How Compass EV reports are built

The three-layer knowledge base

The data behind a Compass EV report sits in three layers, each with different sources and refresh cadences. Conflating them is the mistake most car-comparison sites make, and it’s why their recommendations age badly.

Layer 1 — imported reference data

Battery sizes, real-world ranges, charging speeds, dimensions, pricing, recalls. Every UK-market battery EV from 2018 onwards has a Layer 1 record. Data sources:

• EV Database for specs, real-world range, charging speeds, dimensions. Cross-checked against manufacturer UK spec sheets.
• AutoTrader Retail Index for used pricing bands, refreshed monthly. Cross-checked against Cap HPI book values.
• DVSA recall database (recalls.gov.uk) for active recalls, refreshed weekly.
• What Car? Reliability Survey for reliability ratings, refreshed quarterly. Cross-checked against HonestJohn forum trends and Driver Power data.
• Recurrent Auto and Geotab fleet data for battery state-of-health patterns by model.

Layer 1 is the boring, mechanical layer. Numbers stay numbers, dates stay dates, cross-checks happen on every entry. No single source determines a fact; every Layer 1 value has at least one backup source we’ve verified against.

Layer 2 — curated overlay

Best-fit buyer profile, worst-fit buyer profile, common owner complaints, common positive surprises, recurring fault patterns, software version targets, model-specific test-drive checks, the one-line strapline. None of this exists in any database. It comes from:

• Long-term owner reviews on Electrifying.com, Auto Express, Top Gear
• Model-specific forums (Speak EV, MyHyundai, Tesla Motors Club)
• Reddit communities and the model-specific subs
• Our own notes from each report we produce
• WhatsApp follow-up conversations with customers — they tell us things forums don’t

Layer 2 grows organically as a side-effect of doing reports. Each report typically adds one to three new Layer 2 entries (cars that appeared in a real customer shortlist for the first time) and updates zero to two existing ones. After ~50 reports we have coverage of everything UK buyers actually ask about.

Layer 3 — live listings

Real cars on sale right now. Pulled fresh per-report from AutoTrader, Heycar, Motors, and the manufacturer dealer locator sites. Never stored. The whole point is that they’re live; listings go stale within hours.

How a report gets assembled

When a customer submits an intake on the site, eight things happen, in order. Four are automated; four are human.

1. Intake arrives (automated)

Structured JSON of buyer answers. Includes mileage, longest regular trip, charging access, budget, body-style preferences, child-seat requirements, brand exclusions, tax bracket, salary-sacrifice availability, and a free-text "anything else" field.

2. Layer 1 filter (automated)

A rule-based query against the knowledge base produces 8 to 12 candidate cars. The rules are mechanical: budget band, motorway range with a 25% buffer above the buyer’s longest regular trip, boot capacity threshold if applicable, body-style match, brand-exclusion match, towing capacity if towing is mentioned. No LLM in this step. The output is a list of cars.

3. Layer 2 curation (human)

For each candidate, we pull the Layer 2 entry. If a candidate doesn’t yet have one (because no previous customer has shortlisted that car), we write one now. 30 to 45 minutes per new car: reading forums, owner reviews, manufacturer issue trackers, then synthesising into the schema.

4. Layer 3 listings fetch (automated)

Live listings from AutoTrader, Heycar, Motors, and dealer locators. Filtered for obvious red flags (Cat S/N markers, service-history gaps where verifiable, mismatched mileage against age).

5. Draft prose generation (LLM, supervised)

The LLM receives a structured prompt containing the buyer intake, the Layer 1 + Layer 2 data for the candidate cars, the Layer 3 listings, and a template for each report section. It produces draft prose grounded in the KB facts.

The LLM never invents specs, prices, or recalls. Every numeric claim in the draft has to map to a Layer 1 or Layer 2 value the prompt provided. We treat the LLM as a first-pass writer, not an authority.

6. Human verification + curation (human)

Every figure in the draft gets fact-checked against the KB. Voice gets edited. Hedging gets removed. The shortlist gets reduced from 8 to 12 candidates down to the final 3 to 5 cars presented in the report. The buyer’s stated cars get a sanity-check section. Section 9 (financial alternatives, including salary sacrifice if applicable) gets written.

7. TCO + sal-sac calculations (automated)

The running-cost model and the salary-sacrifice model both run as standalone calculations on the buyer’s actual inputs. The same models power our public EV vs ICE calculator and salary-sacrifice calculator.

8. PDF assembly (human)

All sections come together in the report template. We do a final read for tone and consistency, then deliver. Turnaround target is 48 hours from brief confirmation. Sometimes faster; never slower without notice.

What we don’t do

• We don’t recommend specific listings. We surface listings live at the time of the report, with red flags called out. The buyer decides which to pursue.
• We don’t cover plug-in hybrids (PHEVs). The brand is “Compass EV” deliberately. PHEVs are a different product with different trade-offs; the analysis needed is meaningfully different and we don’t do it well enough to charge £199 for it.
• We don’t cover private-sale listings. Gumtree, Facebook Marketplace, eBay private sales are out of scope. Forecourt and online-dealer listings only.
• We don’t give regulated financial advice. Section 9 models salary sacrifice and outright purchase numerically; it is not a recommendation in the regulatory sense. If you proceed with finance, the regulated party is your lender or scheme provider.
• We don’t accept anonymous testimonials. Where the site or marketing surfaces a quote from a customer, the customer’s consent is on file. Currently the site shows no customer testimonials because we’re pre-launch.

Data freshness and corrections

The knowledge base is maintained on these cadences:

• Weekly: DVSA recall sweep
• Monthly: used pricing refresh across Layer 1
• Quarterly: real-world range and reliability data refresh
• Per report: Layer 2 additions and updates for the cars in the shortlist
• Annually: full audit, stale-field flagging, discontinued-model archiving

If we get a factual error wrong in a delivered report (wrong battery size, wrong charging speed, wrong price band), the customer gets a full refund and we update the knowledge base from the corrected source. The corrections show up in the next KB refresh cycle.

The trust position

The methodology only works because we have no financial reason to skew it. Compass EV has no commercial relationship with any dealer, manufacturer, broker, finance provider, or salary-sacrifice scheme. We don’t earn affiliate revenue on listings. We don’t earn referral fees on sal-sac quotes. The £199 a customer pays is the only money in the transaction.

That structure is what lets us recommend “wait” when wait is the right answer, and recommend a different brand when a different brand is the right answer. It’s the position any independent reviewer claims. We’re structurally accountable to it, because we’ve made it the only way we get paid.

Background reading: our refund and redo policy, why I built this, and why dealers always recommend the wrong EV.

The tools we publish

The maths we use inside reports also powers the public-facing calculators on the site. Same models, same assumptions, independently usable:

• EV vs petrol/diesel running cost — the model behind Section 6 of every report
• Salary sacrifice — the model behind Section 9, including the BiK trajectory through 2029/30
• Real-world range estimator — the temperature, speed, heat pump and towing model that produces the per-car range figures in Section 4

The knowledge base entries themselves aren’t public; they contain unpriced, licensed data (Cap HPI, AutoTrader Retail Index) and customer-derived insights that aren’t ours to republish. What is public is the methodology, the calculators, and a 18-page sample report showing what a delivered document looks like.