Why It Exists
For most of Amazon's history, "visibility" meant search rank: a shopper typed a query, scanned a page of results, and compared options themselves. A product's job was to earn a spot on that page and then win the comparison on its own merits — photos, price, reviews, bullet points.
Amazon's AI shopping assistant — Alexa for Shopping (formerly Rufus) — changes what "visibility" means. Instead of a page of results, a shopper gets a single conversational answer, and that answer names a handful of products, sometimes just one. The comparison shoppers used to do themselves is now something the AI does for them, based on reading and comparing listing content.
That shift creates a gap organic rank alone can't measure: a product can rank identically to a competitor in search and still be recommended at a completely different rate once an AI is the one choosing. Recommendation Share exists to measure that specific, new thing — not to replace rank, but to answer a question rank was never designed to answer.
How Recommendation Share Is Measured
A Recommendation Share report is built from three ingredients:
- A frozen panel of purchase-intent questions — a fixed set of realistic buyer questions for the niche (for example, "which of these is best for sensitive skin?"), generated once and reused for every listing being compared, so the comparison is apples-to-apples.
- A recommendation simulation — each question is put to an AI model multiple times, alongside the listing content for the product and its named competitors. Competitor identities are anonymized and their presentation order is randomized on each pass, so the simulation can't learn to favor a position or a brand name instead of the actual content.
- A share, not a single verdict — because the simulation runs multiple passes per question, the result is a share (e.g., "won 2 of 3 passes") rather than a single win/loss, which is what makes close calls visible instead of hidden behind a coin-flip.
The output is two numbers: overall Recommendation Share (share of all questions won across all passes) and First-Rec Share — the stricter cut of questions where the product was the AI's very first pick, not just one of several mentioned.
Recommendation Share vs. Organic Rank vs. Share of Voice
These three metrics are often confused because they all describe some form of "visibility" — but they measure different things, at different levels, using different methods:
| Metric | What it measures | How it's produced |
|---|---|---|
| Organic rank | Where a listing appears in a page of search results for a keyword | Amazon's own search ranking (not disclosed to sellers) |
| Share of voice | A brand or listing's visibility across a set of keywords or ad placements, market-wide | Third-party keyword/ad tracking tools |
| Recommendation Share | Whether an AI assistant names a specific product over specific named competitors, question by question | A recommendation simulation, run by Keoxs |
A product can lead on all three, lead on one and lag on another, or lead on none — they aren't derived from each other. A listing that ranks well but reads generically to an AI comparing it against named competitors can have a low Recommendation Share despite a strong organic rank.
What Moves Recommendation Share
Four factors tend to move the needle most, based on what Keoxs's reports surface repeatedly:
- Content coverage of buyer questions — whether the listing's title, bullets, and description actually address the specific purchase questions being asked, not just contain related keywords.
- Verifiable claims — an AI comparing listings favors specific, checkable claims over vague ones; a claim with no source behind it is a claim an AI has less basis to repeat.
- Price and positioning coherence — a listing whose price doesn't match the register of its claims (a budget price with premium language, or the reverse) reads as a mismatch.
- Reviews as evidence — review content can function as corroborating evidence for a claim the listing makes, or contradict it.
None of this is a guarantee — it's a description of what correlates with a higher share across the reports Keoxs has run. See how Amazon's AI reads your reviews and what Amazon COSMO is for more on the underlying research this draws on.
The recommendation simulation is calibrated on the frameworks described in Amazon's published research — COSMO (SIGMOD 2024) and the SPN Shopping Agent paper (WSDM 2025) — not on access to Amazon's live systems. Amazon has not published an equivalent official metric.
Limitations
Recommendation Share is a calibrated simulation grounded in Amazon's published research — it is not a readout of Amazon's live systems, and it should not be read as one. A few things follow from that:
- Results are directional and comparative, not an absolute prediction of what Amazon's actual AI will say to a real shopper at a given moment.
- The simulation reflects the model and question panel used at the time of the report — it can shift when either changes, independent of any real change to a listing.
- A high share is evidence of strong, well-covered content relative to named competitors — not a guarantee of rank, recommendation, or sales on Amazon.
- The metric compares a product only against the competitors named for that report, not against the entire category.
Treating a report as a confirmed forecast of Amazon's behavior would be an overclaim. Treating it as directionally useless would ignore that it's grounded in the same published research Amazon's own AI work draws on.