How to Pick Your First 10 SKUs for Agentic Checkout (Without Wrecking Your Margins)

What does “picking agentic SKUs” actually mean?

In Shopify’s Agentic Storefronts world, not every product should be eligible for native, in‑chat checkout.

You effectively have three states for each SKU:

• agentic_mode = 'checkout' – AI can recommend and complete checkout natively
• agentic_mode = 'discovery' – AI can recommend, but must link to your site
• agentic_mode = 'off' – AI shouldn’t sell or even highlight this SKU

Your first 10 agentic SKUs are the ones you allow into full autonomous selling. Everything else either stays discovery‑only or off while you learn.

This is not a creative decision. It’s a unit‑economics and risk decision.

Why SKU selection matters for profitability

Agents don’t care about your margin. They care about satisfying constraints: price, specs, availability, policies.

If you let them:

• Push low‑margin SKUs, your Contribution Margin and LTGP per AI‑origin customer will lag other channels.
• Push high‑return SKUs, your “AI revenue” will come with ugly refund costs and customer complaints.
• Push fragile or complex SKUs, you’ll eat mis‑sells, support load, and chargebacks.

That shows up as:

• Lower contribution per order vs Meta/Google/email
• Higher effective discount rate and return rate by AI channel
• LTGP:CAC that looks worse, even if CAC is “cheap” on paper

Selecting only clean, high‑margin, low‑risk SKUs for checkout lets you:

• Keep LTGP:CAC ≥ 3:1 for AI‑origin cohorts
• Maintain healthy CAC payback (close to or better than other channels)
• Learn agentic behavior on your safest, most profitable products first

The 4 variables that should drive your first 10 SKUs

For each candidate SKU, pull these numbers (last 60–90 days):

1. CM1 per unit
   • $$ CM1 = Net Revenue – (COGS + Fulfillment + Pick/Pack + Fees) $$
   • You want strong, stable CM1, not razor‑thin dollars per order.
2. Return‑adjusted margin
   • Subtract refunded COGS and return shipping / processing from CM1.
   • If return‑adjusted margin is weak or negative, that SKU is off‑limits for agentic checkout.
3. Weeks of cover (inventory reliability)
   • $$ Weeks of cover = On‑hand units / Average weekly net units sold $$
   • You want SKUs with reliable stock, not constant stockout risk.
4. Mis‑sell / complexity risk (qualitative score)
   • Sizing nightmares? Fragile or regulated? Contraindications?
   • Rate 1–5:
     • 1 = dead simple (coffee beans, pet treats)
     • 5 = requires human consult (complex sizing, medical‑adjacent)

The “10‑SKU Agentic Scorecard”

Create a simple scorecard with columns:

• SKU
• CM1 per unit (score 1–5)
• Return‑adjusted margin (score 1–5)
• Weeks of cover (score 1–5)
• Mis‑sell risk (score 1–5, but inverted)
• Total score

Scoring guidance:

• CM1 per unit
  • 1 = <$5, 3 = $10–$14, 5 = $20+
• Return‑adjusted margin
  • 1 = barely positive, 3 = solid, 5 = excellent, near zero returns
• Weeks of cover
  • 1 = <2 weeks, 3 = 4–6 weeks, 5 = 8+ weeks
• Mis‑sell risk (invert)
  • 1 = high risk, 5 = extremely simple to sell correctly

Then:

• $$ Total score = CM1 + Return‑adjusted + Weeks of cover + Mis‑sell $$

Sort descending and pick your top ~10 SKUs, with:

• High total scores
• CM1 and return‑adjusted both ≥ “3”
• Mis‑sell risk ≥ “3”

Only these SKUs should get agentic_mode = 'checkout' in your metafields or mapping layer.

How to apply this in your business (step by step)

1. Define your pilot scope
   • Choose 1–2 product families you’re comfortable testing: simple, high‑margin, lower return SKUs.
2. Pull 60–90 days of data for those SKUs
   • For each candidate SKU, calculate:
     • Net revenue
     • COGS
     • Fulfillment + pick/pack + payment fees
     • Returns: units, refunded COGS, return shipping/processing
     • On‑hand inventory and net units sold per week
3. Build the 4‑column scorecard
   • Compute CM1 per unit, return‑adjusted margin, weeks of cover.
   • Assign a mis‑sell risk score (1–5) based on operational judgment.
4. Rank and select your first 10
   • Sort by total score and pick the top ~10 SKUs that pass your thresholds.
   • Flag them as agentic_mode = 'checkout'.
   • Flag complex or low‑margin SKUs as 'discovery' or 'off'.
5. Turn on Agentic Storefronts for a small channel set
   • Enable 1–2 AI surfaces (e.g., Gemini + ChatGPT) with only these SKUs eligible for native checkout.
6. Monitor AI‑origin performance separately
   • Track contribution per order, return rate, and early LTGP for these SKUs by AI channel vs your existing funnels.
   • If they look equal or better, expand to the next tranche of SKUs; if worse, tighten or revert to discovery‑only.

How this fits into the Agentic Commerce framework

On the Agentic hub, we talked about:

• Not dumping your whole catalog into AI channels
• Starting with a curated subset of simple, high‑margin SKUs
• Using agentic_mode = 'checkout' | 'discovery' | 'off' to control exposure

This SKU selection framework is the first gate in that system:

• It keeps AI agents selling products that your P&L likes
• It protects you from early mis‑selling and ugly return spikes
• It gives you a clean, measurable pilot slice so you can answer:
  “Do AI‑origin orders have better or worse contribution than Meta/Google?”

How MarginOS operationalizes this

You can do all of this in spreadsheets. It just gets brittle fast.

Inside MarginOS:

• Standardized CM1 and return‑adjusted margin
  • We compute CM1 and return‑adjusted profit per SKU × channel, so you can see which SKUs are true “profit engines.”
• Weeks of cover and stockout risk
  • We track weeks of cover and highlight stockout risk on high‑margin SKUs, so you don’t send agents to sell what you can’t ship.
• Signals for agentic SKU decisions
  • Signals like “Negative margin SKUs,” “Return spike on SKU X,” and “Margin dogs with heavy paid spend” make it obvious which SKUs should be checkout, discovery, or off as you scale agentic channels.

If you want to see this 10‑SKU scorecard built automatically on your own data, connect Shopify, Google Ads, and your 3PL to MarginOS and let the Profit & Inventory Command Center do the boring math.