Brand Voice at Scale: Keeping AI Copy On-Brand Across 500 Products

Nobody complains about the grammar. When a merchant says AI-written catalog copy isn't working, the sentences are almost never broken. The complaint is some version of "it doesn't sound like us." And when you open the export and read forty descriptions in a row, the problem is usually not that they sound wrong. It's that they sound like each other.

Elevate your morning routine. Discover the perfect blend of style and function. Whether you're heading to the office or out for the weekend. Four openers, four hundred products.

That's the real failure mode. Not errors. Sameness.

Why it happens

A language model with no instructions writes the average of everything it has read about your product category. That average is not neutral, it has a personality, and the personality is a mildly enthusiastic stranger who has never touched the product. Ask ten different stores in ten different countries to describe a linen shirt with no further context and you get ten variations of the same paragraph, because you asked the same question.

The fix isn't a better prompt on the day. It's a set of constraints that gets applied every time, including on the Tuesday six months from now when someone bulk-generates 80 new SKUs and doesn't think about voice at all.

Sameness is also a business problem, not just an aesthetic one

Two reasons to care beyond taste.

Google's spam policies define scaled content abuse as generating many pages "for the primary purpose of manipulating search rankings and not helping users," with "little to no value to users, no matter how it's created." That last clause matters. The policy is not about whether a machine wrote it. It's about volume without value, and generic copy across 500 products is precisely that shape. Copy that says something true and specific about the product doesn't have this problem, no matter what wrote it.

The second reason is newer. Shopify Catalog now passes product titles, descriptions, images, pricing and inventory to AI surfaces automatically, and Shopify's own explainer confirms merchants are included by default, with products syndicated to ChatGPT, Microsoft Copilot, Google AI Mode and the Gemini app. Your description is no longer just the text under the buy button. It's the raw material an assistant paraphrases when a shopper asks it to compare three options. Interchangeable copy gets you an interchangeable summary, and then the shortlist gets decided on price.

What a voice profile actually needs

Four inputs do the work: tone, audience, banned words, and a catch-all for house rules. Each is only as useful as it is specific.

Tone first, and this is where most people write something worthless. "Professional and friendly" describes roughly all commercial writing ever produced. Compare it to "flat and factual, short sentences, no exclamation marks, spec before benefit." One of those is a rule a writer can follow and an editor can check. The other is a mood.

Audience has the same problem. "Our customers" tells you nothing. "Home baristas who already own a grinder and have opinions about burr size" changes what the copy leads with, which is the entire point of naming an audience in the first place. So does "parents buying a first pair of hiking boots for a child who will outgrow them in a year."

Then banned words, which does more work than the other three combined and is the field most people leave empty. Start with the obvious offenders: elevate, unleash, indulge, game-changing, must-have, look no further, perfect for, seamlessly. Then add yours. Words legal has ruled out. Competitor names. Claims you can't substantiate. Terms your customers don't actually use, like calling it a "carryall" when everyone on the internet calls it a tote.

The reason banned words beat tone settings is boring and practical. A negative constraint is checkable. You can search an export for "elevate" and get a number back. You cannot search an export for "on-brand."

Everything else goes in guidance. Always name the material in the first sentence. Never say "sustainable" without naming the certification. Metric first, imperial in brackets. British spelling. A decent test for whether a rule belongs here: would a new copywriter need to be told this on day one?

The fields voice should not touch

Two output types should be exempt from styling entirely: keywords and image alt text.

Keywords are matching tokens. They exist to line up with what someone typed into a box. Running them through a tone filter is like applying your brand font to a barcode, and at best it does nothing.

Alt text is more important and more often ruined. It has two real jobs: telling a screen reader user what's in the picture, and telling search engines and agents what's in the picture. "Elevate your mornings with our artisan pour-over kettle" fails both. "Matte black gooseneck kettle with wooden handle, 1 liter, on a wooden counter" succeeds at both. Brand voice makes alt text worse in direct proportion to how strong the voice is. We go further into the mechanics in the image SEO guide, but the short version is that alt text is a factual field wearing a marketing costume, and you should take the costume off.

This is why Seokai applies the brand voice profile to titles, descriptions, tags and social posts, and deliberately excludes keywords and alt text. It isn't an oversight. Factual outputs get worse when you style them.

Drift is what actually kills it

You can set all of this up correctly and still end up back where you started in eight months, because voice drifts. It drifts in specific, predictable ways.

New products get optimized weeks after launch, sometimes by a different person, sometimes with the profile half-filled. A seasonal campaign comes through and someone loosens the tone for the holidays and never tightens it back. Translations get generated and the voice quietly flattens, because tone is the first thing lost in translation and nobody on the team reads the German copy. Then the slow one: the brand repositions, the profile doesn't, and eighteen months of copy is now describing a company that no longer exists.

Multi-language is worth its own attention here. A voice profile written in English does not automatically survive being rendered in eight locales, and the fix is a per-locale review rather than trust. Our multi-language SEO guide covers the setup side of that.

The twenty-minute drift audit

Do this quarterly. It's tedious and it takes less time than one meeting about brand guidelines.

• Export 40 product descriptions at random. Put the first five words of each in a column and sort. If more than a quarter of them start the same way, you have a sameness problem, and no amount of tone tuning fixes it.
• Search the export for every word on your banned list. Any hit means the constraint isn't being applied or the copy predates it. Both are fixable in bulk.
• Read five of them out loud. Voice problems are audible long before they're visible on a screen.
• Compare three bestsellers against three products from the long tail. Drift always shows up in the tail first, because nobody reads the tail.
• Repeat the banned-word search on one non-English locale.
• Diff your newest ten products against your oldest ten. If they read like two different companies, the profile changed and nobody backfilled.

When you find drift, fix the profile first and re-run in bulk. Hand-editing individual descriptions treats the symptom and guarantees you'll be back here next quarter.

Voice at scale isn't a creative achievement. The stores that hold a consistent voice across a thousand SKUs are not more talented than the ones that don't. They wrote the rules down, put them somewhere a machine reads them, and checked occasionally that they were still being followed.

Seokai keeps that profile in Settings and applies it to every title, description, tag and social post it generates, and there's a free plan if you want to hear what your catalog sounds like under one set of rules.