Search Query Performance: the four-index diagnosis

The only report Amazon gives you that shows the whole funnel, one query at a time, for your ASIN and the market side by side. It is also the report most commonly read backwards.

Search Query Performance is the only report Amazon gives a seller that shows the entire funnel, one query at a time, for your ASIN and for the whole market side by side. Amazon confirmed That second half is what turns it from a report into an instrument. An absolute click-through rate of 0.4% is uninterpretable on its own. A click-through rate of 0.4% on a query where the market runs at 1.1% is a diagnosis, and it points at one specific surface.

Everything else in the discipline is a proxy. Third-party volume estimates are models built on other models. Rank trackers tell you position without telling you what position bought you. Ad reports tell you about the traffic you paid for. SQP is the market's own arithmetic on demand that actually happened, and it is the spine of every measurement decision in the six-layer stack.

It is also the report most commonly read backwards. The export ships with percentage columns that look exactly like conversion rates and are not conversion rates. Analysts who read them as stage rates reach conclusions that are not merely imprecise but inverted, and then spend a quarter fixing the wrong layer. That single error is the most consequential measurement mistake in Amazon SEO, and it is the reason this page spends its second section on arithmetic before it gets to anything strategic.

What follows is the full measurement system: how to pull and prepare the data, the four splits you apply before reading anything, the four indices and what each one localises, the sample floors below which a number is not evidence, and the register that makes before-and-after measurement possible at all.

Why SQP is the spine

Three properties make this dataset structurally different from every other input available to a seller, and each one closes off a class of error that the rest of the industry lives inside.

Because the market columns exist, every performance question becomes a ratio rather than a judgement. You never have to decide whether a 12% cart-add rate is good. You divide it by the market's cart-add rate on the same query and read the answer.

The caveat that inverts diagnoses: the percentage columns are not conversion rates

The SQP export contains share and rate columns expressed as percentages. They are normalised against total query volume, not against your own funnel. Observed Read as click-through rate, cart-add rate or conversion rate, they do not merely mislead at the margins. They routinely point at the opposite layer from the one that is failing.

A worked illustration

The numbers below are arbitrary and chosen to make the arithmetic legible. They are not drawn from any account.

Take one query. Across the whole market it generates 100,000 impressions, 5,000 clicks, 1,000 cart adds and 500 purchases. So the market clicks at 5.0% of impressions, adds to cart at 20.0% of clicks, and purchases at 50.0% of cart adds.

Your ASIN on that same query records 4,000 impressions, 320 clicks, 70 cart adds and 30 purchases.

The share column climbs steadily from 4.0% to 7.0% and then dips slightly. Read as a funnel it looks unremarkable, maybe mildly encouraging. It suggests nothing worth acting on.

The real reading is completely different. Your tile beats the market by 60% on this query. Your page holds a small edge. Your offer is a whisker under the action threshold and worth a price check. And the actual problem, the one costing money, is that you are collecting 4% of the impressions on a query you out-click by a wide margin. This is a findability failure on a query you are demonstrably good at, and the correct instrument is indexation, attribute coverage and rank depth, not a single word of copy.

An operator reading the share columns sees a flat funnel and does nothing, or rewrites bullets because that is the default reflex. An operator computing indices from raw counts sees an under-served winner and goes to work on retrieval. Same export, opposite decision.

The two ways it inverts

Both readings are common. Both are produced by the same normalisation. The SQP index calculator on this site takes the raw counts and returns the four indices directly, if you would rather not maintain the spreadsheet.

Exports and preparation

The dataset has a minimum shape below which the method does not work.

Scope and data integrity

Four properties of this export are not documented in the interface, and each one reverses a conclusion if you miss it. Most analysts make at least two of these errors in their first month with the report.

The mandatory splits, before any index is read

Aggregate SQP hides the business. An account-level CTR index is a blend of traffic types that behave nothing like one another, and the blend is dominated by whichever type carries the volume. Four splits, applied in this order, before a single index is computed.

1 · Branded versus generic

Match brand tokens and their misspell variants, and separate the two sides completely. Destination traffic and discovery traffic are different products. In one portfolio we measured branded queries at roughly 5% of volume carrying about 58% of purchases at around 8.5 times the generic click-through rate. Observed That is one account and not a benchmark, but the shape recurs.

Why it changes the decision: a blended CTR index of 0.9 can be a branded index of 3.2 sitting on top of a generic index of 0.35. The blended number says "roughly fine". The split says the tile is failing everywhere it matters and the brand is carrying the account. Every downstream decision, from image testing to spend allocation, differs by side.

2 · Constraint-modified versus commodity

Split generic queries by whether they carry a modifier signalling a deciding constraint: material or quality, origin, size or fitment, audience, use case. A commodity query names the product. A constraint query names the product plus the thing the shopper will not compromise on.

Why it changes the decision: this is where the "our actual business is 2% of our volume" finding lives. A premium or specification-led product judged against commodity queries reads as broken on both price and click rate. Judged against the constraint queries it is built for, the same product often reads as market-leading. The split decides whether you are fixing a listing or fighting on the wrong battlefield.

3 · Head versus family

Group queries by shared root plus modifier class. Year variants, size variants, colour variants and spec variants of one intent are members of a family, not independent queries. Below-floor queries are read only as families: sum the raw counts across the family, then compute the family's indices.

Why it changes the decision: the wrong unit of analysis manufactures noise. A family member with 300 impressions and two clicks produces a CTR index that swings between 0.2 and 2.4 week to week and means nothing. The family it belongs to, with 14,000 impressions, produces a stable number you can act on.

4 · Price-normal versus price-mismatched

Compare your ASIN price to each query family's purchase median, per-unit normalised. Flag any family where the gap exceeds roughly 1.5 times. Model

Why it changes the decision: a price-mismatched family is not a copy problem and cannot be made into one. At large multiples of what a query actually buys, no image and no title recovers the click rate, because the shoppers filtering that query have already excluded your price band. The correct output is battlefield reallocation, not a listing brief.

Classifying the queries you keep

Not every row in the export is the same kind of object. Tag four classes at ingest, because each is judged by a different rule.

The four indices

Every index is constructed identically: your stage rate divided by the market's stage rate on the same query, from raw counts. Parity is approximately 1.0. Our action threshold is approximately 0.85, chosen to sit outside normal noise at the volumes we typically operate on. Model

Four numbers, four different fixes, and on any given query three of the four are wasted effort. The default industry behaviour, which is to rewrite copy whenever performance dips, is correct roughly a quarter of the time by chance.

The signature reads

Four patterns recur often enough to be recognised on sight.

Significance floors and the corroboration doctrine

No standard for this exists in circulation, so these are ours. They are working floors calibrated to our portfolio: tight enough to exclude noise, loose enough to keep most of the query set actionable. Model

The last row is different in kind from the others. A below-floor CTR index is a weak signal. A price comparison across pack architectures is not a weak signal, it is an arithmetic error dressed as a finding.

This matters most when the number is dramatic. A query where your purchase index reads 0.18 on nine clicks will get attention in any review. It should get none, unless something else in the data agrees with it.

Structural diagnosis: did the impressions disappear?

Before any four-index decomposition, ask one binary question of the affected queries. It sorts the problem into two classes that share no fixes at all.

Yes, impressions went to zero or near zero

This is structural, not performance. Something removed your eligibility. Work the causes in this order, because this is roughly their frequency order in practice. Observed

A query sitting at zero after a structural edit is a candidate orphan until proven otherwise. Confirm the market demand still exists, re-home the term according to the placement hierarchy, and then verify indexation directly rather than assuming the edit took.

No, impressions held but position is worse

This is performance, and the four-index decomposition applies. Identify the stage that fell, fix that stage and only that stage, one variable per measurement window. If rank moved without impressions moving, start with rank drop diagnosis and work the causes in frequency order.

The delta method for before-and-after measurement

Most claimed Amazon SEO results are unfalsifiable: no baseline, no window definition, no statement of what else changed in the account during the period. The delta method is the minimum procedure that produces a verdict rather than a story.

Judge-by windows and the rollback doctrine

Amazon's systems settle at different speeds at different layers, and reading a layer before it has settled is the second most common measurement error after the percentage columns. These are the windows we hold to. Model

The five requirements for any change you ship

The Search Protection Register

Every rule above depends on a baseline existing. "Archive the export" has proven too vague to be executed consistently, so the baseline is specified as an artefact with a name, a shape and a cadence.

One sheet per brand. One row per protected query or query cluster. Weekly refresh. Append-only history, meaning last week's row is never overwritten, only added to. Built before any title, copy or coverage change ships.

Intake feeds, in priority order

SQP is the spine. Everything else supplements it, in this order of trust: SQP, then the advertising Prompts report Observed which is sparse with a short lookback and should be treated as directional only, then ad search-term reports, then Brand Analytics top search terms, then Opportunity Explorer, then third-party master lists and competitor ranking sets, then mining your own detail pages for the questions shoppers ask, then review and Q&A language, then return reason codes.

What "protected" means, and how the project is judged

Protected does not mean never change. It means do not change casually. A title already performing strongly is the worst place to test speculative conversational language. Validate a pattern on the tail first, then move it to heroes family by family. The 75/125 title system covers how that sequencing works in practice.

Advanced reads

Three uses of the same export that most operators never run, each of which converts an argument into arithmetic.

Delivery-speed conversion analysis

The highest-leverage under-used report in the suite. SQP can be columned to show clicks and purchases segmented by delivery speed. Amazon confirmed That makes the conversion penalty of a slow delivery promise directly measurable on your own ASIN rather than argued from principle, which is what turns an inventory conversation into a ranking conversation.

Branded share as a spend-control instrument

Branded advertising spend is the largest reliably recoverable line in most managed accounts, and SQP supplies the only honest way to test it. Revenue is useless as the judging metric because external traffic, seasonality and deals contaminate it. Branded purchase share in SQP is the correct instrument, because it isolates the question actually being asked: when somebody searches the brand name, are you still capturing the purchase?

Run it as a defined test, never as a standing instruction.

The supporting evidence in circulation amounts to isolated accounts in which a full cut cost only a few points of branded share against material savings. Observed That is a strong argument for running the test and a weak argument for the policy. Note also the structural incentive: branded spend flatters reported ACOS, and any party compensated on ACOS has a reason to defend it. State that conflict openly, then let the test decide. More on the interaction between spend and organic position in PPC and organic rank.

Outlier and category-term positioning for event days

On major deal days shopper behaviour shifts from targeted search to broad discovery browsing, and volume floods into short category terms that are unremarkable the rest of the year. An ASIN already ranked and converting on those terms before the event captures a disproportionate share of the surge. An ASIN that starts competing during the event does not, because rank cannot be built inside a 48-hour window.

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. Observed

Delivery promise and inventory depth

The standing policy is 90 or more days of cover in the fulfilment network, never below 75, distributed across five warehouses where the programme allows. Model Low stock silently degrades the delivery promise long before it produces a stockout, and can strip the Prime badge across whole regions with no notification. The badge is tied to processing status rather than physical arrival, so units sitting in a receiving queue during peak are not eligible regardless of when they shipped. Build receiving lag into peak planning, not just transit time. Monthly ZIP-panel audit on Tier 1 ASINs, and an immediate audit on any stock dip.

Price level, and price history

Two distinct mechanisms, and the second one is newer.

Reviews as launch physics

Review count and text quality gate both conversion and the evidence layer an assistant reads. At launch the decisive tactic is marketplace-stacked Vine: the same hero child ASIN launched across every active marketplace, roughly 30 units to fulfilment per marketplace, enrolled in each, with reviews consolidating to the single ASIN. Target 90 to 150 reviews at launch across a minimum of two marketplaces. Model The change in click-through and conversion during that window is the whole point: it separates a launch that compounds from one that stalls before PPC has anything to amplify. Full sequence in the launch sequence.

The offer envelope

Returns, NCX and the negative-signal chain

The causal chains in this system run both directions, and this is the direction most agencies never model. A return costs far more than the refund line suggests, because it is a ranking event with a delay.

The chain starts with a claim the product cannot keep: a placeholder attribute value, an overstated line of copy, a wrong fitment entry. What follows is predictable and slow.

The original defect was a data-layer decision. The penalty arrives at the content and evidence layers months later, where nobody connects it back. A guessed attribute value is worse than a blank precisely because it passes gates on a promise the product does not keep, and this is where that promise gets paid for. Attribute census discipline is not compliance hygiene. It is return-rate management executed twelve weeks early.

What this means in practice

Five things you can do this week, in order.

The discipline here is unglamorous and it is the difference between a measurement system and a collection of anecdotes. Compute from raw counts, split before you read, respect the floors, judge in the declared window, and never let a single below-floor number set the agenda.

Sources