The launch sequence, day 0 to day 90

Eligibility before relevance, relevance before conversion, conversion before spend. Every failed launch we have reviewed inverted that order, usually by activating spend against a listing that was not yet eligible.

Every other procedure in this system assumes an ASIN with history. You pull four weeks of Search Query Performance, compute the indices, name the failing layer and fix it. A launch has none of that. There is no pre-export, no baseline, no market comparison built from your own funnel, and for the first fortnight there is barely enough data to justify an opinion, let alone a decision.

That absence changes what discipline means. When you cannot measure, the sequence itself becomes the control. A launch run in the right order produces evidence you can trust by day 45. A launch run in the wrong order produces three weeks of expensive noise and a listing that has already taught the system something inconvenient about itself.

What follows is the day 0 to day 90 sequence this framework uses, the failure signatures that tell you which phase actually broke, the conversion levers that produce rank at launch volumes, and the other end of the lifecycle nobody writes about: how to tell a recoverable decline from an unrecoverable one, when to stop spending, and how to retire a record without leaving a landmine in the catalogue.

It also settles the honeymoon question, because that piece of folklore is responsible for more misallocated launch budget than any other single idea in this industry.

The launch doctrine

Treat a launch as the deliberate manufacture of the four ranking terms in order: eligibility before relevance, relevance before conversion, conversion before spend. It is an engineering sequence rather than a marketing event. Model Those four terms are the spine of the six-layer stack, and at launch they have to be built rather than diagnosed.

The order matters because each term is the input to the next. Eligibility determines which queries you can be retrieved for at all. Relevance determines how deep in the result set you sit once retrievable. Conversion determines whether the position holds. Spend amplifies a conversion that already exists and cannot manufacture one that does not. Skip a rung and the rungs above it have nothing to stand on.

There is no usable public standard for launch sequencing. What circulates instead is a mixture of launch-service marketing and inherited habit. The sequence below is this framework's own, and it is the procedure listing teams ask for more often than any other.

There is no honeymoon period. There is cold-start handling

Start here, because the answer changes what you do in Phase 0.

The folklore says Amazon grants new listings an artificial visibility boost for a fixed window — thirty days, two to three weeks, or ninety days, depending on who is selling the launch package — and that generating enough sales inside that window locks in a permanent position.

Observed Amazon has never documented any of this. No announcement, help page or publication uses the term or describes a time-boxed new-product boost, and the specific figures have no traceable source. The idea survives because a closing window is an excellent thing to sell against.

Amazon research What the published research supports is ordinary cold-start handling. A new product has no behavioural history, so the ranking model has no observed values for the features it normally uses. It substitutes prior prediction values estimated from the historical performance of similar items, drawn from catalogue and new-item-setup attributes. As real interaction data accumulates, those priors are progressively replaced by posterior values based on your own impressions, clicks, cart adds and purchases. Observed Published analysis of the mechanism places the start of that transition around day five, gradual rather than expiring on a date.

That is the practical consequence, and it is why the catalogue and attribute work in Phase 0 is not administrative overhead. At launch, before any behavioural data exists, your structured data is substantially what you are being ranked on. Complete, specification-accurate attributes are the highest-leverage launch action available, they cost nothing but an hour of attention, and almost nobody frames them that way because the honeymoon story generates a different instruction entirely — drive volume before the window shuts. The full treatment of the honeymoon claim and the research behind cold start is here.

The second consequence is a caution rather than an instruction. What the system learns about you in the first weeks is what it will assume about you later. Model Launching with thin attributes, no reviews or an unstable delivery promise is expensive far beyond the launch window itself, because you are not merely performing badly, you are supplying the evidence that becomes the baseline. Get the evidence right before you get the volume up.

Phase 0 — pre-launch, before a single unit ships

Everything in this phase happens before publication. It is the cheapest work in the entire sequence and the only work that is genuinely irreversible in cost terms, because every item on it becomes exponentially more expensive once the ASIN has history attached to it.

The launch set — teaching the system what you are

The launch set is a separate object from your full query universe, and it is deliberately small. Model Its purpose is to teach the system what need you satisfy rather than to establish coverage. The first purchase cohort writes the behavioural association between your ASIN and a set of queries, and broad early traffic teaches a blurred identity that is expensive and slow to correct.

Select five to fifteen exact-match terms that meet every one of the following. Terms failing any single criterion are excluded from the launch set entirely, not weighted down.

Execution is narrow on purpose. Exact-match campaigns only for weeks one to three. Structured attributes complete before the first impression. Conversion is the goal metric, not ACOS, because purchase-after-query is the event that writes the association. Accept a high ACOS on the launch set and size it as tuition in advance, so nobody panics and pauses it in week two.

Phase 1 — days 0 to 14: indexation and evidence

The first fortnight has two jobs and neither of them is sales. Job one is confirming that the retrieval layer works. Job two is starting the evidence corpus that everything downstream depends on.

The day 2–3 check is the one teams skip, and it is the one that invalidates the rest of the quarter when it fails silently. A branch that is not retrievable is not underperforming, it is absent, and no amount of bidding fixes an absence. Indexation troubleshooting covers the diagnostic order when a branch does not come back.

The day 14 archive matters more than it looks. It is the only baseline this ASIN will ever have from a clean state, and every measurement verdict for the next year compares against it. Export it, date it, and store it where it will still be findable in nine months.

Phase 2 — days 14 to 45: conversion proof before scale

This is the phase where launches are won or quietly lost, and it is the phase with the strongest temptation to skip straight to spending.

Compute the four indices as soon as the data floors allow. Each index is your stage rate divided by the market's stage rate on the same query, computed from raw counts. Parity is approximately 1.0 and the action threshold is approximately 0.85. At launch volumes most individual queries sit below the significance floor, so read at family level and resist acting on single-query noise. The four-index diagnosis explains the construction, and the SQP index calculator does the arithmetic.

Spend stays gated. The purchase-index gate of approximately 0.9 applies at launch exactly as it does in steady state. A query you cannot yet convert does not receive rank spend. It receives page work. Below the gate, spend buys data rather than position, and it should be labelled a data-buy and capped as such. PPC and organic rank covers the gate in full.

Fix the upstream-most failing index first. A failing tile starves every layer beneath it of the data required to evaluate them, so working bottom-up wastes the window. At launch specifically, weak click-through is almost always the main image, and weak cart-add is almost always evidence density on the page. Both are cheaper to fix now than after the account has learned from them.

One variable per measurement window still applies. The temptation to change five things at once is highest at launch, and the cost is identical: an unattributable result, which manufactures false confidence and is therefore worse than no test at all.

Phase 3 — days 45 to 90: scale what converts

Only now does spend become a ranking instrument rather than a discovery instrument.

Queries crossing the 0.9 gate graduate to rank-driving spend, concentrated on top-of-search placement, which converts best in roughly seventy per cent of brands and is best on ACOS in forty to fifty per cent. Observed Build the push list from the top three to five query families scored on volume, realistic headroom and winnability. One family is one rung. Shift budget by no more than twenty points per week, and hold one variable per window.

Coverage expansion is selective. Branches that proved out get reinforced. Branches that produced impressions but no conversion get re-examined for truthfulness before being reinforced, because the most common cause of impressions without conversion at launch is a branch you should not have claimed.

Run the first deliberate review-evidence pass. The early corpus now exists, which means the objections real buyers raise are legible for the first time. Every recurring complaint gets a direct answer on the page, or it becomes the answer an assistant gives on your behalf.

The launch failure signatures

By day 30 a launch that is going wrong has usually declared itself, and the shape of the failure names the layer. Read the signature before you choose the instrument.

The first signature is the one most often misread as a marketing problem, and it is the cheapest to fix. Near-zero impressions on a branch you deliberately covered means the retrieval layer never accepted the coverage. No bid solves that.

The last signature is the expensive one, because the instinct it triggers is to spend harder. Flat performance with climbing spend means the sequence was inverted. The correct response is to stop the spend, return to Phase 0, verify eligibility branch by branch, and restart. Continuing generates below-market conversion evidence on every query you touch, which makes the eventual recovery slower than the restart would have been.

Reviews as launch physics

Review count and review text gate two things at once: conversion, and the evidence layer that assistant-generated answers are composed from. At launch you have neither, and no amount of copy compensates for their absence.

Observed The decisive tactic is marketplace-stacked Vine. The same hero child ASIN is launched across every active marketplace, roughly thirty units are sent to fulfilment in each, the ASIN is enrolled in each programme, and the reviews consolidate onto the single ASIN. The target is 90 to 150 reviews at launch across a minimum of two marketplaces.

The change in click-through and conversion during the launch window is the entire point. It is the difference between a launch that compounds and one that stalls before advertising has anything worth amplifying. Spend applied to a listing with four reviews is spend applied to a conversion rate that has not yet been allowed to exist.

There is a second-order reason to build the corpus early. The review text determines which positive product and brand aspects are even available for the system to surface, which makes the review corpus part of your advertising infrastructure rather than a passive social-proof asset. Model A thin corpus limits what can be said about you, by any surface, paid or organic.

The conversion levers that actually produce rank

Conversion is the term the ranking system watches most closely, and the levers that move it are mostly not copy. Ordered by observed effect size and speed, the offer comes first and the words come last.

Delivery promise and inventory depth

The standing policy is 90 or more days of cover, never below 75, distributed across five warehouses, with a monthly ZIP-panel audit on Tier 1 ASINs and an immediate audit on any stock dip. Model Distribution is the part teams under-plan: total units are meaningless if they are concentrated in one region, because the delivery promise a shopper sees is regional and the promise is a direct input to both conversion and featured-offer ranking.

Treat inventory planning as a ranking workstream owned jointly by the inventory lead and the brand manager. When it reports elsewhere, stock decisions get made on working-capital logic alone and the ranking cost lands on someone who was never consulted.

Price level and price history

Two distinct mechanisms, and the second is newly consequential.

The offer envelope

Three things sit around the price that most launches leave untouched.

That last point reframes what a price event is. If a segment of demand triggers on base price with no page involved, then your floor is not a discount decision, it is a switch. Know where it is before you move it.

Deals, launch and event architecture

Deal events matter to ranking because they concentrate conversion volume into a window the system reads as velocity. They only work if the ASIN enters the window already ranked. An unranked ASIN entering a peak event gets a discount and very little else.

The keyword work around an event has its own shape. From three weeks out to one week out, push outlier category terms where the price gap temporarily closes under deal pricing, and freeze all structural copy and attribute changes. Never migrate a title inside a deal window. After the event, protect the new behavioural association with maintenance spend on the families that converted, for at least two weeks. Otherwise you buy an association and then let it decay.

Diagnosing decline correctly

The other end of the lifecycle, and the one nobody writes about because it does not make a good case study. Knowing when to stop is a ranking discipline: effort and spend sustaining an unrecoverable ASIN are effort and spend not compounding somewhere else.

Decline is not one condition. Separate the four before deciding anything, because three are recoverable and one is not.

The distinction between the first and the second is where most teams go wrong, and it takes two minutes to settle. Pull Search Query Performance against your last archived export. Queries that went to zero impressions are indexation loss. Queries with impressions and a worse position are ranking loss. These have entirely different causes. Indexation loss points at your own change log first, then at contribution acceptance. Ranking loss decomposes with the four indices. Rank drop diagnosis walks the full order, and the scenario runbooks cover the incident cases.

The stop rules

Two conditions in that rule do the work, and both are routinely dropped. Three consecutive periods, because one period is a data point. And the fixes have been shipped and judged, because withdrawing spend from a query whose page problem was never addressed is not a stop rule, it is an abandonment.

Consolidation over proliferation, and how to retire a record

The most common portfolio pathology is a catalogue that has accumulated children nobody would create today. Consolidation recovers attention, concentrates the conversion signal onto fewer records and simplifies eligibility work.

The retirement checklist

Portfolio tiering and prioritisation

Tiering is the mechanism that decides where finite hours go. Without it, a forty-procedure register produces uniform shallow coverage instead of concentrated advantage, and a launch competes for attention with a legacy SKU on equal terms.

Note where a launch sits. A strategic launch under active investment is Tier 1 by definition, not by revenue, because the whole point of the ninety-day sequence is that it needs Tier 1 attention during the only window in which that attention compounds.

V × R × W — prioritising work inside a tier

Within a tier, the queue is ordered by three multiplied factors rather than by whoever shouted loudest.

Capacity rules

What this means in practice

Five things you can do this week, whether you have a launch in front of you or a decline behind you.

If a launch of yours is currently in the flat-and-spending signature, stop the spend today rather than at the end of the month. Every additional week of that pattern is another week of below-market conversion evidence attached to the queries you eventually want to win, and the unteaching takes longer than the restart. The operator's kit holds the census templates, the register format and the launch checklists in executable form; the complete guide covers the layers this sequence is manufacturing.

Sources