AI in Marketing: What Scales and What Backfires

Marketing is where generation is cheapest to apply and easiest to apply destructively, because the output is public, permanent and attributed to you. The interesting question is not whether the writing is good. It is which failures compound.

Why a thousand pages hurts, structurally

Search engines publish policies about content produced at scale primarily to manipulate rankings, and those policies change; this page does not try to summarise anyone’s current wording. It does not need to, because the damage from a mass-generated page set is a property of the pages and holds regardless.

• Cannibalisation. Pages generated from one template across a keyword list overlap heavily in what they are about. They compete with each other, split whatever authority the topic accumulates, and the site ends up with five mediocre pages ranking nowhere in place of one that would have ranked.
• Near-duplication. Same structure, same transitions, same closing paragraph, different nouns. Whatever a crawler does with that, a reader who lands on two of them learns something about you that is difficult to unlearn.
• Crawl and index dilution. A site’s crawl attention is finite. Thousands of low-value URLs consume it and slow the discovery of the pages that matter.
• The step change is itself a signal. A site that published two pages a week for three years and then publishes nine hundred in a fortnight has told a story about how those pages were made. Publishing to a cadence is not superstition; it is the only way the shape of your publishing history stays consistent with the claim that a person was involved.

The test that survives all of it: does this page answer a question that the current results answer badly, and can the person who published it say what that gap was. If not, the page is a liability whether a model wrote it or a contractor did.

Brand voice as a specification, not an adjective

“Friendly but authoritative, professional yet approachable” in a system prompt produces the house style of every model, which is the thing readers have learned to recognise. A voice specification that actually constrains output has three parts, and all three are checkable:

• Exemplars. Five to ten paragraphs of real, already-approved copy in the target voice, in the prompt. This does more than any amount of description — see few-shot prompting.
• A banned list. The specific constructions your editors always strike: the em-dash-heavy triads, “in today’s fast-paced world”, “it’s not just X, it’s Y”, “delve”, “unlock”, “robust”. A regex over the draft catches these before a human reads it, and it costs nothing to run.
• Measurable constraints. Sentence length distribution, reading grade, second person versus first person plural, whether numbers are spelled out. These are computable, so they can fail a build rather than be argued about in review.

Put together, that is a lint that runs on every draft, and it converts a subjective review into a short list of specific violations. The general pattern — treat prompt outputs as artefacts under test — is in prompt versioning and regression testing.

The sentences that need a human

Advertising claims about a product’s performance, superiority or effects generally have to be substantiable, and the rules differ by market and by product category. A generator has no idea which of its sentences is a claim. It will write “the fastest way to reconcile your accounts” and “clinically proven to reduce redness” with the same fluency it writes a greeting, because both are common patterns in marketing copy.

The practical control is a claim detector on the output — superlatives, comparatives naming a competitor, numbers with a unit, the words proven, guaranteed, certified, safe, and any health, financial or environmental assertion — which routes those drafts to a person who can check the substantiation file. Everything else can flow. This is a cheap classifier and it is the difference between a generator that produces work and a generator that produces regulatory correspondence. It is also not legal advice: what counts as a claim, and what substantiation is needed, is a question for whoever advises the brand.

Personalisation and the confident wrong detail

The classic personalisation failure was the unfilled merge field: “Hi {first_name}”. It was embarrassing and obviously a bug. The generated equivalent is worse because it is not obviously a bug — a model given a thin CRM record writes a warm, specific opening referencing a project the recipient never worked on, at a company they left, and it reads as though a person got it wrong, which is a different and more damaging impression.

Constrain personalisation to fields you hold and can verify, and have the generator emit which record fields it used so a spot check is possible. Where the record is thin, the correct output is a shorter, less personal message. A template that degrades gracefully beats one that invents rapport.

Measuring the thing rather than the output

Content volume is not a result, and it is the metric that a generation project produces effortlessly, which is why it gets reported. The measures that mean something are the ones that were already used to judge the work: for organic pages, entries and assisted conversions per page over a ninety-day window, with a holdout of pages you did not publish so seasonality does not get credited to you. For email, the usual randomised split, with unsubscribe rate weighted as heavily as click rate, because the fast way to raise clicks is to spend list health. For paid, the existing experiment infrastructure with variants as the unit.

None of that is specific to generation, which is the point: if a content programme could not have been measured before, generating it faster does not make it measurable, it just makes the unmeasured thing bigger.