Foundations before amplification: the order AI marketing skips

Almost every AI-marketing engagement I get asked to price starts at the wrong end. The brief is amplification: more posts, more citations, more ads, more agents. Amplification is step five. When it is step one, it multiplies a number nobody has defined yet, and the result is a busier month with the same amount of information in it.

I wrote the five steps down on my own domain, with what each one has to produce before the next one is allowed to start: the FREEMAN framework.

F — Foundations: the audit that decides everything after it

Before any channel work there is an unglamorous week: what is tracked, what is duplicated, which conversions are double-counted, which form fields never reach the CRM, what the actual baseline is. The deliverable is not a slide. It is a short document that states, in numbers, what the funnel currently does and which of its numbers are trustworthy.

This is the step that gets skipped because it produces nothing visible. It is also the step that decides whether the next four produce readings or noise.

R — Relevance: the message that survives contact with a list

Relevance means one formulation per channel, tested against a defined audience, with the reply rate as the reading — not impressions. If nobody can say which message is currently winning, there is no relevance step, there is copywriting.

I — Integration: schema, discovery, one system

This is where the machine-readable layer goes in: structured data, an llms.txt, a sitemap that matches reality, pages that answer a question in the first paragraph because that is the part an answer engine quotes. Integration is not an integration project. It is the decision that the system reads from one set of definitions.

M — Measurement: one owned number per channel, monthly

Each channel gets exactly one number it owns, stated before the month starts, plus a stop rule expressed as a cost per qualified conversation checked on a fixed date. A channel that reports four numbers owns none of them.

The arithmetic here is unfriendly and worth knowing in advance: one strict two-variant test at a 3% base rate with a +20% relative lift needs 13,914 contacts per arm — 27,828 in total at 80% power. At 500 contacts a day that is 56 days. Media money is rarely the constraint; volume is. You can check it yourself with a free test planner and a queue planner.

A — Amplification: only what already proved itself

External signals last: distribution, listings, posts, PR, archive snapshots. By the time you get here you know which message to amplify and which number is supposed to move — so amplification has a kill rule instead of a content calendar.

The order is the product

The steps are cheap to name and expensive to reorder. A team that starts at A will spend the quarter learning that its baseline was wrong; a team that starts at F and R will finish the quarter with one verified message and a channel that can be scaled or stopped on evidence. Neither quarter costs more. Only one of them ends with a decision.

I run this as a fixed-scope engagement: Sprint $900 one-off (foundations, definitions, tracking, stop rules — yours whether or not you continue), Engine $1,900/month (one channel end to end, a written record and an owned number each month), Full Build $2,900 (the whole loop including the pipeline). Prices and deliverables are published on the offer page, and the artifacts this produced on my own domain — packages, accepted third-party listings, archived snapshots — are on the proof page. The method also ships as an npm CLI and a public playbook repository if you would rather read the operating file than take my word for it.