Skip to content
allanninal.dev/build
Collaborate

Part 3 of 7 · Address verifier series ~5 min read

How a correction gets suggested

The suggestion is the only moment this system is visible to a customer, and it has about two seconds of their attention. Everything here is about spending those two seconds well, which mostly means offering one thing rather than several.

Key takeaways

  • One suggestion or none. A list of candidates is worse than no help at all.
  • Offer only when there is a single clear candidate; ambiguity means say nothing.
  • Show the specific difference, not two full addresses to compare.
  • Declining is one tap and is never questioned.
  • A declined suggestion is recorded, and repeated declines mean the rule is wrong.

One, or none

How a correction is decided and offeredA vertical chain of five steps entered by a box labelled A near match exists, from the lookup. Step one asks how many candidates the reference returned; two or more exits to Say nothing, because ambiguity helps nobody. Step two asks whether it is close enough, meaning one field with a small change; if not it exits to Say nothing, because it is too different to be a typo. Step three builds the difference, working out what exactly changes. Step four offers it as one line with two buttons. Step five records both what was offered and the answer. A note says two of the three exits are silence, and suggesting badly is worse than not suggesting.AWS ACCOUNTA near match existsfrom the lookupHow many candidates?from the referenceSay nothingambiguity helps nobodytwo or moreClose enough?one field, small changeSay nothingtoo different to be a typonoBuild the differencewhat exactly changesOffer itone line, two buttonsRecord bothoffered, and the answerTwo of the three exits are silence. Suggesting badly is worse than not suggesting.
Fig 1. When a suggestion is offered at all. Both silent exits are common, and choosing silence over an uncertain suggestion is what keeps the ones that are shown trusted.
  • App integration
  • Security & identity
  • Management
  • Analytics
  • Front-end & mobile

Why not a list

The instinct with several candidates is to show them and let the customer pick, and it produces a measurably worse outcome than showing nothing. A dropdown of ten similar addresses at checkout is a cognitive task at exactly the wrong moment, and a meaningful proportion of people will pick the wrong one because it looks close enough.

A wrong address chosen from a picker is worse than the wrong address they typed, because it now looks verified. So the rule is one candidate or silence, and silence is the common case.

Close enough

DifferenceOffer it?Why
Postcode, one characterYesThe classic typo; the rest of the address confirms it
Street type (Road vs Street)YesOnly if one street exists at that postcode
House number, one digitNo14 and 24 are both real houses
A missing flat numberNoThe system does not know which flat
Town nameYesThe postcode is authoritative; the town is a label
Two fields at onceNoThat is not a typo, it is a different address

The house number row is the important one. A one-digit difference looks like the most obvious typo in the world and is the most dangerous suggestion available, because both numbers usually exist and both have somebody living in them. Suggesting a different house number sends a parcel to a real person who did not order it.

Showing the difference

What the customer sees

  • Did you mean TN24 8QL?
  • You entered TN24 8QC — we have 14 Chestnut Road as TN24 8QL.
  • Use TN24 8QL · Keep what I entered
  • Nothing else. No full address comparison, no confidence score, no explanation of what a reference file is.

Highlighting the single changed character is what makes this a two-second decision. Showing two complete addresses side by side and asking somebody to spot the difference is the same information presented as a puzzle, and puzzles at checkout cost orders.

Declining is never questioned

“Keep what I entered” takes one tap, produces no warning, no second confirmation and no red text. The customer knows where they live and the reference file does not always. A system that argues with somebody about their own address is both wrong and infuriating.

Declines are the tuning signal

How repeatedly declined suggestions are handledA horizontal row of five boxes. Suggestion offered: a specific change. Declined: by this customer. And by others: the same suggestion, six times. The rule is wrong: not the customers. Stop offering it: for that address. A note says six people insisting on the same address is not six mistakes.A SUGGESTION EVERYBODY DECLINESSuggestion offereda specific changeDeclinedby this customerAnd by otherssame suggestion, 6 timesThe rule is wrongnot the customersStop offering itfor that addressSix people insisting on the same address is not six mistakes.
Fig 2. What repeated declines mean. A suggestion nobody accepts is a fact about the reference data rather than about the customers.
  • Machine learning
  • Security & identity
  • Analytics
  • Front-end & mobile

This happens most often where the reference file records an address differently from how residents write it — a development with a name the file does not use, a building split into units the file records as one. Six declines is conclusive, and suppressing that specific suggestion is both correct and something no third-party validator will do for you.

Next: the address that cannot be verified at all.

All posts