A good product recommendation does not need to sound breathless. It needs to answer three reader questions quickly: Who is this for? What useful job does it do? Why is it appearing here? Source-backed creative answers those questions with claims the product owner can verify and the host is willing to stand behind.
This matters when AI acts as the middleman. A model can read a landing page and produce polished copy in seconds, but fluency is not evidence. Without constraints, it may strengthen a careful statement into a guarantee, infer an integration that does not exist, or describe the host as a happy customer.
Use AI to shorten the path from evidence to a draft—not to manufacture evidence.
Start with a claim ledger
Before drafting, collect a small set of allowed claims. Each claim should include the exact source URL, a short supporting excerpt or structured fact, the date it was checked, and whether it needs qualification. This is a claim ledger: a review aid, not a hidden dossier.
| Claim | Acceptable source | Qualification |
|---|---|---|
| “Exports reports as CSV” | Current product documentation | None if generally available |
| “Starts at $9 per month” | Current public pricing page | Include billing period |
| “Built for solo consultants” | Owner-approved positioning page | Positioning, not a usage statistic |
| “Used by 5,000 teams” | Owner-supplied evidence plus public source | Date the number; omit if unverifiable |
Do not scrape private dashboards, customer records, or gated materials to make the recommendation more impressive. Public, owner-approved product material is usually enough.
A five-stage creative workflow
1. Normalize verified facts
Turn the submitted page into a structured product profile: product name, one-sentence purpose, target audience, supported capabilities, exclusions, source URLs, and the date each source was checked. The owner should be able to correct this profile before any matching occurs.
2. Match audience before writing
Creative quality cannot rescue a bad match. First ask whether a meaningful segment of the host’s audience has the problem the promoted product addresses. Exclude direct competitors, unrelated categories, unsafe content, and pairings the host has declined.
3. Draft inside hard boundaries
Give the drafting model only approved facts and a strict output shape. A compact placement usually needs a headline, one supporting sentence, a call to action, and a disclosure. Tell the model not to claim personal use, results, rankings, awards, customer counts, or security properties unless those exact statements appear in the ledger.
4. Run automated checks
Before a human sees the draft, compare every factual phrase with the allowed evidence, validate the destination URL, enforce length limits, and scan for prohibited language such as “guaranteed,” “best,” or “risk-free.” Automated checks reduce review fatigue; they do not replace judgment.
5. Require two approvals
The product owner confirms factual accuracy and destination. The host confirms that the message fits the audience and the location. Either person can edit or reject it. Publication should require the approved version’s immutable fingerprint, preventing a later background rewrite from changing live copy.
Anatomy of a trustworthy placement
| Element | Job | Example pattern |
|---|---|---|
| Context | Explains relevance | “Need a simpler way to…” |
| Specific benefit | States one supported capability | “Turn approved notes into…” |
| Audience cue | Lets readers self-select | “Made for independent…” |
| CTA | Describes the destination | “See how it works” |
| Disclosure | Makes the relationship clear | “Partner recommendation” |
Specificity beats hype. “Create a shareable checklist from your audit notes” is more useful than “supercharge your workflow.” A recommendation should help the right reader decide, not push every reader to click.
Turn risky copy into bounded copy
| Risky draft | Source-backed rewrite |
|---|---|
| “The #1 platform trusted by thousands.” | “A focused workspace for independent teams who need to organize client approvals.” |
| “We use this every day and love it.” | “Partner recommendation: explore a tool designed to simplify recurring reports.” |
| “Guaranteed to double your traffic.” | “Track which partner placements send qualified visits to your product.” |
| “Enterprise-grade security.” | Use the precise published controls—or omit the claim. |
The 60-second human review
- Can I point to a source for every factual phrase?
- Would a reasonable reader understand the partner relationship?
- Does the copy avoid implying personal use or endorsement?
- Is the product genuinely relevant to this audience?
- Does the call to action describe the page it opens?
- Would I still publish this if there were no credit attached?
If any answer is no, revise or decline. Fair systems make declining inexpensive. For the network model around this workflow, read verified cross-promotion: a fairer growth loop. For search and disclosure boundaries, continue with the ethical cross-promotion playbook.
Let AI handle the draft, not the decision
Bring one product and its public source page to Project Relay. You retain approval over every recommendation that enters or leaves your site.
Request access →