Find the motorcycle priced below its market.

Full inventory, not a niche

Every body type the dealer's own search offers — adventure, naked, sport, cruiser, classic, scooter and the rest. 1208 listings captured across 13 categories.

Mileage-aware valuation

Expected price is fitted on price vs mileage within the bike's own model and era, not a flat category average. Low mileage is rewarded, but only up to what the market actually pays.

False positives are the enemy

A deal-finder that cries wolf is worthless. Every scoring stage has a bail-out: refuse to score rather than guess, and label the confidence when it does.

Fit chain — measured, not assumed

The scorer values a bike from comparables of its own registration year first, widening to ±1, ±2 then ±4 years only when the own-year pool is too thin. That order was chosen by benchmark, and the honest reading of the benchmark is narrower than it first looks: a same-year fit scores 5.7% median error, the ±1-year chain it replaced scores 5.8%, and the shipped own-year-first estimator scores 5.9% — a tie inside the per-fold spread, not a win. What the numbers do support is that the year restriction carries the accuracy: ±3-year medians degrade to 9% and a year-blind median to 12.3%. (tools/benchmark_valuation.py re-runs it.) If no pool qualifies the scorer returns no data instead of a number — 44 bikes in this run are in that bucket, shown as such rather than quietly dropped.

Self-ad exclusion

BigMoto advertises on Auto Trader too. Any comparable matching the bike's own price, year and mileage is dropped, so a listing can never be compared against itself.

Cheap is not automatically a deal

Every score carries an 80% interval — a range the market level for that model plausibly sits in, given how many comparable listings went into it and how tightly they agree. It is not the spread of individual asking prices, which is much wider; it is how well the model knows the average. A bike is only called a deal when its price falls below the bottom of that range, so 109 bikes here are cheap but inside the model's own margin of error — labelled ? rather than sold as bargains. Where a pool is too thin to narrow the range at all, the model is saying "I don't know", and a wide interval must not read as a big discount.

Coverage by category

• Naked6,592
• Adventure6,548
• Sports Tourer1,908

Beating the 100-page cap

Auto Trader serves at most ~2,000 listings per query, so the crawl splits each body type into year bands that each stay under the ceiling. The only remaining gap is the busiest 2026 adventure band, which loses roughly the last 100 pages' worth of ads.

Dealer and private, matched like for like

This run scored 14,116 trade and 932 private ads. Private asking prices sit well below dealer ones for the same model, so each listing is valued against a pool of its own seller type wherever that pool is large enough — otherwise the comparison measures the seller, not the bike. The dashboard shows which is which on every row.

Two sources, one shell

BigMoto stock scored against the market, and Auto Trader deals scored against their own model — 1208 + 15,048 rows. The source switch sits with the scan's numbers, so it never competes with the motorcycles for the top of the page.

Four destinations

Deals to browse, Shortlist to keep, Compare to decide and Market to understand the scan. The card view is the default; the dense sortable table is one click away and remembers the choice.

Filters in two levels

The seven questions asked on every visit — make, model, category, max price, min deal %, mileage and year — stay in the open. Confidence, evidence, first seen and the four toggles fold behind More filters. On anything narrower than a laptop the whole panel becomes a sheet. Filter state lives in the URL, so a view can be bookmarked or sent to someone.

Every category, not two of them

The crawler tags twelve categories — naked, adventure, sport, classic, scooter, cruiser, sports tourer, tourer, motocross, supermoto, enduro and three-wheeler — and the page reads that list out of the data. The filter carries a count per category, the summary strip names the two largest, and the Market view has a category board showing each one's median discount and best recorded deal. Add a category to the crawler and it appears here by itself.

Search that reads a sentence

Type Triumph under £8,000 or adventure under 30,000 miles and the price and mileage clauses are lifted onto the real filters and shown as chips you can remove. Nothing is invented: a clause can only ever narrow the list, and free text still behaves exactly as it always did.

A shortlist and a comparison

Star any bike to keep it, then filter to the shortlist. Pick up to four — explicitly, or by falling back to the shortlist — and Compare puts them on one shared price axis and then in a table that marks the row that differs and the value that wins.

The deal, explained

Every bike has its own page: listed at, estimated market value, and the difference — underneath, the comparable count that produced it, the model's 80% range drawn as a distribution, the other examples of that make and model in the scan, and the price history if the snapshots can prove one.

Price history

Rows that moved since an earlier run are flagged with the change, using the dated snapshots the pipeline already writes — and saying nothing at all where the snapshots cannot prove a claim.

Export and print

Download whatever you have filtered to as CSV, or print the current view as a clean paper shortlist with no chrome.

MOT, from the record

The DVSA MOT History API now supplies each bike's real test record — expiry, defect and advisory counts, and the odometer at every test — shown as a badge on the row. The highest recorded odometer is compared with the advertised mileage, but it is evidence, not a verdict: an odometer can be replaced and a private plate moves between bikes, so a contradiction is flagged for a human and never asserted as a clocking claim.

BigMoto · ClickDealer

Plain requests gets a 403 challenge; headless Playwright Chromium clears it reliably on a fresh browser process. Re-navigating during a challenge resets it — wait on the same page instead. A page titled "Just a moment…" is a challenge; an empty page still says "Bigmoto".

Auto Trader · Vite SPA

Legacy REST endpoints 404, the GraphQL endpoint is Cloudflare-protected. The winning approach is parsing the rendered DOM by data-testid. A postcode parameter is mandatory or the page only shows a prompt, and make must accompany keywords or the search returns unrelated bikes.

The stale-DOM trap

If a navigation fails, the previous query's cards stay in the DOM and get parsed as results. Guard: assert the URL-encoded keyword appears in the final URL before trusting the cards.

Facets are case-sensitive

body-type=Adventure returns thousands of bikes; adventure returns zero. The full facet list is saved in html/facets_resp_0.json.

Shipped

• ✓Dual-source ingestion with challenge handling
• ✓Mileage-and-era-aware valuation with confidence tiers
• ✓Self-ad exclusion and false-positive guardrails
• ✓Uncertainty intervals and abstention — a cheap bike is not automatically a deal
• ✓Admin-fee-aware pricing, shown as a drive-away total
• ✓Distance to the dealer, reported but never used to filter the market
• ✓Self-contained dashboard with detail enrichment
• ✓Scheduled pipeline with price-history alerts
• ✓Data-quality gates — every crawl checked against a stored baseline
• ✓One shared band rule, so the page, dashboard and alerts cannot disagree
• ✓Published figures checked or generated rather than maintained by hand
• ✓Real MOT history per bike, through the DVSA MOT History API — the credentials were approved, retiring the gov.uk hand-off

Next

• Calibrate against real sold prices — the tooling and an external MCN band check exist, but every automated source for achieved prices is now shut, so sold_prices.csv has to be filled in by hand
• Publish the dashboard to a host so it is one click away

Closed

Investigated properly, no route exists. Recorded here so nobody spends another afternoon on them.

• ✕eBay UK sold-price signal — eBay walls its sold/completed pages behind a sign-in and captcha challenge: 15 attempts across four strategies returned zero rows while ordinary eBay pages rendered fine from the same browser. The only API with sold data is the restricted Marketplace Insights one, and automating a signed-in session past the wall would breach eBay's terms and risk the account