The Marketing Engineering Loop
Evidence, decision, implementation, measurement, next decision. The loop is what separates a marketing engineer from someone who generates more work, and keeping it unbroken is the hard part.
Why the loop matters more than the output
AI makes producing work cheap. It does not make choosing the right work cheap, and it does not make you remember why you chose it. A marketing engineer's advantage is a single operating loop that holds its context from the first piece of evidence to the next decision.
Without that memory, you get an endless stream of intelligent opinions. Every new audit can reopen a question that was settled last week.
The five steps and what to record
| Step | What happens | What to record |
|---|---|---|
| 1. Evidence | Gather what the data and the site show | The source of each finding, and whether it is observed fact or interpretation |
| 2. Decision | Choose what to change first | Options considered, the choice, why, and the confidence level |
| 3. Implementation | Make and review the change | The change itself, who reviewed it, and how to undo it |
| 4. Measurement | Compare against a baseline | The baseline, the date, what moved, and what the data cannot prove |
| 5. Next decision | Let the result shape what comes next | Whether to keep, adjust or reverse, and what it changes about the plan |
Step 1: Evidence
Start with what can be observed: whether a page is indexed, what a crawler actually receives, which queries a page appears for, and what visitors do. Keep that separate from interpretation. 'The metadata is the priority' is a judgment. 'This page has no description and is not in the sitemap' is evidence.
Several AI models will usually give you different readings of the same site. One focuses on metadata, another on rendering, another on structure. All can be reasonable. Labelling each finding as evidence or interpretation is what lets you compare them.
Step 2: Decision
Pick the intervention with the best likely commercial effect and write down why it beat the alternatives. Attach a confidence level, so a hunch is not recorded as a fact. Include what you decided not to do, since that is what stops the same question coming back.
Step 3: Implementation
Make the change as a reviewable unit: a pull request with a private preview, automated checks, and a person who approves it before it goes live. Plan how to undo it. A change that cannot be reversed is a bigger decision than one that can.
Step 4: Measurement
Capture a baseline before you change anything, and measure against it. Be plain about causation: a rise after a launch does not on its own show the launch caused it. Report what moved, and what the data cannot prove.
Step 5: The next decision
The result should change the plan. Keep the change, adjust it, or reverse it, and record what you learned. This is the step teams skip, and it is why the loop stops being a loop.
A worked example
An important product page is invisible to Google and to ChatGPT. Evidence: the page has no description or structured data, and the crawler receives almost no text. Decision: fix the rendering and add the markup before writing anything new, with medium confidence that rendering is the cause. Implementation: a pull request with a preview, reviewed and approved. Measurement: compare indexed status and impressions with the baseline over the following weeks. Next decision: if the page is indexed but not cited, look at the content next.
What the loop looks like when it is working
On the ARRIBA site, changes were asked for in plain English, opened as a pull request with a private preview, and approved by the team, with a median of 3 minutes from a change being proposed to its approval. Eight proposals never went live, and one change was undone in a single step. The case study shows how that control worked in practice.
On the DACPros site, 24 reviewed changes shipped in the first month, each approved before going live.
Controls that keep speed safe
The same tools that let you move fast let you implement the wrong idea fast. A technically perfect page can target a query nobody valuable searches, and an agent can optimise a conversion event with no link to revenue. Put controls above the models: previews and approval, recorded decisions, a way back for every change, and measurement against a baseline.
Frequently Asked Questions
What is the marketing engineering loop?
It is the cycle a marketing engineer runs: gather evidence, decide what to change, implement it, measure against a baseline, and let the result shape the next decision. The value is in keeping continuity across all five steps.
Why do different AI models give different recommendations?
Each model weighs the same site differently. One prioritises metadata, another rendering, another information architecture, and changing the reasoning level changes the answer again. The fix is to separate evidence from interpretation and to record why you chose what you chose.
What should be in a decision record?
The evidence found, the options considered, the one chosen and why, the confidence level, who approved it and what baseline it will be measured against. Short entries are enough if they are consistent.
How do you keep fast AI-driven changes safe?
Put controls above the models: previews, review and approval, automated checks, a way to undo any change, and measurement against a baseline. Speed without those makes bad decisions faster.
Run the loop on your site
Chad finds the work, prepares reviewable changes, ships on your approval and measures the result.
Book a walkthrough →