Skip to main content

Brand Audit

In plain English

When AI writes your product descriptions, two things can go wrong.

The first is tone. It uses words you'd never use. Maybe you've decided never to say "revolutionary" or "game-changing," or there's a competitor name you don't want appearing. You can give the app a list of banned words, and this screen tells you every time one slipped through.

The second is more serious: AI can make things up. Asked to write about a jacket, it might helpfully add "with a 2-year warranty and free next-day delivery." You never said that. It sounds plausible, so it's easy to miss in a review — and if you publish it, you may well have to honour it.

Seven kinds of fact get checked specifically, because these are the ones that become promises:

Invented factWhy it's dangerous
Price or discountYou may have to sell at it
Warranty or guaranteeYou may have to honour it
Delivery or shipping timeA promise you might not meet
Returns or refund policyLegally binding
Stock or availability"Only 3 left" when there are 300
WeightWrong shipping costs, wrong expectations
DimensionsReturns, and complaints

The app's writing tools are already told not to invent these. This screen is the safety net that catches anything that gets through.

How to use it

  1. Click "Brand Audit" in the left-hand menu.
  2. Look at the violations count for the period. Zero is what you want.
  3. If there are any, click through to see the exact phrase, where it appeared, and which tool generated it.
  4. Fix or remove that copy before it goes live.
  5. Add banned phrases in your brand voice settings — competitor names, words you never use, claims you can't support.
  6. Check the trend — if violations are rising, your brand rules and your writing tools have drifted apart.

⏱ ~10 min/week · 💳 Pro+ · 🎯 Never publish a promise you didn't make

Why this matters for your business

There's a specific risk that comes with generating copy at scale, and it isn't about quality. It's about liability.

If your product page says "2-year warranty," that's a term of sale. A customer who bought on the strength of it is entitled to it, whether or not you meant to offer it, and whether or not a machine wrote the sentence. "Our AI generated that" is not a defence anyone wants to test.

The same applies to delivery times, returns windows, and prices. Each is a commitment. Each is exactly the kind of plausible detail a language model adds to make a description feel complete.

When you're writing ten descriptions by hand, you'd catch it. When you're generating four hundred, you won't. That's the gap this closes.

The tone side matters too, just less urgently. Consistent vocabulary is most of what makes a brand feel like a brand, and it's the first thing to erode when generation scales up. A banned phrase list is a blunt instrument, but it's an effective one.

One deliberate design choice: the fact-checking warns rather than blocks. It flags for a human to confirm or remove, instead of refusing to publish. Sometimes the AI is right — if you do offer a 2-year warranty and it's in your source data, the sentence is fine. Blocking would create busywork; warning puts a person in the loop where judgment is needed.

What this typically unlocks

What you getTypical result
Invented promises reaching customersCaught before publish
Brand-voice consistency at scaleMeasured, not assumed
Legal exposure from generated copySubstantially reduced
Time spent proof-reading AI outputLower — you review flags, not everything
Visibility of brand driftA trend line, not a feeling

What you actually get

Violations dashboard

NumberWhat it tells you
Violations in periodHow many generations broke a brand rule
Distinct AI sourcesWhich tools are producing them
Distinct banned phrases hitHow many of your rules are being broken

Filter by last 7 days, 30 days, 90 days, or all time, and by which tool produced it.

The "distinct sources" number is the useful diagnostic. If every violation comes from one tool, that tool needs its configuration checked. If they're spread evenly, your banned list may be too aggressive.

The seven merchant-owned facts

These are checked separately from banned phrases because the consequences are different.

How the check works: it looks for a policy word combined with a number or unit — "2-year warranty," "ships in 3 days," "30-day returns." It's tuned to be precise rather than exhaustive, so it rarely cries wolf.

It warns, it doesn't block. If the fact is true and you supplied it, keep it. If the AI invented it, remove it. The decision is yours; the flag makes sure it's a decision rather than an oversight.

On-brand rate

A simple measure with a useful trend:

on-brand rate = 1 − (generations that broke a rule ÷ all generations)

A daily snapshot builds a drift line over time.

PatternWhat it means
Steady and highYour rules and your tools agree
FallingDrift — usually a new tool, new prompt, or a rule added without updating settings
Sudden dropSomething changed. Check what shipped that day

Brand Authority score

A separate 0–100 measure answering a different question: how strong does your brand look to AI shopping assistants?

It fuses five signals:

SignalWeightWhat it measures
AI visibility35How often you appear in AI answers
Review strength20Your average rating (needs at least 5 reviews)
Citations15How often AI answers cite you — 10+ scores full marks
Answer rank15Your position when listed. Each place below first costs points
Semantic readiness15How machine-readable your content is
BandScore
Strong70–100
Building40–69
Weak0–39

A signal you don't have doesn't count against you. A shop with no reviews is scored on the other four, re-weighted — it isn't penalised for an absence. The screen shows how many signals are present so you know how complete the picture is.

This score also appears on your Daily Ops Executive Pulse.

How it works (without the technical bits)

Real merchant scenarios

Scenario A — The warranty nobody offered

Setup. Electronics retailer generated 340 product descriptions in a bulk run.

Flagged before publish:

Warranty / guarantee claim detected — 23 descriptions

"...backed by a comprehensive 2-year manufacturer warranty."
"...comes with a 12-month guarantee for total peace of mind."

They offer 90 days.

What would have happened. 23 product pages promising up to eight times their actual warranty. Any customer buying on that basis would have had a strong claim.

Their estimate of exposure, at ~40 units per product per year and an average repair cost of £85: potentially £78,000 across those 23 lines.

Time to fix: about 20 minutes, because the flags pointed straight at the sentences.

Scenario B — Delivery promises they couldn't meet

Setup. Homeware brand, AI-generated collection pages.

Delivery / shipping time claim — 41 pages

"...with free next-day delivery on all orders."

Their actual delivery: 3–5 working days, and next-day wasn't offered at all.

Why it happened. The model had been given competitor pages as style reference. It absorbed a competitor's delivery promise as a category norm.

Result of catching it. No customer complaints, no chargebacks, no advertising-standards exposure. The pages went live with accurate information.

Their comment afterwards was that they'd have spotted it in one page but not in 41 — and they'd only have read a handful before publishing.

Scenario C — Drift after adding a new tool

Setup. Merchant's on-brand rate sat around 98% for months, then fell to 71% over three weeks.

The dashboard showed all new violations coming from one source: a newly-enabled ad-copy tool.

Cause. Their banned list included "cheap," "discount," and "sale" — brand rules for a premium positioning. The ad tool was producing promotional copy that naturally used all three.

Two legitimate options. Either the ad copy was off-brand, or the banned list was too strict for advertising.

What they decided. The rules were right for product pages and too strict for ads. They kept the rules and rewrote the ad prompts to work within them.

On-brand rate returned to 96%.

The value was the diagnosis — one tool, one cause, visible in a trend line.

Scenario D — A flag that was correct to keep

Setup. Merchant saw a warranty claim flagged and nearly deleted it automatically.

Warranty claim — "...covered by our 5-year frame guarantee."

They checked. It was true. The 5-year frame guarantee was real, in their product data, and a genuine selling point.

They kept it.

Why the warning design is right. A blocking system would have stripped a true and valuable claim. The warning put a person in the loop, they confirmed it in ten seconds, and the good copy survived.

Scenario E — Brand Authority with missing signals

Setup. New brand, six months old, saw:

Brand Authority 62 · Building
Signals present 3 of 5

AI visibility 68
Semantic readiness 71
Review strength — (only 3 reviews)
Citations — (not yet scanned)
Answer rank 41

They assumed the missing signals were dragging the score down.

They weren't. Absent signals are excluded and the rest re-weighted. The 62 was based purely on the three they had.

What it actually told them. Answer rank at 41 was their weakest present signal — they appeared in AI answers but low down. That was the thing to work on.

Action. Improved structured data and content depth on their key categories. Answer rank rose to 68 over four months, and overall authority to 74 — "Strong."

Best practices

Build your banned-phrase list early. Competitor names, words you never use, claims you can't support.

Review flags before every bulk publish. That's where the risk concentrates.

Treat the seven fact types as urgent. They're commitments, not style.

Watch the on-brand trend, not just today's count. Drift is gradual and easy to miss.

Check which tool is producing violations. One source means a configuration problem, not a brand problem.

Keep true claims when flagged. The warning asks a question; sometimes the answer is "that's correct."

Don't publish AI copy about warranties, delivery, or returns without reading it. These are the expensive ones.

Don't ban so many words that everything trips. Rules people ignore protect nothing.

Don't assume a low authority score means missing signals are the problem. Look at your weakest present signal.

Don't delete flagged claims automatically. Scenario D.

Plan tiers

CapabilityFreeStarterProAgencyEnterprise
Banned-phrase scanning
Merchant-owned fact detection
Violations dashboard
On-brand rate + trend
Brand Authority score
Filter by source and period
Multi-store brand governance

Frequently asked

Does this stop bad copy being published? Banned phrases are logged as violations. Fact claims produce a warning for a human to judge. Neither blocks automatically — see Scenario D for why.

Where do I set banned phrases? In your brand voice settings. See Brand voice & personas.

Why do fact claims warn instead of blocking? Because sometimes the claim is true. Blocking would strip accurate copy and create work; warning puts judgment where it belongs.

Does it catch every invented fact? It's tuned for precision — a policy word plus a number or unit — so it rarely produces false alarms. It's a strong safety net, not a guarantee. Read anything making a commitment.

Why is my Brand Authority score low with no reviews? It isn't penalised for that. Missing signals are excluded and the others re-weighted. Look at your weakest present signal instead.

How many reviews before review strength counts? At least five. Below that it's too noisy to be meaningful.

What is "semantic readiness"? How well machines can understand your content — structure, clarity, and markup. See Schema markup.

My on-brand rate dropped suddenly. Where do I start? Filter by source. A single tool producing all the new violations points straight at what changed.

See also