Implemented in CEILR
- Profit and PPC guardrails from your product costs, Amazon fees, tax treatment and target profit.
- Index checking for supplied keywords and ASINs using observed organic and sponsored visibility in Australia and the United States.
- Imported seller reports and permitted third-party CSV evidence.
- Steal, Defend, Test, Listing Only, Monitor, Avoid and Do Nothing Yet decisions with reasons and evidence.
- Loss caps, stop rules, controlled campaign planning, Waste & Harvest and proof history.
- Listing Lab for keyword cleanup, exact/close/component listing-text coverage, listing gaps, repetition review and a 249-byte backend search-term builder.
- Template-assisted Amazon AU Sponsored Products XLSX generation from a fresh user-supplied bulk template; every Create row is paused and CEILR never uploads or applies it automatically.
- Controlled AU and US keyword expansion, competitor-keyword observation, sponsored visibility, rank checks, market exploration and listing snapshots when the licensed live Amazon provider returns the required fields. CEILR shows field availability for each run.
- Opt-in scheduled Rank Tracker and Listing Monitor jobs can capture fresh provider evidence on daily or weekly schedules after the hosted scheduler is explicitly enabled; every run remains billing-gated, quota-gated and spend-capped.
- US-only DataForSEO Labs tools for ranked keywords, volume enrichment, related keywords, competitor finding, multi-ASIN keyword gaps and product rank overview, subject to provider coverage and spend controls.
- Discovery evidence can be sent into CEILR's existing profit engine so observed or database-derived keywords receive seller-specific profit-safe CPC limits and decision labels.
- P3 Product Discovery ranks sampled Amazon product observations using demand, review moat, price fit, seller economics and cash affordability; it does not invent exact monthly sales.
- Product Tracker stores timestamped price, rating, review, availability, rank and bought-past-month observations from the moment a product is tracked; missing history is never backfilled.
- The Supplier & Landed-Cost Workspace stores seller-entered suppliers and quotes, compares MOQ/lead-time/incoterm/landed-cost economics and can apply a seller-selected landed cost to the active CEILR Profit Profile.
- Competitive Market Graph, Seller Financial Command, Recovery Intelligence, Customer Voice/Product Gap, Review Request Decision Centre and AI Shopping Discoverability join available evidence without buyer PII, causality claims or Amazon mutations.
- A packaged read-only CEILR Amazon Companion detects Amazon product context and opens the authenticated CEILR analysis workflow. It does not read Amazon credentials, inject seller-state controls or copy CEILR session tokens.
- P4 Daily Decision OS condenses current CEILR evidence into a ranked seller-review queue, leads with the best-supported next move and exposes Show Working evidence, assumptions and claim boundaries. It does not execute Amazon actions.
- Launch-to-Profit Wizard and Scenario Studio model explicit seller-entered cash, inventory, price, landed-cost, conversion, CPC and spend assumptions through the current CEILR economics engine. They are planning tools, not launch or revenue forecasts.
- Proof Score summarizes only completed Proof Center grading cycles as later evidence association, never proven causality. Authenticated JSON/CSV decision reports and a stable read-only decision-brief API are available for seller reporting.
- Every money figure CEILR shows carries whether it was measured in your reports, derived from your figures, or modelled from assumptions. Two equal numbers can rest on very different evidence, and CEILR does not present them as the same kind of truth. Where an engine did not say, CEILR reports that rather than claiming the figure was measured.
- A Decision Compiler is the only surface that promotes an action to a seller. Each promoted decision carries its bounded change, its economic case, the constraint gates that were checked, its evidence, what would make it wrong, and how it will be measured. One economic exposure is counted once, even when several engines describe it.
- Workflow progress reports which stage of each workflow a product is actually on and the single next step. A stage CEILR cannot evaluate is reported as unknown rather than passed, and a stage that does not apply says so with the reason.
- Amazon Ads review worksheet, listing worksheet, repricer-floor, replenishment, support dossier and decision/outcome exports are generated from the same compiled decisions shown on screen, and stamped with the compile and evidence they came from. A decision CEILR held back never becomes an actionable row, every Create row is paused, and an export CEILR cannot stand behind is refused with the reason rather than delivered empty.
- A Prediction Ledger records what CEILR expected before the seller acted, and scores it afterwards against the seller's own later report. A prediction CEILR cannot score is reported as unscoreable rather than as a hit or a miss, and the difference between what CEILR said and what happened is shown.
- Read-only MCP access tokens let a seller connect an MCP client without handing it a CEILR session. A token is scoped to chosen tools, limited to one workspace, expires within 90 days, is revocable immediately, and is refused by every CEILR route except the MCP endpoint.
- Workspace access supports owner, analyst and viewer roles. Shared analyst/viewer access is administratively provisioned from trusted server-side Supabase metadata; it is not a browser-selected workspace or automated invitation claim.
Not claimed by CEILR
- The Index Checker does not automatically discover a competitor’s complete keyword set; use the separately bounded discovery tools for observed crawl or US database results.
- Observed crawl discovery is bounded by seed terms, call depth, result limits and provider coverage; it is not claimed to be a complete historical keyword universe.
- Listing-text coverage does not prove Amazon indexing, ranking or demand.
- Australia does not receive a third-party Labs search-volume estimate; CEILR shows Amazon's observed bought-past-month signal only when Amazon returns it.
- Competitor bid, budget, sales, revenue, units or exact targeting data are not claimed.
- CEILR does not automatically change Amazon campaigns or bids.
- Provider-backed usage is capped; it is not unlimited.
- The P3 browser companion is packaged as a read-only extension foundation and opens authenticated CEILR analysis; it does not yet claim a browser-store distribution channel or direct in-extension authenticated metric streaming.
- Supplier discovery from customs/trade records remains unavailable until a licensed provider contract, commercial rights and an explicit enable flag are present. Seller-entered supplier and landed-cost workflows remain available without it.
- A point estimate is scored on direction only. CEILR does not claim its magnitude was close, because it never stated how close counted as close, and it publishes no single accuracy figure for itself.
- Prediction reliability is reported per action class on one product, only once enough graded outcomes exist for that product. It is never presented as a general accuracy rate.
- The Proof Score is withheld until the same number of graded outcomes exists. Below it CEILR reports how many outcomes it has and lists them, rather than dividing a handful and publishing the result as a percentage about itself.
- The Amazon Ads worksheet is not an uploadable bulk file. It carries CEILR columns and a provenance header and is not built from Amazon's own template, so its rows are transferred into a fresh Sponsored Products bulk template rather than uploaded directly.
- No exported bid ever exceeds the seller's own profit-safe ceiling. An engine recommendation above that ceiling is capped, and both numbers are shown.
- Amazon's own Total line at the foot of a Sponsored Products export is rejected rather than imported, in both the search-term and targeting files, because importing it doubles every figure in the report. A row is treated as the summary only when its name reads like one and its metrics equal the sum of every other row, so a genuine keyword or target called "total" with its own numbers is kept.
- An Amazon Sponsored Products report whose cost column CEILR cannot read is reported as unreadable, not as zero spend. Amazon has used more than one name for that column; if it uses another, the totals would otherwise show 0.00 and be indistinguishable from a period in which the seller genuinely spent nothing.
- A provider call CEILR cannot price is charged against the monthly provider budget at CEILR's own estimate, never at zero, and the spend ledger records whether the figure came from the provider or stood in for it. A charged call recorded as free would consume none of the cap.
- A percentage field can be sent as a whole percent or as a decimal, and the caller declares which. A declared format CEILR does not recognise is refused rather than falling back to inferring from the magnitude — inference is what declaring a format exists to replace, and getting it wrong is a hundredfold error in the referral fee, the tax rate or the assumed conversion rate, each of which sets the exported bid ceiling.
- Decisions that name the same subject are merged into one. Decisions that describe the same evidence in different words are not merged — CEILR cannot prove from the wording alone that they are the same problem, and dropping a real recommendation is worse than repeating one — so each names the others resting on the same evidence, and the count is reported. CEILR does not claim one economic exposure is always counted once.
- The PPC action list is capped at the highest-priority hundred, and says so whenever the cap bites. The workspace table shows the top rows of that list and states how many of how many it is showing, so a seller who works the visible list to the bottom is never left believing there was nothing else.
- The Figure Basis on an exported row is always one of the four bases CEILR defines. An engine word CEILR has not vetted is recorded as unstated rather than printed into a customer file beside the vocabulary a seller has been taught to read.
- A recommendation CEILR’s own provenance graph has blocked for missing or stale evidence never reaches the seller as ready to act. The block is matched to the keyword and action class it was issued against, not to the sentence the recommendation is titled with — the profit engine gives every keyword over its ceiling the same title, so a title match landed the block on whichever keyword happened to come first and left the rest reading as ready. A verdict CEILR cannot interpret blocks rather than promotes.
- Every queue row names the keyword it is about. Two keywords over the same profit-safe ceiling are two rows, not one, because a seller cannot act on “review paid CPC” without being told on what, and collapsing them understated how many keywords were exposed.
- The US ↔ AU Transfer Radar says how the two markets were paired. An exact keyword match and an approximate one are different claims, so an approximate pairing carries the token overlap, both keywords and an instruction to confirm they describe the same product, and is never reported as high confidence — evidence completeness says how much was observed about each keyword, not whether they name the same market. Keywords are compared on every token, including the short ones: dropping them discarded exactly the model number, size, capacity or pack count that separates one market from another, so “iphone 15 case” and “iphone 16 case” matched as the same market.
- Keyword relevance is measured on every word and every character, and a hyphen is treated as the word break it is. A seller’s own keyword now matches their own listing — “omega 3 fish oil” against a title written “Omega-3” scored 66.7 out of 100 and “vitamin d3 5000iu” scored 33.3, close enough to the relevance floor that CEILR was steering sellers away from the terms that describe their product. A keyword naming a different variant no longer scores full relevance either: dropping single characters made “size 4 diapers” a perfect match for a size 6 listing. Relevance remains a lexical measure — it can say fewer of the keyword’s words appear in the title, not that 4 and 6 are competing sizes — and CEILR does not claim otherwise.
- Branded traffic is decided on whole words, by one shared rule. A brand name appearing inside an unrelated word is not the seller’s brand — “Ora” inside “storage box”, “Nest” inside “nesting bowls” — and the first word of a listing title is treated as a brand name only where the seller has recorded no brand to use instead, so a title beginning “Premium” no longer makes every keyword containing it branded. Branded share is the figure a seller reads to judge whether advertising is finding new customers or being paid for people who already knew them, and the traffic split and the keyword decision now answer it from the same definition rather than two copies of it.
- Capital is allocated in one currency, the currency of the capital the seller entered. A catalogue spanning marketplaces used to have each product’s share stamped with that product’s own currency, so a 20,000 pot became 10,000 USD to a US listing and 10,000 AUD to an Australian one, totalled as 20,000 — an exchange rate of 1.0, invented. Products reporting in another currency are now excluded, counted and named, the total carries its currency, and a mixed catalogue with no stated currency allocates nothing rather than guessing. This is the same rule the Portfolio Firewall already applied to the same catalogue: CEILR does not convert currencies it was never given a rate for.
- Days of cover on the Monitor route is counted from today, not from the day the stock was counted. A 45-day-old count of 600 units selling 20 a day read identically to one taken yesterday — 30 days of cover, healthy, no action raised — for a product that had run dry a month earlier. The count is now projected forward at the observed velocity, the counted figure and the projected figure are both reported with the assumption named, and a count carrying a date CEILR cannot place in time is not projected at all. This is the rule the replenishment file already used; both surfaces now answer the question the same way.
- The conversion rate a bid ceiling is computed at is decided by one rule, wherever the ceiling is shown. PPC Waste & Harvest used to accept three orders on ten clicks as a measurement and raise the ceiling to 2.30 where the keyword analysis on the same report gave 0.61; a row carrying more orders than clicks — which Amazon’s attribution window makes legitimate at a period boundary — raised it to 7.67, offered as “review for controlled Exact promotion at or below 7.67”. A rate is a measurement only above CEILR’s own calibration threshold; below it a sample may argue the ceiling down but never up. A measured rate still moves the ceiling in either direction.
- No orders yet is an absence, not a conversion rate of zero. A single click at a seller’s own 8% assumption is expected to produce 0.08 orders, so seeing none is the ordinary outcome — but it was read as a measured rate of zero, and because the ceiling scales with the rate, one click with no order set the bid ceiling to zero. Every keyword in a new advertiser’s first campaign was zeroed that way. Above CEILR’s calibration threshold a zero rate is a real finding and is kept. A keyword genuinely burning money is still caught where it should be — on spend against the seller’s own maximum-loss cap — because that rule weighs what was actually spent.
- When CEILR scores its own recommendation, the periods either side of the action are costed from the revenue the seller’s report actually observed. The proof log used to price every order at the profile’s selling price and ignore the sales it had already totalled, so it silently assumed the seller sold at list: across a period carrying a 20% promotion it reported contribution up 290.60 and graded the recommendation a win, where the observed revenue makes it 634.25 down. A period CEILR cannot cost — no economics, or a report booking revenue against zero orders — is left unquantified and the grading says insufficient evidence rather than reporting a result.
- A seller report set aside because it overlaps another is named, with its window and the reason. Amazon lets a seller download any date range, so a 30-day pull and a 7-day pull ending the same day cannot be added together without counting the shared week twice; CEILR keeps one report from each run of overlapping windows. That was already true and already correct, but it happened in silence — 240 clicks became 56, the conversion rate stopped being a measurement, and the analysis named only the report that survived. Which report to keep is a genuine trade-off between the most recent import and the widest window, and CEILR has no evidence to settle it, so the rule is unchanged; what changed is that the seller can now see what was dropped and re-import one report covering the whole period.
- A report costed at economics that were not in force during it says so. A seller’s landed cost changes with each shipment, so Profit Profiles carry an effective date, and every surface costs a report at the profile in force today — right for “what may I bid next”, and not what a seller reads when a July report tells them what July earned. Which profile to compute with is a real choice CEILR has no evidence to settle, so the selection is unchanged; what changed is that a period lying wholly or partly before the profile’s effective date is named, with both dates, on the keyword analysis and on PPC Waste & Harvest. Waste & Harvest had been printing the report period and the profile’s effective date beside each other without ever comparing them.
- A constraint CEILR could not evaluate is never presented as one that passed. Nothing is blocked on an unevaluated check — a workspace that has not recorded its stock would otherwise never receive a recommendation — but the Ads worksheet and the support dossier both name which checks did not run and say they are not failing, they have not run.
- That includes the cash check on an action that spends capital. A replenishment order or capital allocation carrying no stated cost reported the cash check as passed — for a workspace that had not recorded a cent of deployable cash — and never appeared among the checks named as not run, because that list names unknowns and this was a pass. An action that needs no capital, such as a bid reduction, still passes it. Nothing blocks on an unknown either way: CEILR does not invent a constraint it was never given, it says which checks it could not make.
- An observation dated further ahead than ordinary timezone skew is treated as evidence CEILR cannot place in time, not as evidence observed today, on every surface that reports freshness. The evidence-quality gate accepts only the freshness states it knows; any other value is reported as unknown rather than passed, so a state CEILR cannot interpret never reads as evidence it has checked.
- A row reporting more clicks than impressions is rejected at import, because no attribution lag can explain it. A row reporting more orders than clicks is kept — Amazon's attribution window lets a click from the previous reporting period land an order in this one — but its implied conversion rate above 100% is never treated as a measurement, and the ceiling falls back to the rate the seller assumed.
- A search-term row reporting revenue with zero orders is rejected at import and named, rather than carried. Every unit cost CEILR knows is charged per order, so such a row would be costed as pure margin and shown as a profit on a keyword that sold nothing. Where one already exists in a workspace, its contribution is reported as not quantified rather than as a win.
- The same is true in reverse: a row reporting orders with zero revenue is rejected and named. Contribution falls back to the product’s selling price when a report gives no sales, so such a row booked revenue the report denied. This is the more dangerous half, because a Sales column mapped to nothing makes every row look that way while the conversion rate still computes from orders and clicks — an entire analysis resting on assumed revenue with nothing about it appearing thin. The Targeting Report importer is deliberately different: its metrics are optional there, so a missing column stays absent instead of becoming a zero.
- Days of cover and the stockout date count only stock on hand. CEILR is not told when inbound units land, so it will not report a container that has not shipped as days of cover; the inbound figure and what it would mean are shown separately, and it is still subtracted from the suggested order so the seller does not order it twice.
- The replenishment order quantity runs from the stock projected to be left today, not from the day the seller counted it, and the file carries the counted figure, the projected figure and the assumption the projection rests on. A count old enough to have run out says so rather than projecting negative stock.
- The exported price floor clears cost, Amazon fees and the seller's stated target profit before advertising. Advertising is not in it. The file states the observed advertising cost per attributed order beside the floor rather than folding it in, because a search-term report does not say what share of the seller's orders were advertised.
- A conversion rate is treated as measured only once a keyword has enough clicks for the number to mean something. Below that, the rate the seller assumed is used instead: a thin sample may still pull a ceiling down, but it is never allowed to push one up, because a single click that converted reads as a 100% conversion rate.
- That ceiling moves with the conversion rate it was computed at, so every exported bid names the rate behind it and whether it was measured or assumed. Where the seller's own reports show a lower conversion rate than they assumed, the bid is held to what the measurement supports; where they show a higher one, CEILR keeps the lower ceiling and states the higher figure rather than raising a bid on the seller's behalf.
- A bid change is not exported as a bulk row. Amazon expresses it as an update to a campaign structure CEILR does not hold, and writing it as a new negative would turn a bid reduction into a shutdown.
- CEILR does not schedule stop-losses or rollbacks on the seller's behalf. Safety is a seller commitment, and CEILR will not invent a risk appetite it was never told.
- US Labs results are provider database observations, while AU discovery is a bounded live crawl rather than a complete historical database. CEILR does not claim a 475M-product proprietary catalogue or decade of historical sales estimates.
Marketplace truth
- Australia and United States are the validated live CEILR marketplaces for observed provider workflows. CA, UK, DE, FR, IT, ES, MX, JP and IN adapters are defined but remain visibly gated until provider coverage, Amazon roles, currency/tax semantics and launch tests are completed.
- DataForSEO Labs tools are intentionally restricted to United States/English products before provider spend.
- Observed Amazon visibility, sponsored presence, listing details and bought-past-month signals are timestamped and kept separate from provider database search-volume estimates.
- Raw provider responses are not exposed or retained; CEILR stores response hashes, safe field-presence summaries and normalized evidence rows.
Current plan limits
- One workspace owner, optional administratively provisioned analyst/viewer access, and up to 50 active products.
- An active product is one sellable listing identity in one marketplace. A variation child with its own SKU and economics counts separately, as does the same SKU in another marketplace. Archived products keep their full history and evidence, do not occupy an active slot, and are not monitored.
- Up to 1,000 charged live keyword observations per calendar month.
- Up to 50 keywords and 10 competitor ASINs per scan.
- Cached reuse does not consume the charged-keyword allowance.
- Up to 25 charged competitor-discovery or keyword-expansion jobs per calendar month. Market Explorer searches use the live-keyword allowance instead of the discovery allowance; Listing Monitor snapshots use provider spend only.
- A US$5 monthly provider-spend cap applies per workspace, with per-job and global safeguards. Deterministic analysis across all 50 active products is not limited by the provider allowance.
- Reaching the 50-product limit offers archive-or-replace. CEILR never deletes evidence, never silently stops monitoring another product, and never bills an overage.
Decision control: CEILR may recommend that the evidence is too weak and that no action should be taken yet. It never guarantees rankings, sales, advertising performance or profit.