Tuesday morning, May 12. The seller Slacks lit up.
One of the bigger groups we sit in (300-ish private-label brands, mostly in the $1M-$30M ARR range) opened with a screenshot of a kitchen-gadget listing that had fallen from BSR #847 to #3,400 in five days. Nobody had touched it. No price change, no out-of-stocks, no policy strike. Then more screenshots came in: hair dryers, yoga mats, a couple of supplements. Different brands, different categories, same pattern. Somebody finally typed what everybody was thinking. "Did Amazon push something?"
That afternoon we pulled the panel. 17,932 listings on the ListFocal daily tracker had moved an average of 3.1 rank positions in the prior week. The baseline drift is about 1.4 positions. So yes. Amazon pushed something.
This dispatch is what we've found after three weeks of regressions. There's a TL;DR up top if you want it. The slow version, with the data and the tactics that actually adjust to the new weights, is below.
§ 01A note on what we can and can't see
We don't have Amazon's source code. Nobody outside Bellevue does. What we have is a panel of 17,932 listings sampled across six US subcategories, tracked daily since January. When something in the ranking model shifts, the listings move in patterned ways, and the patterns let us back out the (approximate) weights.
The math is dull and the assumptions are stated in the methodology section at the bottom. The short version: if a thousand listings sharing a feature all drop rank together while listings without that feature don't, the feature's weight probably moved. That's the whole game.
My listings are crashing and I haven't changed a word. Are you seeing this too?
— A panel seller, message from May 16
§ 02The panel
17,932 active listings, drawn from six US subcategories: vlogging microphones, smart light bulbs, cooking thermometers, hair dryers, electric kettles, and yoga mats. Stratified by rank within each: half from the top 100, a third from rank 100 to 500, the rest from 500 to 2000. The point of the rank stratification is to keep statistical power high without overweighting the long tail.
Daily, for each listing, we log: organic rank, inferred impressions, session conversion rate, units, review velocity, returns rate (when visible), title length, bullet length, A+ presence, twelve keyword density features, and PPC spend share. Twenty-one days times 17,932 listings gives us roughly 376,000 daily observations. Enough to do real statistics. Not enough to be certain about edge cases.
| Subcategory | Listings | Median price | Median reviews | Δ rank (21d) |
|---|---|---|---|---|
| Vlogging microphones | 2,847 | $39 | 418 | −5.2 |
| Smart light bulbs | 3,182 | $22 | 1,204 | +1.4 |
| Cooking thermometers | 2,402 | $18 | 2,108 | −3.8 |
| Hair dryers | 3,612 | $54 | 744 | +0.8 |
| Electric kettles | 2,861 | $32 | 1,612 | −2.2 |
| Yoga mats | 3,028 | $28 | 988 | −4.1 |
§ 03Eight signals, four moved
We split the ranking surface into eight signal families. Most of these are familiar from the Helium 10 / Jungle Scout school of thought. The weights we infer for them are not.
1. Title relevance. Keyword coverage in the title, weighted by the search volume of each keyword. The big lever everyone fights over. Still important. Slightly less so.
2. Backend keywords. The 250 bytes nobody enjoys filling out. Has been quietly losing weight for eighteen months. Did again.
3. Conversion velocity (CVR-7 and CVR-14). Trailing-window session conversion, normalized to category median. The thing that moved.
4. Review velocity. New verified reviews per day. The other thing that moved.
5. PPC contribution. Sessions attributable to Sponsored placements. Amazon swears this doesn't feed A10 directly. The inferred weight is small but stubbornly not-zero.
6. Image quality and count. Resolution, count, infographic presence. Modest weight, sat still.
7. A+ content. Mostly indirect. A+ lifts conversion. Conversion lifts rank. The direct A+ contribution is small.
8. Returns rate. Negative-weight. High returns drag rank, and a little more so than before.
Of the eight, four moved meaningfully. Of those four, one moved dramatically.
§ 04The weight shift, in one table
| Signal | April | May | Δ | Direction |
|---|---|---|---|---|
| CVR-14 (conversion velocity) | 0.21 | 0.32 | +0.11 | ↑ Up |
| Title relevance | 0.28 | 0.22 | −0.06 | ↓ Down |
| Review velocity | 0.16 | 0.20 | +0.04 | ↑ Up |
| PPC contribution | 0.08 | 0.06 | −0.02 | ↓ Slight down |
| Backend keywords | 0.07 | 0.07 | ±0.00 | = Flat |
| Image quality | 0.06 | 0.06 | ±0.00 | = Flat |
| A+ presence | 0.05 | 0.04 | −0.01 | = Flat-ish |
| Returns rate (inverse) | −0.09 | −0.10 | −0.01 | ↓ More punishing |
That's the jump in inferred weight on CVR-14 between the April and May regressions. It's the biggest single shift we've measured in three years of running this panel. Honestly, when we first saw it, we re-ran the regression twice because we didn't believe it.
ListFocal · May 21, 2026
The story is: relevance got cheaper, behavior got more expensive. A keyword in your title used to earn its keep just by attracting impressions. Now it has to convert those impressions too, or it eats your rank.
We also looked at how the effect splits by listing age. New listings (under 90 days) got hit harder. CVR-14 now explains roughly 41% of their rank variance, against 24% for listings older than a year. The first two weeks of a listing's life were always sensitive. They just got more sensitive.
§ 05What this actually means
Translated out of regression: your listing wins by being more relevant to the right buyer, not by being relevant to more buyers. The funnel narrowed.
Picture two versions of a title for a wireless lavalier mic. Version A says "Wireless Lavalier Microphone for iPhone Android DSLR Camera Vlog Podcast TikTok Interview YouTube Live Streaming." That title catches impressions from twelve different intents. Most of those intents do not actually buy a $39 clip-on mic. Conversion drops, CVR-14 drops, rank drops.
Version B says "Wireless Lavalier Microphone for iPhone & Android, 2.4 GHz, 65 ft Range, 2-Pack." Same product. Half the keyword surface. The shoppers who click on B actually convert on B, because B was honest about what it is.
We're seeing version-B listings climb in our panel and version-A listings sink. Not a 100% effect. More like 70-80%. There are categories where the long-tail keyword approach still works, mostly the ones where shoppers genuinely use long search strings (specialty supplements, hardware components). But for the median consumer category, the May rewrite punishes keyword greed.
§ 06Six things to actually do
These are in roughly descending order of impact. The first one is the entire game for most sellers reading this.
1. Cut keywords from your title that don't convert. Pull your Search Term Reports for the last 30 days. Find the keywords with click-to-cart conversion under 40% of your category median. Those are the keywords costing you rank. Remove them from the title and watch what happens over 14 days.
2. Front-load the first 14 days of any new listing.CVR-14 has always weighted recency. The May change made that weighting steeper. If you're launching, this is the moment to push real units through the listing (Vine, an aggressive promo, well-targeted PPC, whatever you have). Conversion in week 1 buys you more rank than conversion in week 8.
3. Rewrite bullets from features to outcomes."3,400 mAh battery" tells me a spec. "9-hour shoot day on one charge" tells me what my afternoon will look like. We A/B-tested 540 bullet swaps in the panel last quarter. The outcome-led bullets converted 2.1x better on average. Bullets that convert better lift CVR-14. CVR-14 lifts rank. This is the second-cheapest move on the list.
4. Stop using "best", "ultimate", "premium".Their indexing weight has been decaying for two years. The May rewrite accelerated the decay. They take up real estate in your title and contribute roughly zero to CTR. (Yes, we measured it. The CTR contribution of "best" is now indistinguishable from noise.)
5. Use the description to answer the single most common return reason.Description text is weakly indexed, so you're not losing much SEO value by spending those words on something practical. "Does this fit an iPhone 15 Pro Max case?" if you sell phone accessories. "Is this dishwasher safe?" if you sell kitchen. Returns just got more punishing, by about a tenth of a point. That sounds small until you remember that it's a negative-weight signal compounding daily.
6. Plan a 90-day push, not a 30-day push.Old advice was "crush the first 30 days of a launch". Our data say the CVR weighting now tapers over closer to 90 days. If you front-load for 30 and then ease off, you're leaving the slope mid-climb.
§ 07Things you can stop doing
Some of the long-running cottage-industry tactics aren't worth the effort anymore.
Backend keyword stuffing. Five minutes of effort, near-ceiling weight already. Just fill it once.
Daily title micro-tweaks.Title weight fell. Re-indexing latency didn't. Every nudge costs you a day or two of fresh signal for a marginal gain you can't measure.
Review velocity hacks where you batch-import.Amazon's review-velocity weighting is compounding. Steady arrival beats lumpy arrival. Vine campaigns spread across a quarter outperform Vine campaigns crammed into a launch window. We were surprised by that. So were our sellers.
§ 08A 30-day audit, if you want a runbook
Days 1 to 7.Pull Search Term Reports for your top 30 listings. Score each indexed keyword by click-to-cart conversion. Flag every keyword under 40% of category median. Rewrite the affected bullets from feature mode to outcome mode while you're in there.
Days 8 to 14. Push CVR on your top three SKUs. Whatever you have: a targeted promo, a Vine drop, fresh A+ for the next twenty. The goal is to move 14-day conversion meaningfully against the category median.
Days 15 to 30.Wait. Don't touch anything else. Pull weekly rank for the listings you modified versus the ones you didn't. The lag between CVR-14 movement and rank response is real. Six to ten days is normal. If CVR-14 is up and rank hasn't responded by day 24, write us.
§ 09A confession
An earlier draft of this dispatch claimed the CVR weight had jumped by 14 percentage points. That was wrong. When we re-ran the regression with a slightly tighter outlier filter, the number came in at 11. We caught it before publishing, but it's a good reminder that these inferences are not Amazon's real weights. They're a model of Amazon's model. The direction is reliable. The magnitude has a ±2 point band around it. Read the methodology section if you want the gory details.
We'll re-run the analysis in 30 days, 60 days, and 90 days. If the weights settle differently from the May snapshot, we'll write a correction.
§ 10Methodology
Panel selected by stratified sampling within each subcategory, refreshed quarterly. Daily observations from a private rank-tracker plus inferred metrics for fields Amazon doesn't expose to sellers. Title-relevance scores computed via TF-IDF against the top-50 search terms per subcategory, with intent-matching done by an LLM classifier validated on a 500-listing audit set.
Ridge regression with listing-level fixed effects, ridge penalty 0.1 chosen by 5-fold cross-validation, standard errors clustered by subcategory. We tested ten alternative specifications. The four shifted weights are stable across all of them. The exact magnitudes shift by ±2 points depending on specification.
What we cannot see: Amazon's actual model. Internal CTR adjustments. Personalization. Buy Box logic. Sponsored ranking (a separate model, with some shared features).
If you want the de-identified daily aggregates for replication, email [email protected]. We'll send them if you're doing serious work.
We're rerunning this panel monthly. The next dispatch (out Saturday) breaks down bullet patterns across 540 A/B tests. After that, we owe you the title-length analysis people have been asking for since the rewrite landed.
Cite this work. Figures licensed CC-BY-4.0. Quote any passage with attribution to ListFocal.
Field notes by email, Saturdays.
Algorithm shifts and category dynamics. Free. Unsubscribe in one click.