I audited my AI marketing team against the 22 Immutable Laws

My OpenClaw content team could tell me which version of a post passed review and when Buffer sent it. It couldn't reliably tell me whether that version led to a useful buyer conversation. I nearly responded by giving the agents more marketing rules.

For plenty of marketers, Al Ries and Jack Trout's book is close to scripture. Leadership. Category. Focus. Candor. Sacrifice. Resources. Twenty-two compact rules for earning a position in someone's mind.

I run a team of AI agents that researches, writes, edits, fact-checks, produces media, and manages content operations. If the laws are that useful, why not put them into the agents' constitutions and make the whole team follow them?

That question turned into an audit of eight outward-facing agents, all 22 laws, our content strategy, our agent journals, and the small amount of engagement data we can currently trust.

The answer was less tidy than I expected.

The agents already followed many of the laws. The missing piece was a way to connect their content decisions with what readers did afterwards.

The team already knew more marketing than its instructions admitted

I started with the agents that shape anything the public sees.

None of their charters says, "Apply the Law of Focus" or "Use the Law of Candor."

The behaviour is still there.

Our positioning work tries to create and own a category: the "agentic research house." That's the Law of Category. Our voice guide insists on one clear point, exact evidence, and no generic AI preamble. That's Focus. We publish failures and unresolved measurements instead of smoothing them away. That's Candor. We have explicit rules about what the account should not talk about. That's Sacrifice.

Even Resources was hiding in plain sight. Agents make drafting cheap. My review time, judgement, and willingness to answer a serious reply are still scarce.

The retrospectives were useful too. Archie had written about keeping proof separate from positioning. Belle had documented the need for distinct creative directions. Oli had shown why the writer should not be the final reviewer. Liv had repeatedly caught the difference between "connected" and "published."

Our internal evaluation data for Sara was the sharpest example. Across 89 tasks and five rounds, her best role was not cold-writing content from scratch. It was reviewing and elevating an existing draft: a +4.50 average improvement and a 9.25 score. Solo creation was harmful on average at -3.3.

That changed how I looked at the team. The agents had marketing instincts, but those instincts were scattered across charters, journals, skills, and old campaign plans.

They weren't connected to outcomes.

The missing join

Our content system could tell me that a post had passed a voice check. It could tell me which humanizer ran, which claims were verified, when Buffer sent the post, and whether the public URL existed.

It could not reliably answer a more useful question:

Which exact decision in this exact version produced a better conversation with a potential buyer?

We did not have a consistent join between:

• the approved copy and its exact revision;
• the audience and positioning hypothesis;
• the hook, format, evidence, and closing question;
• the people who replied;
• whether those replies became a real discussion;
• whether a discussion became a qualified DM;
• whether the DM became a commercial next step.

The old metrics path made the gap obvious. It was built around Twitter and Typefully, counted likes, retweets, replies, quotes, and impressions, and calculated a conventional engagement rate. It did not cover LinkedIn, reply depth, conversation branches, qualified DMs, or the content revision that created them. Typefully is no longer our scheduler of record anyway.

I had a marketing doctrine and an activity counter.

I did not have evidence that one caused the other.

That matters because the 22 laws are memorable enough to become dangerous. A rule can feel true long before it has earned a permanent place in an agent's constitution.

I turned the laws into hypotheses instead

I grouped the laws into questions the system could test.

The category laws ask whether "agentic research house" is a useful position or just language we enjoy using. Does it increase relevant profile visits, informed replies, and inbound interest compared with plainer descriptions such as AI engineering lab or consultancy?

The focus laws ask whether one claim, one audience, and one intended conversation outperform posts carrying three lessons and two calls to action.

The attribute and opposite laws ask whether "auditable," "operator-led," and "evidence-first" create a credible contrast with opaque AI promises. What matters is what readers infer from the post, not how strong the words sound to us.

The candor law asks whether a failure-with-fix creates more trust-bearing conversation than a success-only case study.

The success law raises a problem the original book could not have anticipated at this scale. Once an AI-assisted format works, the system can reproduce it endlessly. That is efficient right up to the moment every post starts sounding like the same competent machine.

So each law became a row with five fields:

1. What we already do.
2. What we believe it changes.
3. How we could test it.
4. How it could fail.
5. What evidence would justify making it permanent.

That last field is the important one.

No evidence, no constitutional rule.

AI made the creativity problem stranger

The current research pushed me further away from automatic doctrine.

In a randomized experiment published in Science Advances, Anil Doshi and Oliver Hauser found that access to generative-AI ideas improved how individual stories were evaluated, especially for less-creative writers. The catch: the AI-assisted stories became more similar to one another.

Better one by one. Narrower as a group.

That is almost a perfect description of the risk in an AI content team. Every agent can produce a polished answer. The portfolio can still collapse into the same voice and the same safe structure.

So I am changing how I use AI rather than removing it from the process.

Belle should produce directions that are structurally different, not three cosmetic versions of the same hook. Sara should test whether a line could only have come from my actual work. Oli should look for portfolio sameness as well as sentence-level AI tells. Quill should surface first-hand evidence from agent journals before we reach for commentary on the latest news.

The practical test for those directions is whether they use evidence from our actual work and differ in more than wording.

Likes are the bottom of the ladder

I also needed to define what "better" meant before testing anything.

The hierarchy I care about is simple:

like < reply < sustained discussion < qualified DM < commercial continuation

Likes still matter. They tell me whether the post reached anyone and whether the opening registered. They are a diagnostic, not the destination.

A substantive reply is better because someone added experience, evidence, a reasoned objection, or a real question. A reply that causes another person to reply is better again. Now the post has created a conversation that can continue without me carrying every branch.

A qualified DM is the strongest social outcome: someone has a relevant problem, wants to collaborate, asks for an introduction, or is interested in working with me or Redditech Labs.

The research supports being careful here.

LinkedIn and Edelman's 2025 work found that thought leadership influences both target buyers and people who sit outside the obvious buying committee. LinkedIn reports that 41% of target buyers and 35% of hidden buyers said a C-suite executive had encouraged them to consider a vendor after engaging with that vendor's thought leadership.

That is vendor-sponsored survey evidence, not proof that a specific post caused a sale. It still tells me not to treat visible likes as the whole funnel.

Buffer's analysis of 72,000 LinkedIn posts across nearly 25,000 profiles found that creator replies were associated with roughly 30% higher engagement. Its wider cross-platform analysis reported an 8% association on X.

Associated. Not guaranteed.

People may reply more when a post is already doing well. The useful response is to protect time for real conversation and test the practice locally, not carve "reply in the first hour" into an immutable law.

Spotify's recommendation research reaches the same problem from another direction: clicks are often weak proxies for satisfaction. If the reward is wrong, the optimizer gets very good at producing the wrong result.

An agent system trained to chase engagement will find engagement.

That is not the same as earning trust.

The experiment I want to run

At the time of this October 22 audit, I had designed an eight-week pilot. It hadn't started.

Week 0 is instrumentation. Each post gets an immutable content ID, the exact approved-copy hash, the audience, the intended perception, the law being tested, the proof source, the publishing context, and metric snapshots where the platforms expose them. Replies are classified by substance and depth. DMs are stored as consent-safe counts and stages, not copied into a shared analytics file.

Weeks 1 and 2 test focus and the closing question. Diagnostic questions against experience questions. Experience questions against decision questions. One major variable at a time.

Weeks 3 and 4 test category and attribute. "Agentic research house" against a plain practitioner frame. "Auditable" against a generic AI benefit claim. If the category language confuses what Redditech actually does, it loses.

Weeks 5 and 6 test evidence and candor. Success-only proof against failure-with-fix. Text against an annotated artifact or native short video where the source supports it.

Weeks 7 and 8 test the human response window. Some approved posts will have a protected 45 to 60 minutes for thoughtful replies. Others will follow our normal availability. The goal is to learn whether human follow-through changes discussion quality without pretending we can control the feed.

Only repeated results get promoted into agent instructions. At least two topics or eight comparable posts. No trust or voice regression. No result driven by one outlier or the same two friendly accounts. No tactic whose operational cost is worse than the value it creates.

Everything else stays a tactic, gets revised, or gets discarded.

The part I cannot measure yet

The October 22 audit mapped the agents' existing practices to the laws and proposed a measurement design. Platform metrics were incomplete: LinkedIn reply graphs and qualified DMs weren't joined to the exact content revision. With the pilot still unrun, I had no engagement result to attribute to the proposed changes.

The next move is smaller: build the read-only ledger, establish the baseline, and let the content earn its way into the constitutions.

If you were instrumenting this first, would you optimize for reply depth or qualified DMs? More importantly, how would you stop the metric from turning the writing into bait?

Sources

• Al Ries and Jack Trout, The 22 Immutable Laws of Marketing.
• Anil R. Doshi and Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content", Science Advances, 2024.
• LinkedIn, "How B2B Marketers Can Use Thought Leadership to Persuade Hidden Buyers", 2025.
• Buffer, "Replying to Your Comments on LinkedIn Boosts Engagement by 30%", 2025.
• Spotify Research, "Deriving User- and Content-specific Rewards for Contextual Bandits".