Marketing: AI Agents
Continuous analysis of search, competitors and customer questions, the content it produces, and the editorial line that stops volume becoming the strategy.
Markdown. No sign-up, no email.
Marketing is where cheap generation does the most damage fastest, because the failure looks like success for about two quarters. Output rises, the calendar fills, the dashboard is green, and nothing compounds.
The discipline here is not producing more. It is deciding what deserves to exist.
What is analysed continuously#
| Input | What it produces |
|---|---|
| Search trends | Demand that exists rather than demand we assume |
| Competitors | What is covered well, and the gaps nobody has filled |
| Keywords | Intent and difficulty, which matter more than volume |
| Industry trends | Topics arriving before they are saturated |
| Social trends | What the audience argues about, which is different from what it searches |
| Customer questions | The single best input available, and the most underused |
| Content performance | What actually earned attention, and what quietly did not |
Customer questions outrank every other input. Search data tells you what people type; a support ticket or a sales objection tells you what they actually need to know, in their words, including the part they were embarrassed to ask.
What can be produced#
Articles, LinkedIn posts, video scripts, case studies, whitepapers, newsletters, landing pages, SEO pages and campaigns.
All of it is drafting. None of it is publishing.
The editorial line#
Three rules, and they are the difference between a knowledge platform and a content farm.
1. Nothing publishes that says nothing. The test is whether a reader who already knows the basics learns something. Most generated content fails this while reading perfectly well, which is precisely why it needs a person to catch.
2. No claim without a source or a measurement. Anything numeric is sourced or removed. Generated text produces plausible statistics with total confidence, and a wrong number that we published outlives any traffic it earned.
3. Thin content suppresses the whole domain. Publishing fifty weak pages to fill a calendar does not add fifty pages of value; it lowers the assessment of the entire site. Fewer, better, is not an aesthetic preference here, it is how the mechanism works.
The four agents#
Demand agent#
Watches search, competitors and questions, and produces a ranked list of what is worth writing, with evidence: real demand, an actual gap, and something we can say from experience.
Draft agent#
Produces the draft from a brief and our own published material. It writes; it never publishes.
Performance agent#
Reports what earned attention, and specifically what did not. The list of underperforming content is the more useful half and the one that never gets read, because deleting or rewriting is less appealing than publishing something new.
Distribution agent#
Schedules and reformats across channels. Reformatting one good piece for four channels is worth more than four separate weak pieces, and it is the cheapest way to get more from the same work.
What stays with a person#
- Publishing. Every time. This is the whole safeguard.
- Any claim about the company, what we have delivered, who we work with, what we can do.
- Positioning. What we are and are not, which is strategy rather than content.
- Anything about a named customer. Their approval, in writing, before it exists publicly.
- Deciding not to publish something already drafted. Sunk cost is now nearly zero, which makes killing a draft cheap and therefore the right habit.
The failure this design is built against#
A publishing schedule met every week by material nobody finishes reading, on a domain that slowly stops ranking for anything because most of what it contains is filler.
Cheap generation makes that outcome the default. The editorial line is the brake, and it only works if someone has the authority to say no to a draft that already exists.