Call records you can trust: real conversations only

CallPages call records now separate what really happened from setup noise. Products count as ‘discussed’ only when a real conversation took place; connection attempts that never became calls are labeled ‘Setup failed’ instead of impersonating conversations; and products a visitor browsed live in their own clearly labeled section of the call record.

Owners reading their call history met nonsense: calls that connected for zero seconds — microphone denied, setup abandoned — nevertheless listed products as ‘discussed’, copied from whatever page the visitor happened to be standing on. Hundreds of stored records carried the contamination, and every day quietly added dozens more. A record that invents a conversation is worse than no record: owners act on these lists.

How a call record learned to invent products

When a visitor opens a product page and taps call, the call record is created with context: which product they were standing on. That context is useful — for a real call. The bug was that the context was merged into the record as ‘discussed products’ whether or not a conversation ever happened. A visitor who denied the microphone prompt, or whose setup failed in the first seconds, left behind a ‘call’ that discussed two products at length and lasted zero seconds.

At fleet scale the noise compounded: the contamination was being written continuously, on the order of 75–100 records a day.

Three separations, one principle

The principle is that a record may only claim what it can prove.

First, discussed requires a conversation: the product merge now sits behind a check that the call actually exchanged words. Second, failed attempts are labeled, not dressed up: a microphone denial is useful signal — someone tried to reach you — so it stays in history as ‘Setup failed’, with its reason, instead of impersonating a customer conversation. Third, browsing is browsing: products a visitor viewed in the session around a call are kept as context in their own labeled section of the record, visibly distinct from what the call was about.

The cleanup, and the proof

Fixing the writer does not heal records already written, so the stored contamination was cleaned in the data: 343 records that listed discussed products for conversations that never happened. Since the gate shipped in this release, the daily count of new contaminated records has held at zero — the stream that produced a hundred fabrications a day is off, and the record an owner reads now is a record of things that happened.

How it works

1. The merge that attaches products to a call now runs only when the call actually had a conversation — a real exchange with a transcript. No conversation, no discussed products.
2. Attempts that never became calls — microphone denied, setup dropped — still appear in history, labeled ‘Setup failed’, so the attempt stays visible without impersonating a conversation.
3. Products a visitor browsed before calling are kept, but in their own section of the call record, labeled as browsed in the session — never merged into what the call discussed.
4. The contaminated backlog was cleaned in the data (343 records), and since the gate shipped, the count of new contaminated records has stayed at zero — the previous run-rate was roughly 75–100 invented records a day.

See it live: Mi Junta

Mi Junta — CallPage · Mi Junta — website

From the fix’s verification record. Above: an attempt that never connected (microphone denied) — no ‘Products discussed’ section at all, and what the visitor had looked at sits in its own labeled section. Below: a real call — the discussed products are the ones its transcript actually names, and the lamp the caller only browsed stays in the browsed section, out of the conversation’s record.

Review your calls →