How a product earns a Tick, Clip, or Skip

Every number on this page is the number the software actually uses. Where an older TickClip document disagrees with what follows, this page is correct and that one is out of date.

Current methodology version: tickclip-verdict-v2. Every published verdict carries this identifier, so a verdict cited anywhere can be traced back to the rules below. See Version history.

One decision, two measurements

A TickClip verdict has a decision score from 0 to 100 and a separate confidence percentage.

The score says how good the evidence looks. Confidence says how much evidence there was. They are deliberately not combined: a 90 with 45% confidence and a 90 with 100% confidence are different situations, and averaging them into one number would hide the difference.

Amazon's customer rating stays a separate 1-to-5 marketplace value. It is never presented as TickClip's score, and it is never emitted as our own rating in structured data.

Verdict thresholds

The score maps to a decision by fixed bands:

Four rules override the bands:

• Buyer-quality floor. A product with 5 or more ratings and an average rating below 3.4 is a Skip regardless of score. A large discount on a product buyers dislike is still a bad purchase.
• Manipulation cap. When manipulation risk is assessed as high, a Tick is capped to a Clip. It is never forced to a Skip on that signal alone, because a flag is suspicious activity rather than proof.
• Thin-evidence cap. When confidence is below 50%, a Tick is capped to a Clip. Below that line at most one independent signal is behind the score, and a score of 100 built on one signal is not a reason to buy now.
• Thin-evidence floor. When confidence is below 50%, a Skip earned only by a low score is lifted to a Clip. A low score on thin evidence is not proof the product is bad. The Skip stands when the buyer-quality floor applies, or when high manipulation risk sits on top of the weak score.

Thin evidence moves a verdict toward the middle, never out of it: it can turn a Tick or a Skip into a Clip, and never a Clip into a Skip. When one of these rules changes a verdict, the verdict's own text says which one and why.

So: a product receives a Tick rather than a Clip when its evidence-weighted score reaches 75, its confidence is at least 50%, its rating is not below the buyer-quality floor, and no high manipulation risk caps it.

Scoring rules and weighting

Four dimensions, each scored 0 to 25.

Price. Where today's price sits against its own history.

With no 90-day average but a 30-day one: below 90% of it scores 22, above 105% scores 5, otherwise 15.

With neither average, but at least 30 distinct days of TickClip's own tracked prices for the listing (within the last 180 days), the median of that record stands in for the 90-day average and is scored on the same bands. This is the usual case for a listing on a store other than Amazon.

Timing. How close this moment is to the floor.

With none of those anchors, a listing with TickClip's own tracked record is scored on the same bands, with its lowest tracked price in place of the tracked low and its median in place of the 90-day average.

Reviews. What buyers reported.

Demand. Proof shoppers are actually moving it.

Weighting is by evidence, not fixed. A dimension we could not measure is given zero weight, and the weight it would have carried is shared evenly across the dimensions that were measured. The published score is the share of what was measurable, so it stays comparable between a product with four dimensions of evidence and one with two.

That is why a dimension with no evidence costs confidence, never score. Charging it to the score twice, once by scoring zero and again by lowering confidence, would punish a product for our data gaps.

An advertised discount earns no points anywhere. A struck-through list price is a claim by the seller, not a measurement.

Confidence calculation

Confidence starts at 100% and is reduced for each missing input:

When a listing is scored on TickClip's own tracked record (see Price and Timing above), that record stands in for the 90-day average and the tracked low, so those two deductions do not apply. It has no 30-day equivalent, so that deduction still does.

Confidence is a statement about coverage, not about how likely the verdict is to be right. It is also what the thin-evidence cap and floor above read: below 50%, a verdict is held at Clip rather than stated as a confident Tick or Skip.

Fair-value calculation

TickClip does not publish a fair-value range, and this is deliberate: a range implies a precision the available data does not support.

What is published instead is a baseline comparison: the current price against the product's own 90-day average, as a percentage, plus its distance from the tracked all-time low. Both are facts about that product's own trading history rather than a modelled opinion about what it ought to cost.

When no 90-day average exists but TickClip has tracked the listing itself for at least 30 days, the comparison is made against that record and is labelled as what it is: the price TickClip has typically seen over the days it has tracked the listing, and the lowest price seen in that window. It is never called a 90-day average or an all-time low. With neither, the comparison is published as Price history unavailable rather than substituted with a weaker baseline presented as the same thing.

Result ordering

Everything above decides what a single verdict says. This decides what comes first when TickClip shows several results: a deal list, a comparison, alternatives beside a verdict.

The ordering is published for the same reason the scoring is: an unexplained ranking and a paid ranking look identical from the outside. Ordering version tickclip-ranking-v1.

Discount size is not an input. There is no discount field in the ranking at all. A large percentage off tells you what a seller wants you to compare against, not whether the price is good. So it enters only through the two things that mean something: where the price sits in the product's own tracked history, and what you actually pay after any code.

Seven inputs, weighted to 100:

A Skip never outranks a Tick or a Clip. The weights alone would not guarantee this: the six non-verdict inputs total 60 against the verdict's 40, so a Skip with everything else perfect would score above a Tick with nothing else. It is enforced separately: "do not buy this" is ordered below "this is fine", whatever else is true about it. Within that line, a Tick and a Clip compete on the remaining inputs, because a Clip that is in stock at a historic low can reasonably beat a Tick that is out of stock with thin evidence.

Missing inputs score half, not zero. A product whose reviews have not been analysed is not a product with bad reviews. Scoring absent evidence as zero would quietly push down everything TickClip has simply not got to yet.

Insufficient evidence is excluded, not ranked low. A result TickClip has published no verdict for, or cannot read a price for, is reported separately with the reason, never placed at the bottom of a list, where position itself reads as a judgement. "We have not checked this" and "we checked and it is bad" are different statements and are shown as different things.

The ordering is not for sale. There is no field in the ranking for sponsorship, seller submission, or any commercial relationship. Not a field that is ignored, but one that does not exist, so there is nothing a future change could start reading. A seller-submitted offer is labelled as the seller's claim and has no effect on position. This is enforced by a test, not only by policy.

Personal preferences never touch a public ordering. Any ordering that is cached or indexed is computed with preferences absent. A personalised ordering uses the same weights but is a separate, signed-in path, and it can only move a result down. See the note on preferences in Independence.

Two inputs are weighted above but not yet supplied, and are treated as unknown rather than assumed:

• Seller trust. Sellers have an identity in TickClip but no trust score yet. Nothing in the data measures one, and reading fulfilment facts such as "sold by Amazon" as trustworthiness would both invent a number and tilt every ordering toward one retailer.
• Requirement fit. Supplied only when you have said what you need. It is absent from every public ordering by definition.

Both score as unknown, which is half of their weight, and neither is silently treated as zero.

Data inputs and source hierarchy

Inputs, in the order they are trusted:

1. Marketplace price history (via Keepa): current price, 30- and 90-day averages, tracked all-time low, list price, offer counts, Buy Box and seller observations, sales rank, monthly sales estimates. This is measured history and it is the strongest evidence we have.
2. Marketplace aggregate ratings: the rating and rating-count series. Third-party aggregate data, not our own testing.
3. Retailer-published product data: title, images, brand, category, availability.
4. TickClip's own derivations: the score, the price facts, and the decision, computed from the above.
5. Community mentions, where shown: third-party opinion, labelled as such, linked to the source, and never mixed into structured marketplace data.

Higher-numbered sources never override lower-numbered ones. A community post does not move a price comparison.

Missing-data treatment

Absence of evidence is not evidence of a problem, and the two are handled differently throughout:

• A dimension with no evidence carries zero weight, not a zero score.
• Missing values are excluded from every average, never counted as zero. A product with no rating does not drag a rating average down; it simply is not in it.
• A price of zero, a negative price or a malformed price is rejected, so it cannot move an average or produce a discount.
• A discount is not calculated at all when the reference price is missing, zero, or below the current price. A price rise is never displayed as a negative saving.
• A rate computed over too few records returns "insufficient data" rather than a number. Below the minimum sample there is no percentage to show, and 0% would be a different claim entirely.

Duplicate handling

A product is the unit, not a listing. One product sold by five sellers is one product and five listings, and the two counts are always labelled separately, and a count of products is never shown under the word "listings".

Within the catalog, rows are de-duplicated on three identities at once: the ASIN, the internal product ID, and the public verdict ID. When a re-check resolves several rows to the same product, the canonical row is kept and the duplicates are removed, so a product cannot appear twice in a count or a list.

Amazon parent/child variants resolve to the canonical parent where one exists, so colour and size variants of one product do not each become a separate verdict.

Publication checks

A verdict is computed for every product we analyse. It is published only after passing a gate. These block publication outright:

• no usable price
• no product title
• product identity unresolved
• no product image
• an implausible discount: a struck-through price more than 20% above what the product has actually been trading at
• an internal contradiction between the verdict and its own stated evidence
• the page would not be self-canonical

These are recorded but do not block publication: a non-US currency or marketplace, a variant child, thin price history, a small review sample, low confidence, a thin analysis, or stale evidence. They are quality signals worth improving, and withholding an accurate page over them costs more than it protects.

Deal and promo-code verification

Deals and promo codes are entered and reviewed by a person, not scraped and published automatically. Each carries a status and an optional expiry date.

• A deal only appears publicly while its status is active.
• A daily job expires any active deal whose expiry date has passed.
• A verdict that leaned on a promo code is invalidated the moment that code expires, regardless of how recently the verdict was computed.
• A discount percentage on a deal is calculated from the deal's own two prices, by the same single implementation that calculates it everywhere else on the site.

Update frequency and freshness windows

Every verdict has an expiry, and different facts expire at different speeds. Treating a promo code and a product's brand as equally durable is how an expired offer ends up inside a "fresh" verdict.

A verdict expires at its earliest signal expiry, not its latest. A verdict reasoned twelve days ago on a price last checked three weeks ago is out of date on the price, and taking the latest would hide exactly that.

Windows are adjusted by category and demand. Volatile categories (electronics, computers, phones, video games, cameras) get 0.6× the price-shaped windows; stable ones such as books, kitchen and tools get 1.5×. Further shortening applies when the price is at a historic low or far off its baseline, when an active deal is attached, when the seller is unknown or higher-risk, when the product is selling in volume, and when people are watching it. These can only ever shorten a window, and their combined effect is floored at 0.35× so they cannot compound into an implausibly short SLA.

Each verdict publishes the window it was given and the reasons that shortened it, so the expiry is auditable rather than asserted.

Six freshness states, shown identically on every surface: the page, the metadata, the structured data, the API and the AI tools:

What happens at expiry. The verdict, score and price recommendation are withdrawn from every surface (the page, the title, the meta description, the social card and the structured data) and replaced by "Verdict Needs Refresh", the date the verdict was last true, and the re-check state.

The page itself stays indexed and canonical. Staleness withdraws the answer; it never removes the page.

The re-check cycle. A sweep runs daily and queues verdicts a full day before they expire, so a re-check normally lands before the public state degrades. A drain runs hourly and processes the queue. A re-check that fails is retried up to three times, then parked for a person to look at, and the product's price and availability are marked unverifiable, which withdraws its recommendation rather than leaving a stale one standing.

Beyond the calendar, a verdict is also invalidated early by: a price move of 15% or more, an attached promo code expiring, the product going out of stock, a review count growing by 20% or more, or a sales rank moving by 50% or more. A price move of 5% or more marks it for a background re-check while still serving.

Corrections and rollback

Every published verdict is kept. A publication is written once and never edited, so anything citing a verdict is promised that what sits behind that citation does not change.

That makes a genuine correction possible: a previous version can be promoted back to being the live one, with a recorded reason. A rollback is itself a publication and is refused if the version being restored would not pass today's publication checks.

Verdicts whose answer changed are kept permanently, because those are the record of what we said and when. Routine re-checks that changed nothing meaningful are pruned to the most recent few per product.

Independence

TickClip takes no payment from any seller, retailer, brand or marketplace to produce, alter, rank or withhold a verdict. There is no affiliate revenue in the ranking, the score or the decision.

Sellers cannot pay to alter this verdict, score or ranking.

The full policy, including how we handle commercial relationships with data providers, is on the editorial independence page.

Limitations

Stated plainly, because a methodology that only lists its strengths is marketing:

• Amazon-first. Coverage today is Amazon products identified by ASIN. Other retailers are not yet analysed.
• Price history depends on a third party. Averages and tracked lows come from Keepa. Where Keepa has no history for a product, TickClip has none either, and the price dimension is simply not measured.
• Review integrity is inferred, not verified. We read aggregate rating and rating-count series and flag patterns in them. We do not read or semantically classify individual customer review text, and we do not claim to detect fake reviews.
• No access to returns, defect rates or warranty outcomes. These would materially improve a verdict and we do not have them.
• Manipulation flags are signals, not findings. They mean a pattern deserves caution. They are not evidence of seller intent or of illegal conduct.
• Demand estimates are estimates. Monthly sales figures published by the marketplace are approximations, and sales rank is relative to a category rather than absolute.
• A verdict is about a moment. It reflects the evidence available when it was computed. That is why every verdict carries its expiry and its check dates rather than presenting itself as timeless.

Version history

Two identifiers are versioned separately, because they answer different questions. tickclip-verdict-v* is the scoring methodology, what a verdict says, and is stamped on every verdict. tickclip-ranking-v* is the result ordering, what comes first when several results are shown, and travels with any ordered list. A change to one does not invalidate the other.

Changes to how verdicts are published or how freshness is enforced are recorded here too, dated, even when they do not change any score.

Any future change to a weight, a threshold or the confidence rules will increment the identifier and add a row, so a verdict published under older rules can always be read against the rules that produced it.