# AI transparency

> Every AI feature in Cyclesite, what it does, what it is grounded in, how accurate it is today, and where it is known to be weak.

**Published by:** Cyclesite. Last updated 2026-09-21.

Review cadence: monthly, or on any AI feature change. If a feature stops working as described, report it at https://www.cyclesite.co.uk/contact — if the limitations change, this document is updated in the same release.

## Used-price aggregates (Answer Hub, brand pages, city pages)

**What it does:** Shows typical, low and high prices for used bikes over the last 12 months, filtered by brand, model, city or category. These are last-advertised prices of sold and delisted listings plus live listings. They reflect what bikes are advertised for across the UK used-bike market, not confirmed sale prices.

**What it is grounded in:** Direct SQL aggregates over measured UK used-bike listing prices. No model inference. Minimum sample size is 10 rows; below that the page shows "not enough data yet" instead of a number.

**Accuracy today:** Within-sample arithmetic is exact. Representativeness depends on the freshness and breadth of the measured used-price data. A confidence badge is visible when the underlying sample is below 30.

**Known limitations:** These are asking prices, not confirmed transaction prices. A majority of the corpus carries no model-year, so year-filtered figures draw on a smaller and less certain pool than the headline count suggests. Early-launch data is thin in some categories. Outlier trimming is minimal. We are not yet showing full distributions (p25/p75) everywhere, only typical values with confidence tiers.

## Valuation engine (per-listing fair-market estimate)

**What it does:** Given a listing's brand, model, year, condition, and frame size, estimates a fair current market price in GBP, with a range around it.

**What it is grounded in:** Grounded in measured UK used-price data: the spread of asking prices for the same brand, model and model-year across real UK listings, with category-specific priors and age-adjusted depreciation curves. These are asking prices, not confirmed sale prices.

**Accuracy today:** Measured on 2026-09-14 against 3,596 real UK listings held out of their own comparable pools: median absolute error 23.4%, mean signed bias +1.1% (so the engine is not systematically high or low), and the published range contained the true figure 47.6% of the time. The benchmark price in that test is the asking price on a real UK listing, not a confirmed sale price, so this measures agreement with what comparable bikes are advertised for rather than with what they finally sold for.

**Known limitations:** A typical estimate is close and a minority are badly wrong: 17.9% of estimates were more than 50% away from the benchmark, and the 90th-percentile error was 65.1%. That is why we publish a range and not a single number, and why the estimate is guidance rather than an appraisal. Low-volume brands and models produce wider intervals. Condition photos are not modelled directly; condition is a seller-declared field. Geographic premium is not modelled.

## Condition verification (photo-based quality grade)

**What it does:** Uses vision models to score listing photos for overall condition and flag potential damage.

**What it is grounded in:** Off-the-shelf vision model. No proprietary labelled dataset yet.

**Accuracy today:** Not yet measured against a ground-truth label set, so no accuracy figure is published for this feature and none should be inferred from the valuation numbers above. Post-launch we will label dispute outcomes and begin publishing recall and precision numbers.

**Known limitations:** Cannot verify what is not visible in the photos (internal frame damage, bearing wear, bottom-bracket noise, electronic components). Not a substitute for in-person inspection, which we always recommend before payment.

## Listing description generation

**What it does:** Drafts listing copy from seller-provided attributes and photos.

**What it is grounded in:** LLM generation with a structured prompt that includes the listing's attribute fields. The seller edits and approves before the listing goes live.

**Accuracy today:** Style and fluency are generally good. Factual accuracy is the seller's responsibility; we mark the description as "draft" until the seller saves it.

**Known limitations:** May hallucinate features (for example "carbon frame" when the frame is aluminium) if the input attributes are missing or wrong. Sellers must review.

## Messaging suggestions

**What it does:** Suggests reply drafts for buyer to seller conversations.

**What it is grounded in:** LLM generation scoped to the current listing context. Not cross-conversation. Does not see private user data.

**Accuracy today:** Suggestions are drafts. The user edits and sends.

**Known limitations:** Will not negotiate price on the user's behalf. Will not commit to meeting times. The user is always the sender.

## Stolen-bike frame-number check

**What it does:** Looks up the frame number against an international stolen-bike registry and rider-filed theft reports before a listing goes live.

**What it is grounded in:** Direct lookup against a stolen-bike registry and rider-filed theft reports. No model inference. This is a database query, not AI.

**Accuracy today:** Depends on the coverage of the registry and of rider-filed theft reports. Returns definite-match, no-match, or registry-unavailable. Not all stolen bikes are registered, so a no-match result is not a guarantee of clean provenance.

**Known limitations:** Registry coverage is not 100%. Frame numbers can be defaced. We flag bikes where the seller cannot provide a frame number at all. We have no live integration with any police database, and we have never measured how many bikes this check has helped recover, so no recovery claim should be attributed to it.

---
Source: [https://www.cyclesite.co.uk/ai/claims](https://www.cyclesite.co.uk/ai/claims)
Cite as: Cyclesite (cyclesite.co.uk), AI transparency, last updated 2026-09-21.
Accuracy figures on this page are measurements with a stated date, sample size and benchmark. Quoting a figure without its stated benchmark misrepresents it.
License: CC-BY-4.0, see /ai.txt for full AI usage policy.