logged by the engine and the store's own analytics.
The short version: the engine spent week one listening, weeks two and three shipping and being graded, and week four mostly running clean.
The client and the starting line
The client sells candles and room sprays direct-to-consumer: 214 products, about $1.8 million in annual revenue, and a founder who wrote every product description herself for six years and could not anymore. Baseline on May 31: 4,200 organic sessions a month, two AI citations a month, product pages averaging 84 words, a blog dormant for seven months, and a support inbox answering the same 30 questions on rotation.
Her stated goal was narrow: "I want the store to sound like me without needing me." That sentence became the acceptance test for everything that follows.
She had tried generic AI tools twice and quit both inside two weeks. The output read, in her words, like a candle store run by a realtor.
We took the engagement partly because the archive was rich. Six years of founder-written copy is a fingerprinting gift.
What per-client learning means here
Per-client learning means the engine keeps a store-specific memory that every draft consults and every human correction updates: the brand's voice fingerprint, catalog facts like burn times and wax blends, the customer-question inventory, and a growing rulebook of what her edits keep changing. Nothing in that memory is shared with any other client. The brain is shared machinery. The memory is hers.
This is the difference from the tools she quit. They started from zero every session. The engine starts from everything it has ever been corrected on.
Day 30's drafts are better than day 3's for reasons that have nothing to do with the model, which did not change all month.
The memory is also portable. If she leaves, the voice spec, fact sheet, and rulebook export as plain documents. Lock-in is not the retention strategy.
Week one: the engine mostly listened
Days one through six produced almost no visible content, deliberately. The engine ingested 214 product pages, 3,100 support emails, 41 old blog drafts and newsletters, and 1,900 product reviews, then built three assets: a voice fingerprint, a catalog fact sheet with 63 verified product attributes, and a ranked inventory of 118 real customer questions. The owner's only job that week was 40 minutes of correcting the fact sheet.
Eleven of the 63 attributes she corrected were wrong on her own site. The listening week fixed catalog errors before a single new word shipped.
Total inference cost for the listening week was under nine dollars. Listening is cheap. Unwinding confident, wrong copy is not.
Resisting the urge to ship in week one is the discipline. Drafts without the fact sheet would have been fluent, confident, and wrong about wax.
Weeks two and three: shipping and grading
Days seven through twenty-one, the engine shipped 48 pieces: 31 rewritten product descriptions, nine answer-first blog posts built from the question inventory, six FAQ blocks, and two collection page rewrites. The owner graded the first 15 drafts at about 90 seconds each — 22 minutes total — and her corrections became standing rules: never say luxurious, always state burn time in hours, lead with the scent memory rather than the ingredient list.
Edit rate tells the learning story cleanly. Drafts one through ten came back 70 percent edited. Drafts twenty through thirty, 31 percent. The final ten of the month, 12 percent — most shipped untouched.
By day 18 she had stopped reading every draft and started spot-checking one in three. That transfer of trust was earned in the grade log, visible draft by draft.
Grading never felt like a job, per her review notes, because it was 90 seconds with coffee, not a rewrite.
What the learning loop changed by day 30
The store memory ended the month holding 23 voice rules, 63 verified catalog facts, and 118 customer questions with 41 of them now answered on-page. The practical effect: a new product description that took the owner 45 minutes in May now takes the engine four minutes to draft and her 90 seconds to approve, in a voice repeat customers have twice complimented since — unprompted.
Three rules did the most work: the luxurious ban, burn time in hours inside the first 40 words, and scent-memory openings. All three came from her edits, not from our prompts.
One rule was rolled back. The engine over-applied the scent-memory opening to utility pages, where people just want the shipping answer. Learning includes learning where to stop.
Nothing in the loop required her to learn tooling. She graded drafts in plain English, and the engine did the translating into rules.
The numbers after 30 days
June against May: organic sessions 4,200 to 4,950, up 18 percent. AI-engine citations 2 to 19 a month, mostly Perplexity and Google AI Mode quoting FAQ answers and burn-time facts. The 31 rewritten product pages converted at 3.1 percent against the store's 2.7 percent baseline. Owner content time: six hours a week down to 50 minutes. Edit rate: 70 percent down to 12.
Attribution honesty: June included no paid push and one modest email send, but also seasonal softness — June is historically her slowest month. An 18 percent lift in a down month is the part she quotes.
The citation jump is the least surprising number to us. Forty-one newly answered questions with clean FAQ structure is exactly the surface engines lift from.
Revenue attribution stays humble at 30 days. She tracks it. We report the leading indicators that move first.
What did not work, and the next 60 days
Two things flopped. A three-post gift guide series underperformed everything else the engine wrote — 60 total sessions — because June is not gifting season, and the queue logic now checks seasonality before topic selection. And an early FAQPage schema block mismatched its visible text after a manual edit, which the day-19 audit caught. The engine now re-validates schema after any human edit, automatically.
The next 60 days are scale, not novelty: the remaining 183 product descriptions at roughly 40 a week, a two-week content buffer for the blog, and the first quarterly re-fingerprint to catch voice drift.
The owner's total June involvement: about three and a half hours inside the system, most of it grading.
Her sentence at the 30-day review is the one we keep: "It finally sounds like me on the days I am not there." That is per-client learning doing its one job.