How to run AI in your marketing campaigns without becoming a content mill

This note is background; current work is Web & eCommerce packages and US → EMEA market entry.

Every mid-market marketing team has now met the same temptation: the tools can produce a month of content in an afternoon, so why not let them? The companies that gave in are discoverable everywhere and read nowhere — feeds full of interchangeable posts, blogs that answer nothing, and a brand voice sanded down to the texture of everyone else’s. That is the content mill, and it fails for exactly the same reason most agent projects fail: unscoped generation. Output nobody defined, against acceptance criteria nobody wrote, measured by nobody.

The alternative is not “less AI”. It is the same move that fixes agent projects: put a system around the generation — a campaign operating layer where every step has an owner, an input, and a definition of done, and AI fills only the steps it is demonstrably good at. We run our own marketing this way, so this post can describe the machinery from the inside rather than in the abstract.

Campaign anatomy: five decisions before any tool opens

A campaign is five decisions. Every one of them is thinking work, and thinking work is precisely what generation tools are worst at originating — and best at accelerating once made.

Element	The decision	What AI can do	What it must never do
ICP	Who exactly is this for?	Summarise interview notes, cluster themes	Invent the audience’s problems
Offer	What are we actually selling, at what price?	Draft variant phrasings of a decided offer	Set or imply pricing
Channel	Where does this audience already pay attention?	Adapt one asset to channel formats	Choose the channel by generating for all of them
Proof	Why should anyone believe us?	Format real evidence you provide	Fabricate testimonials, metrics, or client names
CTA	What is the one next step?	Propose copy variants	Promise anything the offer doesn’t deliver

The pattern in the two right-hand columns is the whole post in miniature: AI accelerates steps with clear acceptance criteria and corrupts steps without them. “Draft four headline variants for this decided offer, under 60 characters, no hype words” has a definition of done. “Write us a campaign” does not — and what you get back will read like it.

The content engine pattern: how our own Insights run

Here is the system behind the very blog you are reading — described as a process, because it is one, and any operator can copy it.

1. Topics come from evidence, not brainstorms. Search Console impressions tell you which questions the market is already asking you specifically. Our own data — a domain only weeks old — showed discovery clustering around agents and data readiness rather than the pages we might have guessed. So that cluster gets deepened. A topic backlog built from impression data is slower and duller than a brainstorm, and it converts attention that already exists instead of hoping to conjure new attention.

2. One language is the source of truth. Each piece is written once, properly, in a master language — argument, structure, sources, claims all settled there. Cross-language consistency errors almost always come from parallel drafting; a single source of truth removes the class of error instead of catching instances of it.

3. Other languages are native rewrites, not translations. An AI translation pass is a fine first draft — that is a step with clear acceptance criteria. But what ships must read as if written by a native professional in that market: idiom, examples, and register adjusted, not transposed. A Spanish reader can tell within two sentences whether a text was written or converted, and the conversion tells them what you think of their market.

4. Every piece is built to be quoted by machines. AI assistants are now a real referral channel, and they cite content they can extract cleanly: a 40–60-word citable answer near the top, four FAQ pairs with schema markup, defined terms, named sources with URLs. This is answer engine optimisation, and it is less a trick than a return to writing that says what it means — machines are simply a second audience that rewards precision.

5. Every piece points at a real offer. Content without a next step is a hobby. Each post closes quietly onto something real a reader can act on — in our case an Efficiency Audit — with a tagged link so arrivals are attributable. No pop-ups, no gates, no countdown timers. One honest door.

6. Every piece ships with one distribution atom. Publishing is not distribution. Each post generates exactly one companion asset — for us, a short LinkedIn post in the company’s voice. One, done well, beats ten scheduled fragments, and it keeps distribution effort proportional to how much attention any single piece deserves.

Notice what makes this an operating layer rather than a pipeline of prompts: evidence enters at the top, a human gate sits before publication, and measurement closes the loop at the bottom. Generation happens in the middle, inside guardrails. The same shape as a well-run customer-facing automation — because it is the same problem.

What to automate, what to keep human-only

The dividing line is acceptance criteria. If you can state what a correct output looks like before generating it, automate the draft. If correctness depends on judgment, accountability, or facts only a human can vouch for, the step is human-only — not human-reviewed, human-done.

Automate freely (with review)	Human-only, always
Outlines and structure proposals	Claims about results, savings, or outcomes
FAQ drafts from an approved source text	Pricing and anything implying it
Headline and description variants	Testimonials and client references
Translation and adaptation first passes	Legal, compliance, and regulatory copy
Schema markup, meta descriptions, alt text	Positioning — what the company stands for
Repurposing one approved asset into formats	The decision to publish

The right-hand column is short, but it is where every AI-marketing horror story lives: the invented statistic, the fabricated review, the compliance claim nobody checked. Note that the left column still says with review. Automated drafting is safe; automated publishing is how mills happen. One human gate, at the end, always — the same principle as the human checkpoint before an agent’s consequential actions.

Measurement without theatre

Here is the part most content marketing quietly lies about, so let us not. A new domain earns impressions before clicks, and clicks before leads — usually months before. Our own early Search Console numbers show exactly that shape: hundreds of impressions, single-digit clicks, a young domain doing what young domains do. Publishing honest small numbers internally beats performing big ones, because decisions made on performed numbers are wrong decisions.

What to actually track, as a chain rather than a single hero metric:

1. Impressions per topic cluster — is the market seeing you for the themes you chose? This is your earliest signal, and it arrives first.
2. Clicks and click-through rate — do your titles and descriptions earn the visit once you’re seen?
3. Tagged arrivals — UTM parameters and source URLs on every CTA, so a contact-form submission can say which piece sent it. Cheap to set up on day one, impossible to retrofit honestly.
4. Real inquiries — the number that matters, and the slowest to move. For a mid-market operator publishing weekly, expect a traffic step-change around month three and the first steady content-sourced inquiries a couple of months after that. Anyone promising faster is selling the dashboard, not the outcome.

The discipline this buys you: when a cluster earns impressions but no clicks, you fix titles, not strategy. When clicks arrive but no inquiries, you fix the offer or the CTA, not the writing. A measured chain tells you which step is broken. A vanity metric only tells you to feel good or bad.

A 90-day starter plan for a mid-market operator

Assume a services or hospitality business, one marketing owner, a few hours of leadership attention per week, and standard AI tooling — no new platform purchases.

Days 1–30: build the layer, publish nothing new. Write the campaign brief: ICP, offer, proof inventory (real evidence only — what you can actually show), channel, CTA. Pull whatever search data exists — even a thin Search Console profile shows which queries already surface you. Write the acceptance criteria for each content type: length, structure, an answer box, an FAQ block, a banned-words list, a sources rule. Set up UTM conventions and make sure form submissions record arrival sources. This is the unglamorous month; it is also the one that decides whether you are building an engine or a mill.

Days 31–60: run the engine at low volume. One flagship piece per week, through the full chain: evidence-picked topic → brief → AI-assisted draft → human edit that a named person signs → publish → one distribution atom. Resist volume; you are debugging a process, and debugging is easier at one piece per week. Expect impressions to start moving late in this window and nothing else to move yet. That is the plan working.

Days 61–90: measure, prune, and add languages if they’re real. First honest review: impressions by cluster, CTR, tagged arrivals. Double down on the cluster that shows life; prune the one that doesn’t. If you serve a genuinely multilingual market, add the second language now as native rewrites of your proven pieces — proven content in a new language beats new content in a proven language. End the quarter with a one-page report of real numbers, including the disappointing ones. That report is the asset: it is the difference between a marketing function that learns and one that performs.

The quiet close

The mill and the engine use the same models, cost roughly the same tokens, and take roughly the same hours. The difference is entirely in the system around the generation: evidence in, criteria on every step, a human gate before publish, measurement after. That system is a design problem — the same design problem as any other well-governed automation, which is why marketing teams and operations teams keep arriving at the same architecture from opposite ends. It is also, not incidentally, how we think about our own craft.

If you want a second pair of eyes on where AI fits your marketing operation — which steps to automate first, what the measurement chain should look like for your funnel, whether a campaign engine and an operations agent should share plumbing — start with a Web & eCommerce quote or US → EMEA. No mill required.