Does Amazon Have a Honeymoon Period? What Amazon’s Own Research Actually Shows

Your new product launches well. Two good weeks — impressions you did not expect, a few sales a day, rank climbing. Then somewhere around week three it flattens, and by week five you are buying every sale you get.

The usual explanation is that the honeymoon ended. Amazon gave you a window, the window closed, that is how it works.

That is not what happened. Nothing ended. Something finished — a process you were entered into without being told, which ran until it had enough data to stop, and which you either passed or failed on the numbers.

The difference matters, because a window you cannot influence is something to wait out. A test is something you can prepare for.

01The belief, and the part with no evidence behind it

Ask around and you will get a fairly consistent story: a new ASIN gets 30 days, or 14, or 90 depending on who you ask, of artificially elevated placement. Amazon wants to find out whether the product sells, so it hands you free visibility. Use it or lose it.

Two parts of that are right. New products do get exposure they have not yet earned. Amazon does want to find out whether the product sells.

The part with nothing published behind it is the calendar. Nothing Amazon has released describes a timed window — no 30 days, no 14, no fixed boost that switches off on a date. What is documented is a set of statistical methods for a problem called cold start, and those methods run on data.

That reframe changes what you do. Make the most of your first month is advice you cannot act on. You are being measured from the first impression, and the measurement gets less forgiving as it goes, is advice you can.

02What Amazon’s research establishes, and where it stops

Everything in this section is documented. It comes from papers by Amazon’s own scientists, published at academic conferences, and you can read all three. What the papers do not do is describe a seller programme, promise any individual ASIN anything, or tell us what is running in production today. Those boundaries matter, and they are marked as we go.

The problem. Amazon’s search ranker leans heavily on behavioural features — how often an item gets clicked when it is shown, how often it converts. A brand-new listing has none of that. Its true values are unknown. Feed the ranker zeros and every new product is ranked as though it were a proven failure, and nothing new can ever surface. That is the cold-start problem, and it is Amazon’s problem before it is yours. A marketplace that cannot surface new products slowly stops having new products.

The 2020 paper: predict a prior. Treating cold start in product search by priors, The Web Conference 2020. Instead of zeros, predict a prior — estimate what a product like this one would probably do, using everything known that is not behaviour: category, brand, price, delivery speed, the words in the listing. Offline, those predicted priors tracked where the products eventually landed. Online, across 140 million queries, new products got more impressions and more engagement.

The 2022 paper: the mechanism in detail. Addressing Cold Start in Product Search via Empirical Bayes, CIKM 2022. This is the paper that spells out the machinery. The prior is a Beta distribution, and as real data arrives the estimate of the product’s action rate — the paper’s term, covering the engagement events the ranker uses rather than conversion specifically — is updated:

Read that as a sentence and it tells you most of what a seller needs. Your starting position is what your listing resembles. Your ending position is what your listing does. The handover is driven by how much evidence you accumulate, not by the calendar.

When n is small, alpha and beta dominate and you are ranked on resemblance — category, brand, price band, text, delivery promise. As n grows they are swamped, and you are ranked on your own results. Nobody flips a switch. The weights simply shift. The paper also gives the update cadence: version 1 refreshed every 24 hours, version 2 moved to every 2 hours.

The 2026 paper: it has not stood still. Behavioral Feature Boosting via Substitute Relationships for E-commerce Search, SIGIR 2026. A different approach to the same problem: identify substitutes — products that satisfy similar user needs — and aggregate their behavioural signals, meaning clicks, cart adds, purchases and ratings, to give a new item a warm start. Amazon states plainly that the model has been launched in production and has served customers since 2025.

It sharpens the practical advice rather than blunting it. If a new product can be warm-started from the measured behaviour of the products it most resembles, then which products you look like is not cosmetic — it is your opening balance. Category, attributes, price band and text are the inputs that decide whose track record you inherit.

03The early stop

Here is the part almost nobody discusses, and in the 2022 system it is the reason the honeymoon ends.

Exploration is expensive. Every impression given to an unproven product is an impression taken from a proven one, and Amazon does not run that at a loss indefinitely. So the paper includes a stopping rule. For each item it tracks the observed action rate and the uncertainty around it, and exploration stops when the 90% upper confidence bound falls below the average action rate of established products:

In plain English: stop showing this product when even a generous reading of its performance is worse than a typical established product.

The word doing the work is generous. Early on, with few impressions, the uncertainty term is large and forgiving — a weak rate still produces a high upper bound, so the product keeps getting shown. As impressions accumulate, that forgiveness shrinks toward nothing.

Put numbers on it. Suppose the established-product bar is 8% — Amazon does not publish the real figure, this is illustrative — and your new listing is running at 4.8%.

Nothing about that product changed between impression 50 and impression 250. It ran at 4.8% throughout. What changed is that Amazon stopped being able to tell itself the low number might be noise.

The same arithmetic run the other way is the good news. A product running at 10% never trips the rule. Its optimistic reading stays above the bar indefinitely, and there is nothing to expire.

Scope, plainly stated. That rule is what the 2022 paper describes, in the system it describes, in a test on 50 million queries. It is not established that this exact rule governs every ASIN on Amazon today. What generalises is the shape of the idea — that a new product is ranked on an estimate until its own results replace it, and that exposure producing no result is not neutral. That shape is what you plan around.

It also explains why the anecdotes on both sides of this argument sound equally convincing. I had three great weeks then it died, and new products get no special treatment they just struggle, are the same mechanism seen from either side of a stopping rule.

04Counting impressions tells you almost nothing

Everything above is written in terms of impressions and actions, and that is how Amazon’s papers describe it. It is also where most sellers go wrong, because they read their own reports the same way — as a count.

An impression is not a unit. It is a location.

Top of search is seen before the shopper has scrolled or compared anything. Rest of search is seen after the decision has largely been made. A competitor’s product page is seen by someone already reading about something else. Amazon’s own advertising console treats these as three separate placements and lets you bid up to 900% differently on each. A platform does not build separate controls, at that range, for things that are interchangeable.

And this feeds straight back into the mechanism. The estimate is fed by what actually happens, which is actions against impressions. If your impressions migrate to places where nobody clicks, your measured rate falls, and the rule reads a product people do not want.

So a blended click-through rate is not a measurement of your listing. It is a measurement of your listing mixed with wherever it happened to appear.

05How big the effect actually was

Version 1 of the 2022 system, tested across 8 million queries for a week:

That last line is the one to sit with. Nearly doubling impressions for new products made Amazon less money overall, because established products convert better and the 24-hour update cadence was too slow to correct mistakes.

So Amazon reined it in. Version 2, across 50 million queries over four weeks:

Thirteen and a half percent more impressions than a new product would otherwise get. That is the measured effect of the system Amazon shipped, reported by the people who built it.

That is the documented fact, and it requires no theory about motive. Amazon tested a much larger amount of exploration, found it reduced overall purchases, and shipped a much smaller effect.

One thing to be clear about, because it is the source of most of the confusion here. A brand-new ASIN can easily rack up an enormous impression count in its first fortnight, and sellers point at that as proof the honeymoon is real and generous. It is not the same thing. Those impressions were bought. Advertising volume is unbounded — you can have as many impressions as your budget will pay for, and most of what a budget will comfortably pay for is the cheap real estate. The exploration in the papers is organic, and it is modest.

Confusing the two is how a seller concludes they were handed a huge launch boost while their own money was quietly buying impressions in the places nobody looks.

06Advertising is rent, not a launch cost

Most launch budgets treat advertising as a temporary expense: spend hard for the first months, build rank, then taper as organic takes over.

Roughly 30% of a search results page is sponsored, and sponsored placements are auctioned continuously, which means your competitors are bidding on your keywords every day whether or not you are. Rank earned organically does not exempt you from that page; it determines where you sit among the unpaid slots that remain.

So the spend does not taper because you succeeded. It tapers when you decide to stop paying for position, and the position goes to whoever is still paying.

One consequence is worth stating, because the usual version of it is backwards: items in sponsored placements skew more expensive than the organic results around them. The simple reading, that cheap products are forced into ads, has it the wrong way round. A product needs margin to afford a bid. Products without enough margin to pay rent do not win the slots, so the slots fill with products that can.

A possible economic explanation, offered as such. Amazon has more than one revenue line from your product. In 2025 it reported roughly $172B in third-party seller services and roughly $69B in advertising services. An organic slot earns a referral fee if it converts; a sponsored slot earns a bid whether or not it converts, and the referral fee if it does. It is reasonable to wonder whether that shapes how much free exploration a marketplace is willing to fund. But nothing in the cold-start research demonstrates it, and the test results give a complete explanation on their own: more exploration made Amazon less money in total, so Amazon shipped less of it. Treat the two-revenue-lines argument as one possible reading, not as a finding. It is not needed to make the rest of this article true.

07The second loop, and it runs in days

That where you appear decides what your numbers look like has a second half, and this is where it stops being a reporting problem and becomes a mechanism.

This part is not about motive. It is about how the auction works, and it follows from something Amazon’s advertising side states openly: the highest bid does not win most Sponsored Products auctions. Something else is in the ranking, and that something is how likely Amazon expects the ad to be clicked and to convert. Roughly, what you bid multiplied by what Amazon expects you to earn it.

So advertising has its own version of the cold-start problem, and its own feedback loop, on a much shorter clock than the organic one.

The organic loop is slow and patient: gather impressions, update an estimate, stop exploring when the optimistic reading falls below the bar. That takes weeks.

The advertising loop does not wait. Expected click-through is re-estimated against what just happened. If your click-through falls, your position in the auction falls with it, at the same bid, in the same week. The slots you lose first are the expensive ones at the top. Then it compounds: your impressions reappear further down and on product pages, those click at a fraction of the rate, your measured click-through falls again, and you lose more placement.

This also predicts something specific about being new, and it follows from the same logic as the search papers. If the auction ranks on expected click-through, and a brand-new ASIN has no click history, then its expected click-through is a prior rather than a measurement. A prior is a guess about a product like yours. It will not beat an established competitor’s measured performance.

Which reframes the launch budget. The first weeks are not expensive because you are buying volume. They are expensive because you are buying the evidence that makes everything after them cheaper — and if you spend that budget on impressions that were never going to be clicked, you have bought the wrong evidence, at full price, and entered it against yourself.

08When it helps you, and when it does not

Three deserve emphasis.

The prior is the only part you control before launch. Before a single impression, the estimate of your product is built entirely from non-behavioural features. If those place your listing among products that perform, you start high. If you have miscategorised it, priced it into a band where nothing sells, or written a title the model cannot place, you start low — and you then have to climb out with real results, which is exactly the thing you have least of.

A stockout early is not a pause. Be precise about why, because the loose version of this claim is wrong. An ASIN that is not buyable may simply not be served, so it is not that Amazon records a run of failures against you. It is that the weeks when your prior is being replaced by a measurement are the weeks you cannot afford to be absent, and nothing in the system treats an absence as a successful launch. You do not come back to where you left off. You come back with the same prior and less of the runway you were spending it on.

It gets clicked comes before it converts. Most launch advice jumps to conversion rate, because that is what the seller sees on the product page. But the first filter is the tile on the results page — image, title, price, rating — and a listing that is never clicked never gets the chance to fail at conversion. Both are in the mechanism. Only one of them happens first.

09The case against

It would be dishonest to lay all this out without the counter-arguments, some of which are strong.

Amazon denies it. Amazon staff on the Seller Central forums have stated there is no official documentation or policy confirming a honeymoon period. Take that literally: there is no policy. What exists is ranking research, which is a different kind of thing, and which Amazon is under no obligation to describe as a seller benefit.

Experienced practitioners say the timed version does not exist. Danny McMillan and Oana Padurariu have said there is no evidence for a honeymoon period as commonly described, and that the system does not automatically favour new items. On the timed, automatic reading of it, they are right, and the research agrees with them.

Some say the whole thing is an illusion. Chris Turton has called it marketing hogwash, and the argument is a good one: strong early results are adequately explained by launch PPC spend, giveaways, novelty and survivorship bias — nobody writes a blog post about the launch that did nothing. Add that most buyers rely on reviews, which a new product does not have, and the case for early advantage gets weaker still.

Part of where the word comes from is a fee programme, not a ranking one. Amazon runs new-selection programmes that waive certain storage, removal and return-processing fees for new ASINs within an enrolment window. Those windows are real and they have dates on them. They are not organic ranking benefits, and conflating the two is a large part of why the X-day honeymoon language persists.

The evidence is dated in one direction and incomplete in the other. The detailed mechanism is from 2022. The 2026 paper confirms Amazon has kept working on cold start and shipped something else, but does not publish comparable lift figures. And nothing public tells us what Rufus and semantic search changed.

Some widely quoted figures are not verifiable. You will see specific numbers repeated — five days to reach the posterior, ninety days to become purchasable. Those come from agency readings of Amazon patents rather than from the papers, and the patent numbers are generally not given. Do not plan on them.

Stated as precisely as the evidence allows:

10What to do with this

Get the prior right before you launch. Correct category and browse node, accurate attributes, a brand that exists, a price inside the band where the category actually sells, a realistic delivery promise, and text that places the product unambiguously. It is the only input you control before any data exists, and it decides whose track record you start with.

Do not launch into a stockout. Over-cover a new ASIN relative to what the reorder maths would say for an established one. The cost of being out of stock in week two is not just a week of lost sales — it is a week in which the listing cannot accumulate the evidence that replaces its prior, and no system anywhere reads missing evidence as a successful launch.

Buy placement, not volume. A budget that produces a large number of cheap impressions is not a launch. Concentrate it where the impressions get looked at, and accept that this means fewer of them.

Do not buy traffic you cannot get clicked. Broad and loosely matched targeting is the usual culprit: it converts adequately on the fraction that is relevant and buries the measurement under everything else.

Budget advertising as rent, and budget the first weeks as evidence. Model the spend as a permanent line, not a launch expense that tapers. And treat the launch premium as the cost of replacing a prior with a measurement — a real thing you are buying, and one you only get to buy once per ASIN.

Price where you can afford to bid. Margin is what buys placement, and a price that leaves nothing for advertising concedes the top of the page whatever the organic ranking says. It is also worth knowing what your margin actually is by then, because the figure most sellers plan with was calculated once and never revisited.

Watch click-through in your best placement, not rank and not blended click-through. Rank is the last thing to move and the slowest to come back. Blended click-through is a placement mix and will change for reasons that have nothing to do with your listing.

11Three things this predicts

The useful test of an argument like this is whether it says anything that could turn out to be false. Three things follow, and all three are checkable in reports you already have.

One. A new ASIN pays more per click for the same slot than an established competitor does. Compare cost per click in the same placement, on the same keywords, in the same weeks, between a new listing and an established one.

Two. Most of a new ASIN’s impressions land where almost nobody clicks. Not because Amazon is hiding it, but because the top of the page is the most expensive thing on it and the newcomer is the weakest bidder there. Split a new listing’s impressions by placement and compare the split with an established one’s.

Three. When click-through falls, placement goes before rank does. Line up daily top-of-search click-through against daily top-of-search impressions and cost per click, and see which moves first.

If the third one is right, it changes what you look at every morning. A seller who tracks rank finds out last. A seller who tracks blended click-through finds out late, and misreads it when they do. A seller who tracks click-through in their best placement finds out first, while there is still something to do about it.

Frequently asked