What Maud's Did, and What She Stopped Doing.

An 80-product apothecary in Bristol moved from 1,200 monthly visits and a 1.8 percent conversion rate to citation in three answer engines in 90 days. The interesting part is not what she added.

Maud Carmichael runs an apothecary out of a 19th-century pharmacy on the corner of Park Street in Bristol. She sells eighty products: soaps, balms, a single-rose perfume, a cleanser her grandmother formulated in 1968. The shop has a small e-commerce site that does about a fifth of revenue. Until February, the site got eleven hundred unique visitors a month and converted at 1.8 percent1. By the first week of May, it was getting cited by ChatGPT, Perplexity, and Google AI Overviews, six times in seven days.

Maud did not add a single new product. She did not start a TikTok. She did not run a sale. She did not redesign the homepage, hire an agency, or write a single blog post. What she did is more interesting because it is so much less.

·· IThe shelf moved.

For thirty years, the shelf in commerce has been Google. A buyer types a query, "best rosemary cleanser for sensitive skin", and a search engine returns ten blue links. The link with the best product on the right page wins the sale. Reviews mattered, but they mattered the way a sticker on a packet of biscuits matters: as social proof, sitting on top of the inventory, indexed for SEO if you were paying attention, ignored if you were not.

In 2025, the shelf moved. A meaningful share of buying decisions began to start in an answer engine: ChatGPT, Perplexity, Claude, Google's own AI Overviews. By March 2026, by BrightEdge's count, somewhere between eighteen and twenty-four percent of high-intent commerce queries no longer resolved on a Google results page2. They resolved on a paragraph in a chat window that read like a knowledgeable friend.

An answer engine does not browse a shelf. It reads. It looks at first-party customer language: reviews, product descriptions, the words real buyers use to describe a real outcome, and synthesises that language into a paragraph. The merchant who shows up in the paragraph is the merchant whose customer wrote the most useful sentence.

That sentence is not "five stars, love it" or "great product, fast shipping". It is specific: "doesn't strip everything", "leaves my face tight", "smells like rosemary, not perfume". It names the problem, names the difference, names the sensory cue. It is, in linguistic terms, a sentence that answers a question. The question, in this case, is "what's a sensitive-skin cleanser that doesn't strip the skin".

The brand that gets cited is not the loudest one. It is the one whose customers, by accident, wrote the best answer to a question the buyer was going to ask. ·· The Refusal · No. I · 14 April 2026

·· IIWhat Maud's stopped doing.

In February, Maud was running the standard playbook: a tiered discount on first orders, a Klaviyo flow that sent five emails before the buyer had received the product, a "leave us a review for 10% off your next order" automation, and a homepage hero set in a tasteful but anonymous sans-serif from a free-fonts dropdown. The whole site was, by her own description, "fine".

One. She removed the discount-for-review prompt. The reviews it generated were short, generic, and useless to anyone trying to decide between two cleansers. They were also, by a 2024 FTC opinion she was unaware of, technically not compliant3. She replaced it with nothing: no automation, no prompt, no reward. Reviews dropped from forty-six a month to twelve. The twelve were unrecognisably better.

Two. She stopped writing product descriptions in the brand voice. The brand voice, before, was a marketing voice: "luxurious", "transformative", "indulgent". She replaced it with the language of an apothecary's index card from her grandmother's drawer: "rosemary, lavender, oat. For skin that reacts to most things. Stops short of the skin barrier". The descriptions went from 110 words to 38. Conversion rose to 2.4 percent within three weeks.

Three. She stopped adding new products. She had been launching one a quarter; she paused for six months. The catalogue became smaller and the most-used review words became more concentrated: rosemary, sensitive, no perfume, gentle, expensive but worth it. An answer engine prefers a smaller, denser body of evidence to a larger and more diluted one. Maud, by accident, was tightening the language.

Four. She stopped trying to rank in Google. Not deliberately; she just stopped paying attention. Her SEO traffic was flat. What rose, quietly, was the rate at which her customers' sentences were being paraphrased back to other customers, in a chat window, by a model that had read those sentences six months earlier.

·· IIIThe Maud's numbers, audited.

Three months of data, taken between 14 February and 12 May 2026. Sources: the Shopify analytics export, the BrightEdge GEO panel4, and three answer-engine response logs sampled hourly on Maud's product taxonomy.

The single most important number is the second one. The number of reviews fell by three-quarters. The remaining quarter was the entire engine. The reviews that remained were verified, specific, named, and full of the language a buyer would use to ask a question. Those are the only reviews an answer engine has any reason to quote5.

·· IVThe implication, generalised.

Maud's is not a marketing story. It is a categorisation story. The category of "review platform", defined for a decade by Yotpo, Trustpilot, Stamped, Loox, was built for a shelf that no longer exists. It optimised for volume, automation, AI-generated replies, and decorative star widgets. Those optimisations are now actively harmful. Volume dilutes the language. Automation generates filler reviews answer engines learn to discount. AI-generated replies are statistically detectable. Star widgets are decoration on a shelf nobody is browsing.

The new category, generative-engine optimisation, has different operating principles. Fewer, better reviews. Customer-specific language. A merchant voice indistinguishable from a knowledgeable friend's voice. Schema markup that lets answer engines find the language without making the customer write it twice. And, beneath the surface, an engine that reads the customer's words and routes them to the four queries those words happen to answer.

Maud built none of that on purpose. She just stopped doing the things her category told her to keep doing. The shelf moved, and her customers' language was, by good luck, already the kind of language the new shelf reads.

For the rest of us, the unlearning has to be deliberate.