Say you sell three versions of the same product: a standard, a “Pro” and a smart one with an app. You rewrite the Pro’s listing, run a few test questions in ChatGPT, and the Pro shows up a lot more. Good news?
Maybe. It depends on where those picks came from. If the AI used to recommend your standard model and now recommends the Pro instead, you didn’t win anything. You moved a customer from one of your products to another, and maybe to a lower-margin one.
That’s sibling SKU cannibalization. It’s easy to miss, because almost every AI visibility check looks at one product or one brand name at a time.
It already shows up in AI answers
Mike Taylor, Every’s head of evals, tested how AI agents pick articles to recommend. He rewrote one article’s title and description. That article jumped from almost never cited to cited nearly every time, and Every’s share of all citations for that question went from 35% to about 60%.
Then he noticed where part of that came from. His other article on the same topic fell from being cited in 65% of sessions to 19%. Every still came out ahead, so he called it a good trade. His advice: if you have several pieces on the same topic, decide which one should win and make it the most obvious answer.
Those were articles, not products. But an AI shopping assistant building a short list does the same thing. It has only a few slots to fill, and your products compete with each other for them.
What it looked like on our shelf
We saw the same pattern in our own product-picking simulation. We gave AI models a shelf of eight quiet-humidifier listings and asked each one to pick the best two for a shopper. Seven listings were real. The eighth was a humidifier we invented, and we tested different versions of its copy.
Four of the seven real listings were from one brand, which we’ll call Brand A, and two of those were close siblings: a standard model and its smart version with app control. Same tank, same mist output, USD 10 apart.
With our invented product’s copy vague, it almost never got picked: 3 times in 288 Claude sessions. When we wrote its facts as a spec line, it got picked 213 times. Here’s where those 210 extra picks came from:
| Listing | Picks, vague rival | Picks, rival with spec line | Change |
|---|---|---|---|
| Our invented humidifier | 3 | 213 | +210 |
| Brand A, standard model | 84 | 38 | −46 |
| Brand A, smart model (sibling) | 81 | 20 | −61 |
| Brand A, two other models | 76 | 25 | −51 |
| Brand B | 99 | 48 | −51 |
| A 4 L humidifier at USD 39.99 | 233 | 232 | −1 |
| A top-fill humidifier from a third brand | 0 | 0 | 0 |
Claude Haiku, Sonnet and Opus 5.5, 288 sessions per version, two picks each.
Read it from the brand’s side. Its two siblings went from 165 picks to 58. Across all four of its listings, it lost 158 picks, three-quarters of what the newcomer gained. The cheap 4 L humidifier didn’t lose anything. It had the lowest price and a stated noise level, and it kept getting picked.
The other thing that changed was the shortlist itself. With the vague rival, the standard and smart models were picked together, filling both slots, in 54 of 288 sessions. Once the rival had stated facts, that happened once.
If you were that brand counting picks for the smart model alone, you’d see a drop from 81 to 20 and have no idea whether a competitor or your own standard model took them. Counted as a family, the story is clear.
How to check your own product family
You don’t need a lab for this. You need the same handful of questions, asked before and after a change, with the answers written down by family.
- List your siblings. Every product a shopper could reasonably pick instead of the one you’re changing: other sizes, the Pro, last year’s model, the bundle.
- Pick five to ten shopper questions that any of those siblings could answer. “Quiet humidifier for a big bedroom under $90”, not your exact model name.
- Before you change anything, run each question three times in the assistants you care about. For each answer, write down which of your products appear, which competitors appear, and how many slots there were.
- Make one change to one product. Just one, so you know what caused what.
- Wait until the change shows up where the assistant reads it, then run the same questions again, the same number of times.
- Fill in a family table:
| Before | After | |
|---|---|---|
| Answers where the changed product appears | ||
| Answers where any sibling appears | ||
| Answers where any of your products appear | ||
| Answers where your products fill every slot | ||
| Answers where a sibling was swapped for the changed product |
The line that matters most is “any of your products”. If that went up, the change helped your business. If only the changed product went up and the family stayed flat, you moved picks around. Then it comes down to which product you’d rather sell.
A few answers per question won’t give you precise numbers. AI answers vary from run to run. Look for big, consistent moves, and if the family line barely moved, treat it as no change.
When moving picks is fine, and when it isn’t
Cannibalization isn’t automatically bad. Moving shoppers from a product you’re discontinuing to its replacement is the whole point. Moving them toward the version that actually fits their need is good too: in our receipt checks, the clearest reasons named exactly which sibling stated which fact.
It’s a problem when:
- The shift goes to a lower-margin version, and the family total didn’t grow.
- The AI can’t tell your siblings apart. If two listings say nearly the same thing, the assistant may pick between them almost at random, or borrow one’s facts for the other. Give each sibling the one fact that sets it apart, stated near the top of its own listing. Our guide to variant data covers keeping those facts on the right product.
- Your family used to fill the whole shortlist and now doesn’t. Two of your products in a shopper’s top two is the best outcome there is. Losing that is a real cost, even if your total pick count looks steady.
How we counted
We recounted the saved picks from our October 27, 2026 simulation: every valid Claude session for the humidifier, using the original vague copy and the spec-line copy, across all three shopper briefs and both test rounds. That’s 288 sessions per version, two picks each. This is a bigger set than the held-out subset on our research page, so the counts differ from the ones there. The shelf was eight listings: seven real Shopify catalog listings cut to short snippets, plus one product we invented. We grouped picks by brand and counted sessions where both Brand A siblings were picked.
Your next step
Before your next product page rewrite, write down your sibling list and three shopper questions, and run them once. It takes twenty minutes, and it’s the only way to know afterward whether you grew the family or just reshuffled it.