ShelfSight Blog · 2026-07-02
Every AI-visibility tool faces the same tempting shortcut. You ask ChatGPT "best merino socks?", the answer mentions "premium wool options from established brands", and your customer's brand is an established wool brand. Do you count it as a mention?
We don't. That answer scores as unsure, and unsure is excluded from the score entirely. It never counts as presence or absence.
This makes our numbers smaller than they could be. It also makes them true. Here's the full set of rules we hold ourselves to, published on our methodology page:
Decisive-only scoring. A mention counts only when the engine names your store, your domain, or a specific product distinctively enough that a conservative matcher can defend it. Fuzzy vibes don't score.
Stability ranges, not point scores. Engines vary run to run. The same prompt can name you Monday and not Tuesday. So we report "recommended in 3 of the last 5 scans" and never a precision theater number like "AI rank: 73.4".
Row-level receipts. Every score decomposes into individual scan rows: which prompt, which engine, which date, what the answer said, what it cited. If you ask "why did my score drop?", the answer is a list of specific answers, not a shrug.
Freshness on the label. Every surface says when it last measured. A three-week-old report tells you it's three weeks old and asks to be regenerated. Stale data presented as current is just a slower way of lying.
Why be this strict in a market where competitors show bigger, smoother numbers? Because the entire product is trust. A merchant acting on our report spends real hours on the fixes. Inflated inputs would burn those hours on noise, and they'd churn the moment they noticed. Honest and smaller beats impressive and wrong, every week.
Judge it on your own store. The report is free, needs no install, and shows its work: → Run the free report