Recommendation Design

Edition 2026.08

The Canon of Recommendation Design.

The open description of how the practice is run: the loop, the roles, the rules, and the working vocabulary, kept current as the models change. Follow the method on one brand this week and you're practicing.

Last verified: August 9, 2026

You cannot optimise what you cannot see.
The principle

Why this demands a discipline, and why today

Search engines rank documents. AI engines construct recommendations. The playbooks built for the first barely touch the second: SEO metrics explain only 4 to 7% of why AI cites a page.

And the surface moves. In 2026, ChatGPT's citation volume fell up to 90% and recovered inside eight weeks. A snapshot cannot see that. A quarterly audit cannot see that. Only a running practice can.

The answer about your brand already exists, and it is being revised whether you watch or not. The discipline is how you watch, and how you change it.

The values

The stance every rule descends from

Recommendationoverranking

The unit of work is what AI says about you, not where a page ranks.

Specificityovervolume

One well-decoded gap closed beats ten pieces of general content.

Observationoverassumption

You cannot infer what the model weighs; you watch what it produces and work backward.

Transparencyoverpositioning

Evidence that survives the judge is checkable, dated, and named.

Continuousoverepisodic

Run the loop once and stop, and you did an audit, not the discipline.

The heartbeat

Every engagement runs the same loop.

01

Observe

Run the buyer's actual questions across the engines the buyer uses. Capture the recommendation verbatim: the brands named, the reasoning offered, the risks flagged.

Observe across all five stages of the decision journey, from consideration to advocacy. One stage is a snapshot; five is a discipline.

02

Decode

Extract the structure. What criteria produced this list? What cut brands from it? What is the model treating as fact that is not fact?

The four facets are the parser: elimination triggers, buyer risks, switching hesitations, criteria gaps. Decoding turns a paragraph of AI output into gaps you can close.

03

Seed

Publish specific evidence against a specific decoded gap. FAQ pages against elimination triggers. Methodology pages against criteria. Proof against hesitations. Then echo the claim, in the same words, on the surfaces the models cite.

The content the gap requires. More content is what the channel ignores.

The loop closes on the next Observe. If the recommendation moved, the seed worked. If it did not, the decoding was wrong, the seed was wrong, or the retrieval layer has not caught up. Each of those is diagnosable. The loop, not the seed, is the practice.

The rules

Eight rules. Written for the moments when it's tempting to skip one.

Rule 01

Decode before you seed

Evidence without a decoded gap is content for a channel that will not read it.

Rule 02

Observe all five stages

Winning consideration while losing evaluation is losing the buyer.

Rule 03

Fresh prompts each cycle

Yesterday's decoded criteria go stale within weeks. Rerun; never reuse.

Rule 04

Seed with specificity

One gap, one piece of evidence, one home, plus its echoes. Bulk dilutes trust.

Rule 05

Cite what AI can verify

Dated claims, named third parties, primary sources. Unverifiable is worse than silent.

Rule 06

Compare all four engines

Divergence between engines is a diagnostic signal, not noise. Single-engine RD is not RD.

Rule 07

Never gate the practice

The manual practice is complete. If a step needs a paid tool, it is not the discipline.

Rule 08

Record before you argue

The log holds what the model said, never what you wished it said.

The roles

Three roles. One person can hold all three.

The RD Owner

Owns the discipline across the organisation. Sets the scope and holds every team to one story, because the model reads marketing, sales, and product together, and incoherence is what it recommends against.

The Practitioner

Runs the loop end to end and owns the cadence and the observation log. The person who says "we don't seed until we've decoded." Reports movement, never volume.

The Author

Writes the seeded evidence against a specific decoded gap, on a schedule. A narrower brief than "our content marketer," and a sharper one.

Solo practitioners hold all three roles. Teams should keep them distinct: when the Practitioner and the Author blur on a large brand, the practice degenerates into content publishing that thinks it is decoding.

The cadence

Pick a rhythm you can hold. Then hold it.

Weekly is the aspiration; every fourteen days is the workable floor. Monthly is too slow: when the platform can swing 90% and recover inside eight weeks, a monthly read never sees the middle. Consistency beats frequency, because cycle-over-cycle movement is the read-out, and an irregular log is an unreadable one.

Weekly · the cycle to aim for
14 days · the floor that still works
Quarterly · cold-start re-observation
Annual · scope review

The read-out

How you know it's working, not just running

Presence

Your brand appears in the recommendation itself, not just in the citations under it.

Fidelity

The model recommends you for the reasons you intended. Recommended for the wrong reason is still a gap.

Traceability

Your seeded evidence shows up as reasoning in the answer, not just as a link at the bottom.

Velocity

Your loop is shorter than the platform's swing window, so you can tell your movement from its weather.

Coherence

Sales, product, and marketing describe the brand the same way. The RD Owner's headline metric.

Authority

Other practitioners cite your category framing. The strongest signal, and the slowest.

The boundaries

What Recommendation Design is not

It is not SEO: SEO gets you found; RD gets you chosen. It is not AEO or GEO: those treat the citation as the outcome, where RD treats the recommendation as the outcome and citation as one symptom of it.

It is not content marketing: volume is not the metric, specificity is. And it is not a tool: the discipline runs by hand, in a spreadsheet, in an afternoon, on one brand. An instrument helps when the work outgrows a person, never before.

There is no certification and no gate.

Run the loop on one brand and you're practicing. The toolkit gives you the first cycle: a worksheet to bound your scope and three prompts to run the first Observe and Decode today.

Try it on your own brand today

The first email brings the toolkit: a worksheet and three prompts to run the practice. One email a week after that.

Unsubscribe anytime ·Terms

The full practice, with the prompts and worked examples, is in the book: