AI in Marketing: More Content Isn’t the Goal — Better Content Is

Marketing is the area where AI gets tried first in most companies. The reasons are obvious: it’s about text and images, the results are immediately visible, and getting started costs nothing.

What’s striking is that disillusionment sets in fastest in exactly this area. After a few weeks it becomes clear: the texts all sound similar, the images feel generic, and the time saved gets eaten up again by revisions.

That’s not down to the tool, but to the goal. Anyone who wants to produce more content gets more content. Anyone who wants better content has to go about it differently.

Why generic texts happen

A language model without context produces the average of what is typically written on a topic. That’s exactly the problem: marketing depends on standing out from the average.

A text about „digitalization in the Mittelstand“ with no further input inevitably contains the same phrasing found in a thousand other texts. It isn’t wrong. It’s just interchangeable – and interchangeable content generates neither attention nor trust.

The way out lies in what only you have: your concrete projects, your arguments, your language, the questions your customers actually ask. A system connected to this material writes differently than an empty chat window.

Where AI genuinely earns its keep in marketing

Research and structuring

The underrated use case. Derive from customer inquiries, support tickets, and sales call notes which questions keep coming up – and build a content plan from that which is based on real demand rather than guesswork. That’s analysis work that hardly anyone does manually.

From expert knowledge to readable text

In technical companies, the knowledge sits with people who rarely enjoy writing. A conversation with an engineer, recorded and transcribed, is an excellent starting point: the model structures it, shortens it, and turns it into a readable article – with the technical substance that no outside copywriter could supply.

This approach produces content that is genuinely distinctive, because the substance comes from within the company.

Adapting to channels

An article becomes a short version for a network, a paragraph in the newsletter, a script for a short video. Pure reformatting work where the content has already been reviewed – ideally suited and with no content risk.

Linguistic consistency

Checking whether a text matches the defined tone, whether technical terms are used consistently, whether the form of address and spelling are correct. That’s tedious work that often falls by the wayside in everyday business.

Translation and accessibility

Multilingual versions and more accessible variants of complex texts – with expert review before publication.

Where caution is warranted

Technical claims. Technical specifications, standards, certifications, performance promises: everything that could be wrong needs to be checked. A model writes an incorrect standard with the same confidence as a correct one.

Competitor comparisons. Statements about competitors are legally sensitive. Such passages don’t belong in an automatically generated text.

Customer testimonials. References and quotes must come from real people and be approved. A fabricated quote isn’t a writing problem, it’s misleading – and potentially grounds for a legal warning.

Images of people and brands. With generated images, third-party rights need to be respected. Other companies’ logos, protected designs, and recognizable real people don’t belong in generated visuals.

Labeling obligations since August 2026

One point that deserves special attention in marketing: the transparency obligations under Article 50 of the EU AI Act have applied since August 2, 2026.

Two aspects are relevant for marketing departments. First, people must know when they are interacting with an AI system – so a chatbot on the website needs an appropriate notice. Second, artificially generated or altered image, audio, and video content that deceptively resembles real people or events must be labeled as such. Disclosure obligations also apply to texts intended to inform the public, with exceptions such as for editorially reviewed content.

In practice, that means: define internally how you handle generated images, how you label them, and who checks them. That’s half a page of policy – but it should exist before the first campaign goes live.

Visibility in a changed search landscape

Search engines and AI assistants increasingly answer questions directly instead of just pointing to pages. For companies, that shifts the question from „How do we rank?“ to „Are we being used as a source?“

The practical consequences are unspectacular and mostly classic good craftsmanship: clearly structured content with unambiguous headings, concrete answers to concrete questions, verifiable information, clean technical markup. What has changed is the weighting: substance and verifiability count for more, volume of text for less.

That leads to a consequence that surprises some people: the reflex to use AI to increase output volume actually works against you. Twenty interchangeable posts a month are worth less to your visibility than two that genuinely answer a question.

An approach that works

1. Define your tone. One page: how you sound, how you don’t, which terms you use, how you address customers. This page is the foundation for every request.
2. Build a material base. Approved descriptions, real project reports, reviewed technical texts. The system works with that instead of generalities.
3. Derive topics from customer questions. Not from topic lists, but from what is actually being asked.
4. Substance from people, form from the machine. The technical content comes from within the company; the polishing can be machine-assisted.
5. Sign-off by a human. Always, without exception.

Anyone who works this way doesn’t produce more content, but better content – and still saves time, because the laborious part, structuring and phrasing, goes noticeably faster.

An editorial routine that holds up

To turn individual texts into a reliable rhythm, a fixed routine helps. A monthly process with four steps has proven effective.

First: collect questions. Once a month, compile the recurring questions from customer inquiries, support tickets, and sales conversations. That’s analysis work AI supports well – and the foundation for topics that actually interest someone.

Second: gather substance. A twenty-minute conversation with the person who knows the answer professionally. Record it, transcribe it, use it as raw material.

Third: process it. Structure it, shorten it, bring it into the defined tone. This is where the share of machine assistance is highest.

Fourth: review and publish. Technical sign-off by the person from step two, editorial approval, labeling according to internal rules.

This routine produces fewer texts than a purely machine-driven output – and considerably more impact, because every text contains something that isn’t found anywhere else.

The internal labeling policy

Set out on half a page how you handle AI assistance: which content types get labeled, what the wording is, who reviews before publication, and which content is fundamentally never machine-generated – such as customer testimonials, reference quotes, and statements about competitors.

This policy isn’t bureaucracy for its own sake. It protects against the case where an intern generates a deceptively realistic image and nobody thought to label it.

Frequently asked questions

Does Google detect AI text and penalize it?

Origin isn’t the measure; quality is. Content becomes a problem when it has no independent value – and that’s exactly what results from producing without your own substance. A well-founded, substantive text is valuable regardless of how it was created.

Should we have images generated?

There’s little argument against it for abstract subjects and backgrounds. Caution is warranted with recognizable people, other companies’ brands, and protected designs, and the labeling requirement applies to deceptively realistic depictions.

How often should we publish?

Better two well-founded posts a month than ten interchangeable ones. The effort per post decreases with routine – the quality shouldn’t decrease along with it.

What about the newsletter and social media?

That’s the cheapest part: a vetted article can be converted into several formats with little effort. Start with the substance, not the channel.

Creating a tone-of-voice guide

The guide is the foundation without which every generated text tends toward the average. It fits on one page and answers six questions.

How do we address readers? Formal or informal address, and when we switch. For mid-market B2B customers, the formal form is the norm, even on social media.

How do we sound? Three traits that apply, and three that explicitly don’t. „Matter-of-fact, direct, no marketing-speak“ is useful. „Innovative, dynamic, customer-focused“ is not, because it fits every company.

Which terms do we use? A short list of defined spellings – product names, technical terms, anglicisms we avoid, and their German equivalents.

Which phrasings do we avoid? Superlatives without evidence, promises we can’t keep, comparisons with competitors.

How long is a typical text? A guideline per format, so that not every post ends up a different length.

What’s our point of view? The most important and most often missing point. What do you stand for – and against what? A text without a point of view is interchangeable, no matter how well it’s written.

Include this guide with every request. The difference in the result is bigger than the difference between two language models.

Measuring impact without fooling yourself

The obvious metric – number of posts published – measures effort, not impact, and encourages exactly the behavior that causes harm. Four other measures are more useful.

Dwell time and scroll depth per post. They show whether a text is read or abandoned after two sentences.

Inquiries that reference a specific piece of content. Ask in the initial consultation how someone became aware of you. It’s imprecise, yet still more meaningful than any click count.

Returning readers. A post that gets a lot of attention once is a lucky hit. Readers who come back are a signal.

Internal reuse. When sales starts sending articles to prospects, that’s the most reliable sign the content is working. It costs nothing to measure – just ask.

The mistake almost everyone makes

It consists of starting with volume. The math seems compelling at first: if a post used to take four hours and now takes one, we could publish four times as much.

The result is a website full of texts nobody reads to the end, a newsletter with a declining open rate, and a team that stops seeing the point after three months. The time saved doesn’t belong in more posts, but in better ones – in the conversation with the engineer, in thorough research, in the second round of editing.