Why customers can taste AI-written copy — and the 30-minute fix

If your menu descriptions, product listings and service pages read like everyone else's, our bet is that a 30-minute edit pass per page does more for your bookings than another AI subscription.

The sameness problem, briefly

TechCrunch gave it a name in September 2026, in a piece headlined "The sameness problem behind those unappetizing AI-generated menus". That headline is all we're citing it for — everything after this paragraph is our reading, not a finding from the article.

The problem it points at isn't that the copy is wrong. It's that it's the same: same rhythm, same adjectives, same shape as the other places the customer has open in another tab.

Our read on why: it's baked into how the tools work. A language model produces likely text, and likely text is the phrasing that already appears everywhere. Ask ten restaurants on the same street to generate a dish description with the same tool and you'd expect ten variations on one paragraph, because the tool is drawing from the same well every time.

Customers don't diagnose this. They just skim it. "Crafted with the freshest locally-sourced ingredients" carries no information, so the eye slides off and lands on the next option.

What the flatness actually costs

We're not going to hand you a conversion statistic here, because we don't have one we can source, and an invented one would be worse than none. So treat the rest of this section as our judgment, not a measurement.

A buyer choosing between three plumbers or three lunch spots is looking for something to decide on. Generic copy gives them nothing, so they fall back on whatever's left: price, distance, the photo, the review count. If you're not cheapest and not closest, generic copy has quietly moved you into a comparison you lose. Specifics do the opposite — they hand the buyer a reason that only applies to you.

The 30-minute pass

Do this on one page at a time, before anything goes live. Set a timer; the timeboxes matter more than perfection.

Minutes 0–5: cut the throat-clearing. Delete the first sentence and usually the first paragraph. AI drafts warm up before they say anything. "Whether you're hosting a family gathering or grabbing a quick bite" is warm-up. So is any sentence starting with "At [business name], we pride ourselves on". Start where the information starts.

Minutes 5–10: turn adjectives into facts. Every evaluative word is a slot where a fact should go. "Fresh" → the day you get deliveries. "Experienced" → the year you started and the number of vans. "Fast" → the hour you answer the phone until. If you can't replace it with a fact, delete it; an adjective nobody can check is doing zero work.

Minutes 10–15: add proper nouns. Supplier names, street names, the cross-street, the neighborhood, equipment models, the name of the person who does the thing. Proper nouns are the cheapest anti-sameness tool we know of, and they're the one thing the model cannot guess about your business.

Minutes 15–20: add one honest constraint. What you don't do, what's a trade-off, what's genuinely a wait. "No parking lot — street parking on the west side." "We don't do septic." "Fridays after 6, expect 25 minutes." This reads as human because generated drafts rarely volunteer a downside, and our expectation is that it pre-empts the question that otherwise clogs your phone.

Minutes 20–25: run the competitor test. Read each sentence and ask: could the shop down the street publish this word-for-word without lying? If yes, cut it or make it specific. That one test catches a lot of what's wrong with a generated page.

Minutes 25–30: check the promises. Prices, hours, allergens, licensing, warranty length, service area. A model can write a confident sentence about your gluten-free kitchen whether or not you have one. Verify each against a source of truth — your POS, your license, your calendar — before it ships.

A worked example

Both of these are invented illustrations we wrote for this post, not a real client's copy.

Before (generated): "Our signature pasta is crafted with the freshest locally-sourced ingredients, delivering an unforgettable dining experience for the whole family."

After (30-minute pass): "Cacio e pepe. We roll the tonnarelli Tuesday and Friday mornings, and we cut the pecorino off a whole wheel at the counter. One size, one price, and it's a salty dish — that's the point of it. If you want it milder, ask and we'll go light on the cheese."

Same length. The second one tells a stranger three things they can act on and one thing they'd otherwise have complained about.

Before (generated service page): "At [Company], we pride ourselves on delivering exceptional HVAC solutions with unmatched customer service. Our team of dedicated professionals is committed to your comfort."

After: "Two vans, both on the road by 7am. We answer the phone ourselves until 8pm, including Saturdays. Diagnostic visit is a flat fee that comes off the repair if you go ahead. We service the east side and the county line; we don't do commercial rooftop units, and we'll tell you who does."

On the arithmetic — hypothetical, using your own page count: if you have 24 menu items and six service pages, that's 30 pages at 30 minutes each, or 15 hours. Three hours a day for a week, once. Then the marginal cost of every future page drops, for the reason in the next section.

Build the specifics file once

The reason the pass gets faster is that most of the work is recall, and you only have to do the recall once.

Open a doc and write 15–20 lines of raw facts about the business: suppliers and where they are, equipment and models, years in operation, staff first names and what each is good at, the landmark across the street, delivery days, what you refuse to do, your actual prices, your honest trade-offs, the two questions customers always ask on the phone.

That file is now the input to everything. Paste it at the top of the prompt before you ask for a draft. Give it to whoever writes your listings. Feed it to your chatbot so the bot stops answering like a brochure. Use it for review replies. We haven't measured the time saved and won't pretend otherwise, but the drafts come back closer to publishable, because the model is working from your facts instead of the average of everyone else's.

What we'd build for this

If a client handed us this problem, we wouldn't build software for it. We'd build a house-style pack: the specifics file, a prompt template that loads it every time, a one-page kill list of phrases that get cut on sight, and a saved assistant so staff can't skip the setup step. Our estimate — an estimate, not a quote, and it varies with how many pages you have — is roughly a half-day of work for a single-location business, and the person who owns it afterwards is whoever already writes your posts.

The thing to avoid is the middle path: generating copy, skimming it, and shipping it because it "reads fine". It does read fine. That's the problem. Fine is what everyone else has.

Talk it through

If you want a second pair of eyes on the pages that actually bring in bookings — and a house-style pack your staff will keep using after the novelty wears off — book a strategy call. Bring one page you're not happy with; we'll do the 30-minute pass on it live.