The problem
The second most important service line converted badly, and nobody knew how buyers chose a policy.
The quick route was to patch the comparison page. I chose the slower one: a full study of how buyers decide, numbers first, then interviews, recordings and tickets to explain them, so the answer would serve the next problem too.
The drop-off was concentrated on the comparison page, the step where the decision is made.
The finding
Users read the price gap as a quality gap, and couldn't tell which quality.
The comparison page showed a dozen quotes for the same car, some at twice the price, and nothing on the page to say why.
So they went to a street agent who sells one insurer's policy. He wasn't cheaper or better informed. He offered one quote, and one quote needs no explaining.
Interviews and recordings showed where they'd looked for the reason first: the reviews, read for one thing, what happened when someone claimed.
So why did the prices differ? No one internally could tell me. The knowledge was scattered across sales, underwriting contacts and support, and no one person held it.
A fishbone session with every stakeholder produced the answer: insurers price by their own risk appetite, for the city, the vehicle model and year, and how they assess the driver, and the same appetite shapes how they pay out. Each branch has an appetite of its own.
Underneath all of it was coverage. What an insurer pays on a claim is what its policy covers, and no two policies covered the same things.
Users sensed this from the comments: price hinted at what you'd get at claim time, and the only proof was other buyers who had claimed. But it couldn't be rated per company. It varied by branch, by vehicle, by city.
The decision
Normalized coverage and claim experience as the source of truth
Two things had to happen. Coverage had to be normalized: every insurer's terms restated in one structure and applied to this buyer's car, city and driver profile, so a each qoute could be compared on what each would actually pay.
And real claim experiences had to be collected properly and put where the decision happens.
Three pieces:
A review flow that asks what matters at claim time, tied to the buyer's car, city and branch.
A reviews page where those answers read as claim evidence,
On the comparison page, a suggested policy per buyer, built from normalized coverage and the experiences of buyers like them, with the reviews behind it shown beside it.
The catch
The evidence was thin at the start and skewed toward people who had claimed and bothered to write. Where a car or city had few reviews, the suggestion rested on little, and the design had to say so rather than sound sure. It improves with volume, so the first months were the weakest.
And the insurers never confirmed any of it. Their pricing and payout behaviour was inferred from buyers, never disclosed. The model is a reconstruction, and the only way to find out where it's wrong is the next round of reviews.
The Result
Purchase conversion on comprehensive car insurance rose 2 points.
And support received fewer tickets asking which policy is better, the one question it was never allowed to answer.
