Methodology

This archive is a measuring instrument pointed at Japan's used-car market. This page is the instrument's manual: what we record, what we compute, what we refuse to compute, and where the record is imperfect. It is written to be quoted against us.

§1 The recording

Every day, beginning at 11:00 Japan time, we read the public sitemap that Japan's largest used-car portal publishes of its live listings, and we record which cars are on it. On 2026-08-12, the first day, there were 551,333. We have recorded every day since.

We are a polite, identified visitor. We send an honest user-agent naming this archive with a contact address, we re-read the source's robots.txt on every run and obey it, we space our requests, and we ask for nothing the source does not publish for crawling. If the source ever asks us to stop, we stop, keep what we already hold, and say so here.

We add nothing and interpolate nothing. A day we failed to record stays missing, and is disclosed wherever a window crosses it (§6).

Why this archive exists. When a car sells or is withdrawn, the source deletes its page. Listings from 2022 return "掲載終了" and nothing else. There is no archive of them anywhere, and there cannot be one — the only way to hold this history is to have been watching at the time. We started watching on 12 August 2026. Everything before that date is gone for everyone, including us, and we will never claim otherwise.

§2 Coverage — what is a census, and what is a sample

This is the most important section on the page, because the two halves of the record are not equally complete and it would be easy to imply that they are.

Presence is a census. Every live listing, every day. When we say a model has 412 cars listed, that is a count of all of them, not an estimate.

Everything else is a sample. Asking price, mileage, grade, and the source's disclosure fields live on each car's individual page, and reading those pages costs the source a request each. We cap ourselves at 2,000 car pages per day, spaced at least two seconds apart — a self-imposed politeness budget, not a technical limit. Against a live pool of roughly 550,000 cars, that is a deliberate, permanent constraint.

The selection rule, stated exactly so it can be checked against our behaviour:

• Up to 500 places per day are reserved for re-reading cars we have read before, at least seven days apart, oldest reading first.
• The remaining places go to cars we have never read, newest listing first, ties broken by the source's own listing id.

The bias this creates, and which way it runs. Because newly appeared cars are read before the standing backlog, our observed set over-represents fresh listings — and a fresh listing sits at its opening asking price, before any reduction. A median computed naively from the cars we have read therefore runs above the true median of the live pool. We do not correct for this with a model. We handle it three ways: every price figure states how many cars it rests on and what share of the cohort that is; no price figure describing the live pool is published below the coverage floor in §7; and where coverage is thin we publish the statistic that the bias does not harm — the median opening asking price of newly listed cars — and label it as exactly that.

Coverage as of 2026-08-13: 6,025 cars read at least once, out of 552,591 tracked. That is 1.1%. It will rise by up to 2,000 a day. We publish the current figure on the coverage page and we do not round it in our favour.

§3 Prices

Every price on this site is an asking price. Asking is not sold. The source publishes no sold prices, and neither do we. Any site claiming Japanese sold prices drawn from this source is inventing them.

Which price. Japanese listings quote two different numbers: the vehicle-only price, and 支払総額, the total payable including registration and fees. They differ by a large and variable margin. Every figure on this site names which one it is, and the two are never mixed inside a single statistic.

Cars with no price. Some listings say 応談 — price on request. These are not cheap cars and they are not free cars; they are cars with no published price. They are excluded from every price statistic, never imputed, never treated as zero, and their share is published alongside, because a change in how many sellers withhold prices is itself a finding and averaging it away would hide it. As of 2026-08-13, 460 of 6,025 read listings — 7.6% — withheld a price.

Currency. Prices are recorded and published in yen, because that is the currency the market transacts in. Where we show a conversion, it names the rate and the date the rate was taken, and it is marked when that rate is stale. We never retrospectively re-convert a historical figure: a price recorded in August 2026 is reported at August 2026's rate, so that a movement in our index is a movement in the market and not in the exchange rate.

Price changes. Because we re-read a car at most once every seven days, we can place a price change inside the interval between two readings and no more precisely. We will say a price fell between 7 and 14 August. We will never say it fell on 11 August, because we do not know that.

As of 2026-08-13 no price-change statistic exists. Every car in the record has been read exactly once; the first re-readings become eligible on 2026-08-19. Until they accumulate, pages that would carry price histories say so plainly rather than showing an empty chart.

§4 Completion, and days listed

§4.1 What disappearance means. A listing is complete when it stops appearing on the source. Disappearance usually — not always — means sold. Cars are also withdrawn, expire, or are re-listed fresh. We can never observe a sale, and we have direct evidence that disappearance is not sale: listings sometimes come back, and a car that comes back was not sold. So we measure and name time to removal, or days listed. We never say "sold in", and no page on this site reports a sale.

§4.2 How days listed is computed. Days listed is the number of days between the first day we saw a listing and the last day we saw it. We use the last day we actually observed the car, never the day a run happened to notice it was gone — a day we failed to record would otherwise silently inflate every duration that crossed it.

§4.3 Cars that were already there — the exclusion that matters most. On 12 August 2026 we recorded 551,333 cars that were already listed. We do not know when any of them was listed. It could have been yesterday or three years ago.

Those cars are therefore excluded from every duration statistic on this site. We do not floor them at the start date, we do not estimate their age, and we do not quietly include them and hope the average survives. They are counted in supply figures, where they are accurate, and they are absent from duration figures, where they would be a fabrication.

The cost of this rule is severe and we state it rather than bury it: as of 2026-08-13, 551,333 of the 552,591 cars in the record — 99.8% — are excluded from duration statistics. Days-listed figures are built only from cars we watched arrive, beginning 13 August 2026, and they will remain thin for months. We would rather publish a thin honest number late than a thick dishonest one now.

§4.4 Cars that have not finished. A car still listed has not finished its time on the market. Duration statistics use completed listings only, and every duration figure states how many cars in the cohort are still live.

This creates the opposite bias, and it is worse the younger the archive: early on, the only listings that can have completed are the fast ones, so completed-only medians understate true duration. We handle it with a maturity gate rather than an estimator. No duration median is published until the archive is 90 days old and at least half the cohort has completed; below that, we publish the count completed, the count still live, and the longest completed duration observed — three facts, no average. Where a duration figure is affected by this truncation, the figure says so in its own caption.

§5 Days, windows, and counts

A day is a UTC day. Not Tokyo's day, not Hong Kong's, not yours. This is arbitrary but it is fixed, published, and applied everywhere, which is what a unit has to be.

Windows count days; n counts observations. A window of 30 days is thirty daily recordings. An n of 156 is a hundred and fifty-six observations, which may be of fewer cars if some were read more than once — where that distinction changes a figure's meaning, the figure states both.

The day counter shown on every page is the number of days since 12 August 2026. It is exact, it is never rounded, and its unit is always printed.

§6 Gaps, failures, and corrections

A day on which we did not complete a recording is a gap. Every gap is listed on the coverage page with its date and, once we know it, its cause. Any statistic whose window crosses a gap says so.

Two kinds of gap are possible and they are not the same:

• A missed recording. Nothing was recorded; presence and completion are both unknown for that day.
• A suppressed completion. If any part of the source's sitemap fails to load, we record the cars we did see but we deliberately refuse to mark anything as gone that day — otherwise a source-side hiccup would be recorded as several thousand cars disappearing. Presence data for that day is sound; completion data is withheld. These are disclosed separately, because a reader auditing our disappearance counts needs to know which days could not contribute.

Where the source itself is stale or unreachable, the notice says the record is stale and gives the date of the last good recording. It does not assign blame, to the source or to us, because at the moment of failure we generally do not know the cause, and guessing in public is not a thing an instrument should do.

Corrections. When we find an error we fix it, and we say so in the changelog with the date, what was wrong, and which figures moved. We do not silently restate. If a correction changes a number someone may have cited, it stays visible; corrections are not deletions.

Re-reading our own archive. We keep the original page of every car we read, exactly as it was served. Our published figures are produced by software that reads those pages, and software has bugs. When we improve it we re-run it across the entire history, so a fix reaches the past as well as the future — and the changelog records that this happened. The archived page, not our parsed output, is the authority.

§7 The sample-size law

Thin data gets honesty, not decoration. Counts always render. Everything else waits:

Two further floors follow from §2 and §4:

• Coverage floor. A price figure describing the live pool additionally requires that we have read at least 20% of that cohort's live listings. Below it, we publish the opening-asking-price figure instead, labelled as such, or nothing.
• Maturity gate. Duration medians require 90 days of archive and a majority of the cohort completed (§4.4).

Where a figure is withheld, the page says which threshold it is waiting for and, where we can compute it, roughly when it will be met. A withheld number is a statement about the record's age, not an error.

§8 What we refuse to compute

• Sold prices. We do not have them and never will from this source.
• Valuations. We will not tell you what your car is worth. We publish what other people are asking for cars like it, and the difference between those two things is the entire point.
• Rankings and verdicts. No best, no worst, no buy, no avoid. No dealer league tables. We publish counts, medians, spreads and changes, and we let them be what they are.
• Forecasts. We do not predict prices. A trend we have recorded is a description of the past.
• Causes. Our captions may compare quantities we hold or decompose a total we hold. They may not explain why a number moved, because we did not record the reason and cannot.
• Verified odometer readings. Mileage is what the listing claimed. We report it as the listing's claim and never as a verified fact.
• Accident-history judgements. We report the source's own disclosure field verbatim and we do not interpret it, grade it, or infer condition from it.

§9 Provenance and citation

Every figure on this site is traceable to a dated recording, and every page states the basis, window, sample size and full address of what it shows — deliberately, so that a screenshot of one of our charts carries its own provenance and cannot be stripped of it.

We keep the original served page for every car we read, compressed and dated. That archive is the evidence; our numbers are a reading of it.

To cite this archive:

Japan Car Data, daily record 2026-08-12 →, retrieved (date). Method v1, frozen 2026-08-13.

For research and journalism we provide extracts free — documented to the field, citable to the day, and anonymised of dealer identity. What we ask in return: cite the archive and the window, state that asking is not sold, and never present odometer readings as verified. Our arithmetic is written to be checked; if you find it wrong, tell us and we will publish the correction.

§10 Revenue and independence

We hold no inventory and we sell no cars.

Revenue will come from disclosed referral fees, framed advertising, and the letter. No revenue source can alter, reorder, suppress, or delay a number. Where we earn a referral fee, the amount is printed beside the button that earns it, together with when it is payable and any condition attached to it.

Today none of this is live. The site carries no advertising, earns no referral fee, and has no paid relationship of any kind. When the first one exists it will appear here, and beside the button that carries it, before it earns anything.

We earn the same whether a model holds its value or collapses, and whether a dealer prospers or closes. We have nothing to gain from any particular number — which is the only arrangement under which a number is worth trusting. If that ever changes, you will read it here before any figure does.

Import-rule summaries, where we publish them, are written by a person, dated, and carry their own reviewed date. We publish none yet. When we do, they will be summaries of published government rules and not advice; confirm anything that matters with the relevant authority before you spend money.

Amendments

§2a The model-count record — v1.1, added 2026-08-14

Added under the versioning rule in the header, in the same order discipline as v1: this text was written and committed before any per-model statistic was computed or published from the stream it defines. Nothing above is revised.

Each day, alongside the sitemap recording (§1), we record the source's own published tallies of live listings per maker and per model — the counts the source itself prints in its search menus beside each name. One page per maker, about 120 polite requests a day, the same identified user-agent and spacing discipline as §1.

What these numbers are. They are the source's own counts, recorded verbatim at the moment we read them. They are complete — every maker and every model the source lists, no sampling — so a per-model live count is census-grade: when a model page says how many are listed today, that is the source's count of all of them, not an estimate from our sample.

What they are not. They are not derived from our own listing record. The two instruments count the same market at different moments of the day and can disagree by a small margin — on the first recorded day the source's tallies summed to 549,928 while our sitemap census held 549,800 live listings, 0.02% apart. Every published figure states which instrument it comes from, and the two are never mixed inside a single statistic: a model page's live count cites the source's tally; the archive-wide listing total cites our own record.

Internal checks, stated so they can be audited. The source prints each maker's total and its per-model counts separately; on the first recorded day the per-model counts summed exactly to the maker totals for every maker checked. Our recorder re-verifies this identity daily and logs any divergence. A model absent from the source's menu on a given day is recorded as absent, never as zero by assumption.

What this does not change. Prices, mileage, grades and durations still come only from the sampled detail record (§2) and keep every floor and gate in §§2–7. The model-count record begins 2026-08-13 (UTC); like everything else in this archive, days before it began are not reconstructed.

§10 The owner slots, bound at launch — v1.2, added 2026-08-14

§10 was frozen on 2026-08-13 with five owner-supplied blanks left to fill before publication: the referral fee, when it is payable, any condition on it, the owner's independence statement, and the import-rule review date. This records how they were filled. No figure definition changed: §10 governs revenue disclosure alone, which sits outside the make/model/region/dealer definitions the freeze protects.

The honest binding on launch day is that no paid relationship exists — the site earns nothing, carries no advertising, and has signed no referral deal. So the fee mechanics are stated as a forward commitment rather than filled with invented figures; the disclosure that no revenue is live is stated plainly; the import-rule reviewed-date is deferred until the first summary is published; and the independence statement is bound to its final words. The binding sentence "No revenue source can alter, reorder, suppress, or delay a number" is unchanged from v1.

§1 & §4.3 The first-day baseline corrected — v1.3, 2026-08-17

This is a correction made under §6 ("Corrections"): when we find an error we fix it and say, with the date, what was wrong and which figures moved. §1 and §4.3 gave the 12 August 2026 first-day count as 549,759. The complete figure is 551,333 — the same number §4.3's own exclusion line already used ("551,333 of the 552,591 cars … 99.8%"), and the number every other page on this site has shown since launch. The figure moved 549,759 → 551,333 in both places; no definition changed.

549,759 was the count from that morning's first, partial recording — the run at 09:39 UTC on 12 August, before the day's read of the sitemap had completed. 551,333 is the complete first-day cohort: every listing already up on 12 August 2026, i.e. every listing whose first-seen date in the record is 2026-08-12. As §6 states, the archived record — not a single interim reading — is the authority. This is a supply count and the corrected figure is larger, not smaller; it is not a duration or comparative statistic, and it disadvantages no make, model, region or dealer.

§3a The price-change record — v1.4, added 2026-09-05

Added under the versioning rule in the header, in the same order discipline as v1: this text was written and committed before any per-model price-change statistic was computed or published from the re-readings it defines. Nothing above is revised. §3 committed, on 2026-08-13, that no price-change figure yet existed — every car had then been read exactly once — and that the first re-readings would become eligible on 2026-08-19. They have accumulated. This section defines the one figure we now publish from them, and holds to everything §3 already says about price.

What we compare. For one listing we compare the total payable (支払総額) we recorded the first time we read it against the total payable we recorded the last time we read it. It is the same number §3 governs — the total payable, never the vehicle-only price, and the two are never mixed inside one figure. A listing is compared only when both its first and last readings carried a real published total. A reading marked 応談 (price on request) is not a price: it is excluded at either end, never read as zero and never imputed (§3).

The three outcomes, and the share. The last total is lower than the first (we say the car is asking less), higher (asking more), or the same. We publish all three counts. The reduced share is the number asking less divided by the number we compared. We set no minimum size — a reduction counts whatever its amount — because any threshold below which a change was ignored would be a definition chosen after seeing the data, which we do not do (RULES #8). We publish the arithmetic and draw no conclusion from it.

What the figure is — and the bias it carries. This is the reduced share among cars we read more than once, and those cars are not a random sample of the market. Our re-reading rule returns to cars we have already read, oldest first (§2), so a car earns a second reading only by remaining listed long enough to be chosen again. Fast-moving cars sell or are withdrawn before we return and are under-counted; slow-moving cars linger — and a car that lingers is the more likely to be marked down. The reduced share among the cars we re-read therefore runs above the share of all listings that see a reduction, and we label it as precisely what it is — the reduced share among cars we read twice — never as the market's discount rate. Like the opening-asking median (§2) it describes a population we name rather than the live pool, so §7's coverage floor does not apply to it; its sample-size floor does.

What we do not claim. We do not call a reduction a discount and we do not attribute it to the seller. The total payable bundles the vehicle price with registration, inspection and other fees, and it can move for reasons other than a seller's decision; we record that the number changed, not why and not by whom (§8). We do not date the change: reading a car at most once every seven days, we can place a change only inside the interval between two readings and no more precisely (§3) — so this is a share over a window, never a change on a day. We publish no distribution of the sizes of reductions and no median reduction. A count and a share, and nothing beyond them.

When it publishes. The reduced share waits for the §7 threshold for a share — thirty eligible re-read cars for that model. Below it the model's page says so and names the threshold, and shows no partial figure. The record begins with the first re-readings, 2026-08-19 (UTC); the days before are not reconstructed.

§2b The Internet Archive historical instrument — v1.5, added 2026-09-05

Added under the versioning rule in the header, in the same order discipline as v1: this text is written and committed before any figure is computed or published from the instrument it defines. Nothing above is revised. No figure definition changes — this instrument adds a layer that stands apart from the make/model/region/dealer definitions the freeze protects, and, as this section binds, it may not alter any figure computed from our own record.

This archive has recorded forward since 12 August 2026 through three instruments: the presence census (§1–§2), the sampled detail record (§2), and the source's own model-count record (§2a). This section adds a fourth, and it is unlike the others in one decisive way — it points at the past, and the past is not ours.

What it is. The Internet Archive (web.archive.org) has, on dates of its own choosing, captured pages of the source for years before we began. We query the Archive's public capture index and record what those captures show: the source's per-maker and per-model counts and homepage totals as they stood when the Archive captured the menu; the sets of listing URLs present in archived copies of the source's sitemaps; and which individual listing pages the Archive holds, and from when. Where the Archive captured the source, we can read a little of the market from before 12 August 2026. Where it did not, there is nothing, and we invent nothing.

It is dated by the capture, never by us. Every figure from this instrument carries the date the Internet Archive captured the page — its capture timestamp — and not the date we read the Archive. A count from an archived menu is a fact about the source on the Archive's capture date, and it is labelled with that date and no other.

It is attributed to the Internet Archive, and it is never blended. These are the Archive's captures, not our recording, and every figure drawn from them says so. The discipline of §2a holds here with full force: no statistic mixes this instrument with our own record. Our forward series begins 12 August 2026; the Archive's captures precede it and are sparse, irregular, and outside our control. The two are shown side by side and named for what each is — never summed, averaged, spliced, or drawn as one continuous line. An Archive capture is a single dated point standing on its own.

It may not un-censor §4.3 — this is the rule that matters most in this section. §4.3 excludes the 551,333 cars already listed on 12 August 2026 from every duration statistic, because we do not know when any of them was listed. The Archive tempts us to break that exclusion: if it captured a car's page in, say, 2023, and that same car was still listed when we began, it would seem to tell us the car had been on the market since at least 2023 — and so to license a days-listed figure we have said we cannot compute. We refuse this use, without exception. An Internet Archive capture date is not a first-seen date in our record, and this instrument never supplies one. It may not date, age, or re-admit into any duration statistic a single car excluded under §4.3. It yields Archive-attributed historical facts that stand alone; it is never an input to a days-listed number.

The reason is arithmetic, not squeamishness. The Archive kept a vanishing fraction of the source's pages, chosen for its own purposes and not ours. A duration built from the few cars it happened to capture more than once would be a figure defined by which cars the Archive noticed — a definition chosen after seeing the data, which is the exact move this document exists to refuse (RULES #8). The honest §4.3 number stays thin and forward-only. The Archive does not thicken it.

Politeness and provenance. We query the Archive as the same polite, identified visitor described in §1 — an honest user-agent, spaced requests, its published interfaces only — and we keep the Archive's raw responses as our evidence, exactly as we keep the source's own served pages (§9). The Archive's record of the source is whatever the Archive chose to keep; between its captures the state of the market is unknown, and wherever a figure from this instrument appears it says so.

§2c The model-year breakdown — v1.6, added 2026-09-05

Added under the versioning rule in the header, in the same order discipline as v1: this text is written and committed before any per-model year breakdown is computed or published. Nothing above is revised. This section defines one further reading of the sampled detail record (§2); it adds no new instrument and changes no figure definition.

What it is. Every listing states its 年式 — the model year, the year the car was first registered. For the cars we read (§2, the ≤2,000-a-day detail sample), we record that stated year and, per model, we publish the count of the cars we read in each year: "of the {N} {model} we read, {n} were {year}." It is a breakdown of our reads, the same kind of figure as the grade breakdown (§7) — a count of readings, never of the market.

It is the listing's claim, and it is a sample. The model year is what the listing stated; we report it as the listing's claim and do not verify it, the discipline §8 already applies to the odometer. And it is drawn from the detail sample, not the census — so these are the years of the cars we happened to read, under §2's newest-first reading bias, and not the year profile of every car listed. The per-model live count stays the census (§2a); the year breakdown stays the sample, and the two are never mixed inside one figure.

When it publishes. A model year prints as its own row at the §7 grade-row floor — five reads — because it is that same kind of figure, a labelled count among the cars we read. Years below the floor pool into one "other years" row, so a year resting on one or two cars is never a row of its own (§7, "any statistic describing a single car: never"). The rows sum to the count of reads whose listing stated a year we could parse; reads whose year we could not read are excluded, never guessed, and their number is disclosed.

What it is not. It is not a registration-year census, not the age distribution of the market, not a valuation input, and it is never blended with the census count or with any price figure. A count of the years on the cars we read, and nothing beyond it.

Method v1 · frozen 2026-08-13 · amended v1.1 2026-08-14 (§2a, the model-count record) · amended v1.2 2026-08-14 (§10 owner slots bound at launch) · corrected v1.3 2026-08-17 (§1/§4.3 first-day baseline 549,759 → 551,333) · amended v1.4 2026-09-05 (§3a, the price-change record) · amended v1.5 2026-09-05 (§2b, the Internet Archive historical instrument) · amended v1.6 2026-09-05 (§2c, the model-year breakdown) · changes are versioned, never silent.