Generative AI for content and marketing teams: a working guide
Generative AI helps content teams most with the work around the idea: summarising research, drafting from a detailed brief, repurposing, producing variants, writing product descriptions and localising. It hurts when it is asked to supply the idea, which is how generic copy and factual errors get published. Keep a named editor accountable for every piece, give the model a voice guide it can follow, and measure outcomes instead of volume.
Where does generative AI help content teams?
It works when you supply the substance and ask the model to change its shape. It fails when you ask the model to supply the substance.
| Task | How well it works | What the editor still does |
|---|---|---|
| Research summaries of material you provide | Well | Checks the summary against the source, because models misread with confidence |
| First draft from a detailed brief | Well, if the brief contains the argument | Supplies the angle and evidence, rewrites the opening |
| Repurposing a talk, webinar or report | Very well | Chooses what matters for each channel |
| Variants for testing | Very well | Decides what is being tested, removes lookalikes |
| Product descriptions at scale | Well from structured product data | Checks attributes, reviews a sample from every batch |
| Translation and localisation | Well for widely spoken languages | A native reader checks idiom, legal terms and names |
Take a four-person marketing team at a B2B software company. Each month they run one webinar in which an implementation engineer walks through a real customer problem. That 45-minute transcript contains things nobody else has: the engineer’s examples, the questions customers asked, the mistakes they keep seeing. From it, a model can draft an article, five short posts, a newsletter section and several FAQ entries in an afternoon. The substance is theirs, and the model only reshapes it.
Now ask the same model for “600 words on why invoice automation matters”. It will produce 600 smooth words that could sit on any competitor’s site. Same tool, opposite result.
Three rows in the table deserve a note.
Product descriptions generated from structured data (dimensions, materials, compatibility) are among the safest uses, because every claim can be traced to a field. Generated from a product name alone, they invite invented features.
Variants are the quiet win. For a subject line or an advert, ask for options built on three or four distinct angles, such as price, speed, risk and proof, instead of twenty rewordings of one sentence. The test then tells you which angle works, and that is knowledge you can reuse. A winning sentence is not.
Localisation is more than translation. Give the model the market: currency, units, spelling, the local name for what you sell, examples that make sense there. Keep a glossary of terms that must never be translated, such as product and feature names, and pass it with every request. Brand names, legal wording and humour still need a native reader before anything goes live.
Where does it hurt?
Generic copy. Asked for a topic, a model returns something close to the average of what has already been written about it. There is no point of view and no experience in it. Readers recognise the rhythm within a paragraph.
Factual errors. Invented statistics, features your product does not have, quotes nobody said, links that lead nowhere. Google’s own guidance on generative AI content explains why: these models predict plausible text, they do not look facts up.
Sameness. When every team in a sector uses the same tools with the same prompts, structure and phrasing converge. Your tenth article resembles your first, and both resemble your competitor’s.
Tired editors. Reviewing is harder than it sounds. By the thirtieth fluent draft, people skim. Volume without review capacity is how errors get through.
Fabricated proof. Never generate reviews, testimonials or customer stories. Apart from the damage to trust, the US Federal Trade Commission’s rule on fake reviews explicitly covers AI-generated ones.
A workflow that keeps the editor in charge
- Brief. A person writes it: the reader, the single point of the piece, the evidence, what to leave out. If you cannot write the brief, the piece is not ready, and no model will fix that.
- Source pack. Attach the transcript, product notes, data and approved claims. The model writes from these, not from memory. For a large archive, put retrieval over your own material so that drafts cite it.
- Outline. The model proposes two or three. The editor chooses and reorders.
- Draft. The model writes to the outline and marks any statement it could not find in the source pack with [CHECK].
- Edit. The editor works on argument, voice and the opening, and usually cuts.
- Fact-check. Covered below.
- Sign-off. One named person reads the final version and owns it.
The brief is where the quality is decided, so it is worth showing one. For the webinar article it might read:
Piece: article from the October webinar, about 1,200 words
Reader: finance lead at a 50-person firm, thinking about automating
invoice chasing
One point: late payment often starts with a mistake on the invoice, and
three checks before sending catch most of them
Evidence: the three error types shown in the session (transcript 12:40
to 21:15), the audience question about credit notes (31:02)
Leave out: pricing, the roadmap item mentioned near the end
Sources allowed: the transcript, help-centre articles 14 and 22
Six fields, ten minutes of an editor’s time, and the draft that comes back is about something.
Two rules hold the process together. The model is never the last reader. And the tooling should record which sources, prompt and model produced each draft, so that a correction six months later can be traced.
Budget the editor’s time honestly. A draft built from a good source pack usually takes far less effort to finish than writing from a blank page. A draft built from a one-line prompt often takes longer to repair than it would have taken to write.
Teams that run this daily tend to want it built into the CMS, with brief templates, source packs, the voice guide and logging in one place. That is a small internal product build more than a prompt.
A brand voice guide the model can use
Most voice guides were written for people and rely on adjectives such as “confident but warm”. A model needs rules and examples. Something like this, kept short:
# Voice guide for drafting
Reader: finance leads at firms of 10 to 200 people. They know accounting
well and software less well.
We sound like: a colleague who has done the job. Plain words, short
sentences, British spelling.
Always: open with the reader's problem. One concrete example per section.
Numbers only from the source pack.
Never: hype words (list attached), exclamation marks, claims about
competitors, a question as an opening line.
Format: sentence-case headings, paragraphs of four sentences at most.
## Three passages that sound like us
<example>...</example>
## One that does not, and why
<example>...</example>
The examples matter most. Anthropic’s prompting guidance describes examples as one of the most reliable ways to steer tone and format, and that matches our experience. Choose real passages your best writer is proud of, and vary them so the model does not copy one structure.
Treat the guide as a living file. When editors make the same correction three times, it becomes a rule. Our prompt engineering guide explains how to test changes like this instead of guessing.
How should you fact-check an AI draft?
Assume every specific in the draft is unverified until someone has traced it.
- Numbers, names, dates and product claims are checked against the source pack or a primary source the editor opens personally.
- Quotes come only from a recording or transcript, word for word. A model asked to tidy a quote will improve it into something the person never said.
- Do not ask the model for statistics. Ask it to list every claim in the draft that needs a source, then check that list by hand.
- Open every link.
- Check the parts nobody reads in review: titles, meta descriptions, alt text and structured data. Google’s guidance mentions these specifically.
A second model pass that flags unsupported claims is useful as a filter. It is not a sign-off.
What does Google say about AI-generated content?
Google’s position is consistent across three documents, and it is about value, not tools.
Its guidance on generative AI says the technology is useful for research and for adding structure to original content, that all content must meet its usual standards, and that producing many pages without adding value for users may breach its spam policy. That spam policy defines scaled content abuse as generating many pages “for the primary purpose of manipulating search rankings and not helping users”, and adds that this applies “no matter how it’s created”.
Its page on helpful, people-first content asks who created the content, how, and why, and suggests making the use of automation clear where a reader would reasonably expect to be told.
So an edited, accurate article that began as an AI draft is within the rules. Six hundred near-identical location or integration pages are a risk unless each one carries real, distinct information. One operational detail for retailers: Google’s guidance says AI-generated product titles and descriptions in Merchant Center must be labelled as such.
Search is not the only place with rules. Under the EU AI Act, AI-generated text published to inform the public on matters of public interest must be disclosed, unless it has been through human review or editorial control and a person or organisation holds editorial responsibility. The European Commission lists these transparency obligations as applicable from 2 August 2026. That is one more reason for the named editor.
The same qualities decide whether AI assistants cite you, which we cover in answer engine optimisation.
What should you measure?
Not words produced or pieces published. Those numbers will rise whatever happens to quality.
- Cycle time from brief to publication, by content type.
- How much the editor changed. If nearly everything, the brief or the voice guide is the problem. If nothing, nobody is editing.
- Errors caught before publication, and corrections issued after.
- Performance per piece against your own earlier baseline: engaged reading, enquiries, search clicks.
- Test results for variants: how often a generated variant beats the control.
We would not use AI-detector scores for anything. They say nothing about whether a piece is accurate or useful.
Watch editor capacity as closely as output. It is the real limit on how much you can publish safely.
How to start: the first week
Pick one repurposing job with good source material, such as last month’s webinar. Write the voice guide with three real passages. Name the editor. Run one piece through the seven steps and time each one.
Compare the result with the last piece the team produced by hand, on time taken and on quality as judged by someone outside the team. Then decide the first rule you will hold to: nothing is published that a named person has not read in full.
For images in the same workflow, see AI image generation for business. If you want the drafting, retrieval and review steps built into your own tools, that is the kind of generative AI work we do.