Scenario runbooks

When something breaks, the expensive mistake is not the wrong fix. It is the fast fix — because now the account has a broken listing and a contaminated measurement window.

Fact-checked 22 August 2026·24 min read·5,589 words·By Hymie Zebede
Amazon confirmed Amazon’s own documentation
Amazon research Published Amazon research
Observed Reproducible practitioner testing
Model This framework’s inference
Contents — 13 sections
  1. The diagnostic order of operations
  2. The first question in any drop: did the impressions disappear?
  3. Indexation verification: the ten-second test
  4. Runbook A — rank dropped sharply, no obvious cause
  5. Runbook B — Buy Box or featured offer lost, no notification
  6. Runbook C — listing suppressed or restricted with no obvious cause
  7. Runbook D — ads spending, nothing ranking
  8. Runbook E — incident inside a deal window
  9. Measurement discipline
  10. The experimentation protocol
  11. Constants and thresholds
  12. Confidence grading and known gaps
  13. What this means in practice

When something breaks on a listing, the expensive mistake is not the wrong fix. It is the fast fix. A number falls, somebody rewrites a title, and the account now has a broken listing and a contaminated measurement window. Two weeks later nobody can say what caused the drop, because two things changed and only one of them was recorded.

These runbooks exist to slow the first ten minutes down and speed everything after that up. Each one starts from an observable symptom, runs a fixed check sequence, and resolves to the layer where the cause usually lives rather than the layer where the symptom appeared. That distinction is the whole discipline. Every layer of the stack produces symptoms that look like they belong to the layer above it: a Buy Box loss that is really a rejected attribute contribution, an advertising problem that is really a browse-node misassignment, a traffic collapse that is really an orphaned keyword.

Three rules govern everything below. Name the layer before you act. One variable per measurement window. Baseline before change. Break any of them and the account still might recover, but you will not know why, and you will not be able to do it again.

The final section on this page lists the places where this framework is thin or wrong. That is deliberate. A methodology that cannot name its own gaps is not a methodology, it is marketing.

The diagnostic order of operations

This runs ahead of every runbook on the page and every troubleshooting procedure in the wider system. It is five steps and it takes minutes.

Establish the symptom's layer of appearance. Where was it actually observed — a dashboard number, a rejected submission, a lost placement, a support response? Write it down. Then treat it as a location of appearance, not a location of cause. These are almost never the same place.
Descend one layer and rule out. Buy Box or suppression symptom, check contribution acceptance and variation validity before you touch pricing. Advertising underperformance, check product type and item type keyword alignment and the browse node before campaign structure. Denied appeal, check for a locked field or schema validation failure that precedes human review. Traffic collapse, check indexation and attribute presence before copy. Model
Identify the owning system before choosing a channel. Escalation is a routing problem, not a persistence problem. Resubmitting at the wrong authority level, or escalating to a team with no jurisdiction over the field in question, produces identical failures indefinitely and burns the channel while it does it.
Apply the stop rule. After the second identical response through the same channel, stop. Do not send a third. Re-run step two with a wider aperture, or escalate the routing question rather than the request.
Document as a case, not a ticket. Problem, prevailing assumption, actual system behaviour, escalation sequence, resolution. Numbered and archived. This is the habit that separates operators who compound from operators who repeat.
Why step two is non-negotiable

A miscategorised ASIN makes every downstream optimisation unmeasurable, not merely suboptimal. Copy tests, image tests, bid changes and price tests run against a wrong browse node produce data that cannot be interpreted and conclusions that will be wrong in a direction you cannot predict. Verifying the data layer is not hygiene. It is the precondition for the validity of every experiment in the account.

The first question in any drop: did the impressions disappear?

Before the four-index work, before the copy conversation, before anyone opens the campaign manager, one question splits the entire investigation in two. Pull search query performance against your last archived export and look at impressions on the affected queries.

Yes — zero or near-zero

This is structural, not performance. The listing has stopped being eligible for that query. No bid change, no budget change and no copy change fixes this class of problem, and every hour spent on them is wasted.

No — impressions held, position is worse

This is performance. Decompose with the four indices and fix the stage that fell, one variable per window. Read the method on the four-index diagnosis.

When the answer is yes, work the causes in this order. It is ordered by how often each turns out to be responsible, not by how interesting it is.

1 · Your own changes
Anything published in the last 21 days, including changes nobody logged — a virtual assistant, an agency, a bulk upload someone ran and forgot. Start here every single time. Model
2 · Review Listing Changes
An AI-generated draft that published itself. Amazon holds these in a 14-day window in which an unreviewed suggestion goes live by default. Amazon confirmed If nobody in the account is watching that queue, the account has an author nobody hired.
3 · A silently rejected or reverted attribute
The submission confirmed, the value never changed, no error was returned. Re-pull the attribute 48 hours after submission and compare against what you sent. A silent overwrite presents exactly like a ranking drop. Observed
4 · Variation, product type or browse-node change
A restructure or reassignment moves the product into a different tree. Advertising keeps running flawlessly while customers browsing the correct category cannot find it.
5 · Demand disappearance
Check total market impressions on the query, not just yours. Seasonality is real and it is the one cause on this list that requires no action at all.
Candidate orphan until proven otherwise

A query at zero impressions after a structural edit is an orphaned term until you prove it is not. Confirm the market demand still exists, re-home the term according to the placement hierarchy, then verify indexation. Do not assume the term died. Assume you dropped it.

Indexation verification: the ten-second test

Three states fail separately and are fixed separately: indexed, ranked, visible. Diagnosing a ranking problem on a term that is not indexed wastes the entire investigation, and it happens constantly.

The test itself is trivial. Search the exact phrase plus the ASIN in the marketplace search bar. If the ASIN returns, the term is indexed. If nothing returns, it is not. That is the whole mechanism, and it is more reliable than any tool that claims to check indexation for you.

20 termsthe minimum sample for any indexation verification on a hero ASIN

One term proves nothing. The sample has to be built deliberately, because each slice catches a different failure mode:

5 head terms
Chosen by historical impressions. These catch eligibility loss — an attribute or product type change that quietly narrowed what the listing can be retrieved for.
5 branch terms
Spread across different mission fields — use case, audience, constraint, compatibility. These catch coverage that never existed rather than coverage that was lost.
5 newly re-homed terms
Terms this change moved from one surface to another. This is the migration failure mode: deleted from the old home, never landed in the new one.
5 backend-only terms
Terms that appear nowhere on the visible page. Non-negotiable, and explained below.

Why backend-only terms must always be in the sample

A backend search-terms field that exceeds its byte limit is rejected in its entirety — not truncated, not partially accepted — and no error is returned to the submitter. Observed The field looks populated in the interface. The listing behaves as though the field is empty, because it is.

Terms that also appear in the title or bullets cannot detect this, because they index from the visible copy regardless of whether the backend entry took. Only a term that exists solely in the backend field can prove the field saved. A head-terms-only sample will pass cleanly on a listing whose entire backend allocation was silently discarded. When that happens, trim to 249 bytes, resubmit once, wait the window and re-run the identical sample.

Respect the reindex window

24 to 48 hours for a field update; up to two weeks for a structural change. Observed Testing inside the window and resubmitting because the result looked wrong turns one change into two variables and destroys the measurement. Inside the window the correct action is to stop and schedule.

Record results verbatim, per term, with timestamp and tester — not summarised. Hold the account state, marketplace and posture constant and write down what they were, because results differ by personalisation. Re-verification always runs the identical list, never a fresh one; a new list cannot tell you whether a fix worked. Full detail on the failure modes sits in indexation troubleshooting, and the structured version of this check is one of the procedures in the operator's kit.

When a term comes back not indexed, diagnose by term type rather than guessing:

Term typeMost likely causeNext step
Newly re-homedOrphaned — deleted with no destination, or placed on a surface that did not saveCheck the term inventory, place it, re-verify. This is the migration failure mode.
Backend-onlyByte limit exceeded; the whole entry was rejected with no errorTrim to 249 bytes, resubmit once, re-verify. If still absent, escalate to a catalog check.
Long-standing head term now absentAn attribute or product type change altered eligibility, or a contribution revertedHalt. Resolve the catalog question before any copy work.
New branch term on a new ASINNever placed, or placed only inside an imagePlace it as indexed text. Images do not index as search terms.

Runbook A — rank dropped sharply, no obvious cause

Symptom. Organic position on one or several important queries falls hard inside a short window. Nothing was announced, no notification arrived, and the page looks normal.

Establish whether this is indexation or ranking. Pull search query performance against the last archived export. Queries that went to zero impressions are indexation loss. Queries with impressions and worse position are ranking loss. These have entirely different causes and the distinction takes two minutes. Everything after this step branches on it.
If indexation loss — check your own change log first. Any title, attribute or variation change in the prior 21 days? Treat it as an orphaned-term failure until proven otherwise. Re-home the terms per the placement hierarchy and re-verify with the 20-term sample above.
If no change of yours — check contribution acceptance. Did an attribute revert? Compare current live values against your last submission, field by field. A higher-scored contributor overwriting a value produces no error, no notification, and a symptom indistinguishable from a ranking drop.
If ranking loss — decompose with the four indices. Which stage fell? Impression share is findability, click-through index is the tile, cart-add index is the page, purchase index is the offer. Read the upstream-most failing index and stop there.
Read the sequence, not just the level. Conversion-driven rank loss usually shows as the purchase or cart-add index falling before impression share does. If conversion held and share fell anyway, look at competitive entry and price position rather than at your own page.
Check the confounds before concluding. Stock dip, delivery-promise change, price move, deal event, review shock, competitor event. Log all of them with dates.
Decision point

Do not ship a fix until the cause is named. A fix applied to the wrong layer does not fail neutrally — it becomes a second variable inside the measurement window, and now the drop and the fix are entangled. If you genuinely cannot name the cause, the correct action is to change nothing and collect another week of data. That feels like inaction. It is the cheapest thing on this page.

The extended version of this investigation, including the causes that get blamed most and turn out guilty least, is in rank drop diagnosis.

Runbook B — Buy Box or featured offer lost, no notification

Symptom. The featured offer is gone, or a child ASIN has become unbuyable, and no suppression notice or policy message arrived to explain it. Sessions may look normal while units collapse.

The reflex is to check price. Price is the last check in this sequence, not the first, because the silent failures live in the catalog and the commercial ones announce themselves.

Catalog layer first — was a required attribute contribution rejected? An unbuyable child with no suppression notice is a contribution failure signature. The field looked editable, the submission confirmed, the value never changed. Compare live values against what you last sent.
Variation validity. Did a compliance monitor flag an invalid grouping created weeks or months ago? Amazon evaluates the catalog with two separate systems at two different times — a standard applied at upload and a stricter monitor running continuously afterwards. Observed Accepted is not valid. The upload producing no errors is evidence of nothing.
Vendor parent. In hybrid estates, a parent created through a Vendor record cannot be dissolved or rebuilt from Seller Central regardless of Brand Registry status. If this is what you are facing, no amount of correct flat-file work will resolve it and the route is an ownership case, not a catalog fix.
Only then commercial inputs. Price competitiveness, delivery promise, defect and complaint rates. Under the current model these are graded inputs to the ranking formula rather than a pass/fail eligibility gate. Amazon confirmed That changes the shape of the fix: you are improving a score, not clearing a hurdle.
Decision point

If steps one to three come back clean and the commercial inputs are competitive, you are probably looking at a routing problem rather than a data problem. Apply the stop rule: two identical responses through the same channel and you change the channel, not the wording.

Runbook C — listing suppressed or restricted with no obvious cause

Symptom. A listing is suppressed, restricted or under review. The visible page looks compliant. The flat file shows nothing wrong. Appeals come back with the same response.

The governing assumption

Enforcement runs against catalog data, not against the visible detail page. Every appeal is judged against a backend record you may never have seen. This is why appeals fail repeatedly while the seller argues, accurately but irrelevantly, about the page.

Stop arguing about the page. Before writing a single appeal, find the attribute feeding the decision. If you cannot name a field, you do not yet have an appeal — you have a complaint.
Check foreign-language and global-selling attributes. Including regions where you have never sold a unit. Enabling global selling does not create independent regional catalogs; Amazon spins up international listings, auto-translates content, and populates attributes nobody wrote and largely cannot see. Amazon confirmed One auto-translated term in a foreign-language attribute can flag a domestic listing. Teams get split by region. The catalog never was.
Check for schema validation failure or a locked field. These precede human review. An appeal cannot succeed against a validation error — no human is reading it yet, and the case will loop until the schema problem is fixed.
Route by jurisdiction, then apply the stop rule. Identify which team owns the field in question before choosing a channel. Second identical response ends that channel. Persistence through the wrong door is the most common way weeks disappear here.
Post-reinstatement, verify inventory status separately. Reinstated and sellable are different states. A listing can be restored and still not be purchasable, and nobody will tell you.
Known limit — be honest about it

This framework holds diagnosis and triage for this class of problem. It does not hold the escalation resolutions — the exact routing, the phrasing that moves a case to a team with jurisdiction, the language that unlocks it. That material is not published anywhere and is not discoverable from the outside. Anyone selling you a guaranteed reinstatement script is describing something they cannot possess.

How exposed is your catalogue on this?

Twenty checks across the five layers, scored 0–100, returning a ranked issue list by severity — the same audit we run on client accounts. Free, no signup to see your result, and it runs entirely in your browser.

Score my listings →

Nothing you answer is transmitted or stored. The written report and the six working templates are the optional email step afterwards.

Runbook D — ads spending, nothing ranking

Symptom. Campaigns are delivering, spend is landing, impressions and clicks exist, and organic position on the funded queries does not move. Often accompanied by a request to increase the budget.

Gate check first — what is the purchase index on the funded queries? Below roughly 0.9, the spend is renting traffic rather than buying rank, and the campaign is working exactly as it will always work. Model More budget buys more of the same result. This single check resolves the majority of cases in this runbook.
Category placement. Product type and item type keyword alignment, and the resulting browse node. Perfect campaigns in the wrong tree spend correctly and reach nobody. Nothing in the advertising console will surface this.
Placement mix. What share of the spend sits in product-page placements at high cost versus top-of-search? Top-of-search is the best-converting placement in roughly 70% of brands and the best cost-per-acquisition in 40–50%. Observed A budget concentrated in the wrong placements produces this symptom while every campaign metric looks defensible.
Eligibility. Are the funded queries ones the ASIN is even attribute-eligible for organically? Paid impressions can exist where organic eligibility does not, and that mismatch produces exactly this pattern: money moving, rank frozen.
Decision point

If the purchase index on a funded query is below the gate, the correct action is to stop funding it and route the work to the offer — price, delivery promise, stock reliability. Copy cannot fix a purchase-index problem and neither can a bid. The full argument is in PPC and organic rank.

Runbook E — incident inside a deal window

Symptom. Something breaks during a deal, a promotional event or a major traffic spike. Stock, price integrity, Buy Box, a suppression, a feed failure. Everyone wants to fix everything at once.

The rule for this window

Stabilise, do not optimise. Restore availability, price integrity and the featured offer. Ship no structural change of any kind until the event is over. The window is already contaminated for measurement, so there is nothing to learn from a change made inside it — only damage to accumulate.

Stabilise the three commercial states. Availability, price integrity, featured offer. In that order. Nothing else.
Log everything with timestamps. Every action, every observation, every external event. The measurement window is lost; the log is the only thing that makes the post-event analysis salvageable at all.
Ship no structural change. No title work, no attribute edits, no variation changes, no campaign restructures. A structural change inside a deal window is a variable you will never be able to isolate.
Post-event, re-pull the trailing weeks before judging anything. Search query data reports purchases gross — cancellations and returns restate the numbers retroactively. Observed A verdict formed on same-week deal data will be wrong in a predictable direction: too optimistic.
Resume the change queue only after the event's traffic decays. Typically one full week past close. Restarting into a decaying traffic curve produces a baseline that drifts under the change.

Measurement discipline

This section exists because the industry has no shared standard for it. Nearly every case study in circulation is a success story with a headline multiple attached, no control, no window definition and no statement of what else changed in the account during the period. Procedures either bring their own evidentiary standard or they reproduce the same unfalsifiable claims with a different name on them.

The judge-by windows

WindowWhat is readable
+48 hoursPublication acceptance and attribute re-pull states. Did the change actually take?
Weeks 1–2Indexation settles. Re-check every pre-export query and run the orphan sweep.
Weeks 2–4Click-through. Tile and name effects become readable.
Weeks 4–6Impression share and family coverage. This is the verdict window. Judge the change here.
Months 2–6Assistant-layer effects — answers sourcing from the listing. Review-evidence timelines, then compounding.

The five requirements for any change you ship

A pre-export baseline. Four weeks of search query data minimum for every ASIN in scope, archived and dated, before anything is touched. Without it there is no verdict, only opinion — and someone will eventually ask for the verdict.
One variable per window. One change per product per measurement window. Two changes in one window produce an unattributable result that is worse than no test, because it manufactures false confidence that then gets applied across a catalog.
A declared judge-by date. Set before shipping, using the windows above. Never judge on day three, and never re-judge early because a number looked bad. Early re-judging is how a team talks itself out of a change that was working.
Significance floors and a confound log. At least 2,500 impressions before reading a click-through index on a single query, 100 clicks before reading cart-add or purchase, three periods before calling anything a trend. Record every deal, stock event, price change, external campaign and competitor action inside the window. An unlogged confound invalidates the result retroactively.
A rollback trigger defined in advance. The specific metric, threshold and window that reverses the change — written before shipping, not decided under pressure afterwards.

The rollback doctrine

Reverting is a stronger action than it feels like, and most reverts are made too early, for the wrong reason, and in the wrong way.

Revert only on confirmed indexation loss
And only where the lost term cannot be re-homed. Re-homing is almost always the better move: it keeps the improvement and recovers the coverage. A revert throws both away.
A week-one wobble is not a signal
That is indexation settling. A single-keyword dip is noise until the family says otherwise.
Restore the exact prior version
Byte for byte. Not a tidied version, not the prior version with one improvement kept.
A partial revert creates a third state
Neither the old listing nor the new one. It has no baseline, no comparison, and nothing about it can be interpreted. You now need a fresh four-week baseline to learn anything at all from that ASIN.
The case documentation standard

Every non-trivial problem the team resolves gets written up in one format and archived with a number: problem · prevailing assumption · actual system behaviour · escalation sequence · resolution. This is what converts solved tickets into institutional capability. Treat it as a deliverable of the work, not as documentation overhead.

The experimentation protocol

Measurement discipline compares a change against its own baseline. Where experiment tooling is available, a change can be tested against a live control instead, and that is the only method in this framework that removes the confound problem entirely. It is also available for a narrower range of questions than most people assume.

Good candidates

Hero image variants. Title formulations within the same identity root. A+ module arrangements. Bullet ordering and framing. The common shape is single-variable, reversible, and conversion-visible.

Poor candidates

Anything whose effect is on retrieval rather than conversion — attribute fills, backend changes, coverage expansion. These change who arrives, not what they do on arrival.

The interpretation trap

A conversion test cannot see a visibility gain. An attribute census that makes an ASIN eligible for six new query families will show no conversion-rate improvement whatsoever, and may show a decline as broader traffic arrives. Judge eligibility and coverage work on impression share and branch count, never on a conversion experiment. Teams have killed genuinely successful retrieval work on exactly this reading.

State the hypothesis and the failing index first. "Click-through index is 0.71; we believe the hero image is the cause." A test without a named diagnosis is a guess with a control group attached.
One variable, and enough traffic to resolve it. Low-traffic ASINs cannot resolve small effects in any reasonable window. Test on ASINs with volume, then apply the learning to the tail by pattern rather than testing each one.
Run to the tool's decision, not to a date. Stopping early because a variant looks ahead is the most common way teams manufacture false winners.
Freeze everything else on that ASIN. No price, deal, image, copy or campaign-structure change for the duration. If someone runs a promotion mid-test, the test is over and it produced nothing.
Record the loss as carefully as the win. Losing variants go into the case archive with the hypothesis that produced them. The pattern library of what does not work is the more valuable asset, and it is the one nobody keeps.

For most catalogs, most of the time, the tooling is not available. The fallback is the full measurement discipline above: pre-export baseline, one variable, declared judge-by date, significance floors, confound log, rollback trigger. The fallback is weaker, and it is what most decisions will actually rest on, which is precisely why that discipline is not optional.

Constants and thresholds

Every load-bearing number in one place, so procedures reference rather than restate. The standing column matters as much as the value. Amazon confirmed means it comes from Amazon's own documentation or announcements. Observed means reproducible practitioner testing without official confirmation. Framework standard means it is this system's own operating threshold — a decision rule, not a discovered fact. Needs verification means it rests on limited observation and should be confirmed before it enters anything client-facing.

DomainConstantValueStanding
ListingItem name / item highlights75 / 125 characters; both index; neither prioritisedAmazon confirmed
ListingReview Listing Changes window14 daysAmazon confirmed
ListingBulk upload application time~8 hoursNeeds verification
CatalogContribution scores0–100 scale; circulating figures of 30 without Brand Registry, 52 with it, 60 for a Vendor record; data augmenters and internal teams higher and unpublishedObserved — no Amazon primary source; do not quote to a client
CopyBackend allocation60% specification · 20% general · 10% own-brand equivalence · 10% complementFramework standard
CopySpecification density target≥8 concrete attributes with units per listingFramework standard
Search query dataIndex parity and action threshold1.0 is parity; act below ≈0.85Framework standard
Search query dataSignificance floors≈2,500 impressions for a click-through index · ≈100 clicks for cart-add or purchase · 3 periods for a trendFramework standard
Search query dataAttribution window~24 hours, same-dayObserved
PPCSpend eligibility gatePurchase index ≥ ~0.9Framework standard
PPCTop-of-search prevalenceBest converting in ~70% of brands; best cost-per-acquisition in 40–50%Observed
PPCTurnaround reallocation70–80% of spend to top-of-search; ~80% to hero SKUs per marketplaceTurnaround situations only
PPCNew-to-brand pause triggerBelow 10% new-to-brand on vCPM or displayObserved
PricingPrice gate+10–15% above the market purchase median, tier- and pack-normalisedFramework standard
PricingPrice-increase test5–10%, observed 14 days on BSR, conversion rate, rank and velocityObserved
PricingBusiness pricing floor5% off retailAmazon confirmed
InventoryCover policy90+ days target; never below 75; five warehouses where possibleObserved
InventoryInventory Performance Index0–1,000 scale; 450 risk threshold; quarterly checkpoints; optimise 6–8 weeks priorAmazon confirmed
InventoryPrime-badge postcode audit12-postcode panel; pass at 8 of 12 same-day or next-dayFramework standard
InventoryLong-term storage risk365 days in fulfilment centresAmazon confirmed
LaunchReview-programme stacking~30 units per marketplace on the same hero child; target 90–150 reviews; minimum 2 marketplacesObserved
DealsDeal architecture60-day runway · price discount during · best deal after (~80/20) · 4-day price discount with a 12-hour lightning deal insideObserved
DealsEligibility floorStorefront rating ≥3.5 starsAmazon confirmed
ComplianceFeatured offer modelGate-then-rank moving to rank-only, from July 2026Amazon confirmed
ComplianceAI-people image disclosure"contains synthetic performer" in the dc:subject XMP field, pre-uploadAmazon confirmed
ComplianceBusiness delivery standard90% threshold from 30 September 2026; 14-day rolling; deactivation risk from 30 OctoberAmazon confirmed
WindowsJudge-byIndexation +2 weeks · click 2–4 weeks · visibility 4–6 weeks · assistant layer 2–6 monthsFramework standard

Dated compliance items are tracked in the 2026 changelog. The index calculations behind the search-query rows are worked through in the SQP index calculator.

Confidence grading and known gaps

Nothing on this page is worth much if you cannot tell which parts are load-bearing and which are provisional. This section separates them.

What is settled — build here first

Seven findings hold across unrelated accounts, categories and mechanisms, and are treated as settled:

  • Structured data beats presentation. What the catalog believes the product is governs more than what the page says it is.
  • The title is now two co-equal indexed fields, not one field with a truncation problem. Amazon confirmed See the 75/125 title system.
  • Delivery speed is a conversion and ranking lever, not a logistics detail.
  • Escalation is a routing problem, not a persistence problem.
  • Pricing consistency is an algorithmic signal with memory.
  • An increasing share of purchase decisions is made in advance, by a system, against structured data, with the listing absent. Amazon research
  • Deal architecture is a system rather than a discount.
What counts as confirmation

A finding is only strengthened by evidence that could have contradicted it. Two observations sharing a mechanism are one observation. Agreement between two accounts in the same category, on the same tactic, in the same window, is a single data point wearing two hats — and treating it as two is how a plausible idea quietly becomes an unexamined rule.

Contested questions, and where this framework stands

QuestionThe disagreementThis framework's position
Branded PPCCut it entirely as recoverable waste, versus defend the brand's own search real estateA test protocol, not a policy. Defined window, branded-share metric, pre-declared rollback. Anyone with a universal answer here has not run the test.
How much AI search matters nowRapid growth in assistant-originated traffic, versus conventional rank still deciding most outcomesOptimise for AI retrieval because the work is identical to conventional relevance work — never as a displacement of fundamentals.
Who owns PPCOne senior operator owning the full P&L including spend, versus specialist ownershipA brand-side default does not transfer to an agency-side team. Decide ownership per engagement before writing it into a procedure.
Title change severity"Discoverability is unchanged", versus "it materially alters content architecture"Both, on different clocks. No immediate ranking change; meaningful medium-term architectural change.

Known gaps — where a procedure written today would fail

Being explicit here is not a weakness in the framework. It is the part that makes the rest of it checkable.

No escalation resolutions
We hold diagnosis and triage. The routing language and the phrasing that unlocks a stuck case are undocumented anywhere and not discoverable externally. This is the largest build item, and it can only come from a real case archive.
No failure documentation
Successes are visible everywhere and failures nowhere, including in our own material. There is no public record of where hero concentration or a branded-PPC pause did damage. Every commercial procedure needs a rollback trigger derived from real accounts, and most do not have one yet.
No cost or effort baselines
Hours per hundred ASINs for migration, census and audit work are unknown. Anyone quoting them is guessing. Measure internally on the first full cycle and version the estimates.
Thin category specificity
General guidance covers décor, apparel, pet, baby, supplements and tools reasonably well. Fitment-driven categories — where the deciding attribute is compatibility rather than preference — need their own amendments, and those are still being built.
Concepts without execution layers
Semantic bridging, external authority work and ontological structuring circulate as ideas with no procedure and no acceptance criteria attached. Where this framework offers a procedure for them, that procedure is constructed rather than observed, and it is labelled as such.
Moving targets
AI overviews, visual search, automated buying and scheduled actions were all mid-rollout during observation. Every procedure built on them carries a re-validation date, not a permanent status.
Standing instruction

When this framework and your own account data disagree, the account data wins, and the disagreement gets logged as a case. This is a prior, not an authority. It is versioned, dated, and expected to be wrong in specific places. The discipline is in noticing where, and writing it down.

What this means in practice

Five things you can do this week, in the order that pays back fastest.

Archive a baseline today, before you need it. Export four to eight weeks of weekly search query data for every hero ASIN, ASIN view and brand view, read-only and dated. Every runbook on this page begins with "compare against the last archived export." If you have no archive, the first step of every investigation is a two-week wait.
Run a 20-term indexation check on your top three ASINs. Five head terms, five branch terms, five recently moved terms, five backend-only. Twenty minutes per ASIN. Expect the backend-only slice to be where the failures are, and expect at least one hero to fail something.
Open the Review Listing Changes queue and empty it. Anything sitting unreviewed publishes itself inside the 14-day window. If nobody owns that queue, assign it before the end of the week.
Write a change log for the last 21 days. Every title, attribute, image, variation and price change, with dates and who made it — including the ones a virtual assistant or agency made without telling you. Runbook A stalls at step two without this, and most accounts cannot produce it.
Add a rollback trigger to the next change you ship. One sentence: the metric, the threshold, the window. Written before the change goes live. Then honour the judge-by date instead of checking on day three.

If your diagnosis lands on something structural across a large catalog — orphaned terms after a migration, contribution rejections nobody logged, a browse tree that stopped matching the products in it — that is systems work rather than listing work, and it is what a free account teardown is designed to surface.

Sources

Primary sources for the confirmed and published claims above. Observations and framework inferences are labelled as such in the text and are not sourced here.

  1. Amazon Search: The Joy of Ranking Products — Sorokina & Cantú-Paz, SIGIR 2016 — www.amazon.science
  2. Amazon Science — Semantic product search (KDD 2019) — www.amazon.science
  3. COSMO — SIGMOD 2024, Amazon Science — www.amazon.science
  4. Amazon — What is Amazon SEO (official guidance) — sell.amazon.com
  5. Amazon — Best Sellers Rank — sell.amazon.com
  6. Amazon — Brand Analytics and Search Query Performance — sell.amazon.com
  7. Amazon Seller Forums — 250-byte search terms announcement — sellercentral.amazon.com
  8. Amazon Seller Forums — Featured Offer eligibility update, July 2026 — sellercentral.amazon.com

If the diagnosis lands on something structural

Orphaned terms after a migration, contribution rejections nobody logged, a browse tree that stopped matching the products in it — that is systems work, and it is what the teardown is built to surface.

Get a free account teardown →
ZBD Growth

Full Amazon channel management for established brands. Catalog, advertising, inventory and cases — run by a team you know by name, reported in depth.

© 2026 ZBD Growth. Selling on Amazon since 2012.Run by Hymie Zebede →