Numbers look wrong

More rows than you expected

• You’re counting messages. transcripts has one row per message. A 500-conversation chat file can be 5,000 rows. Ask for counts of distinct conversations.
• Ticket fields repeat on every message. On ticket data, a field like product version sits on every message of the ticket, so counting rows inflates it.
• A conversation has more than one enriched segment. Long tickets, or a survey arriving days after a case, can be enriched as separate time segments under the same key. Ask the agent to use one row per conversation (the earliest segment for the reading, all segments for event flags).
• Re-sent data isn’t the cause. Duplicates you send or upload are removed automatically by ingestion. If a conversation still appears twice in a Dimension or metadata table, it’s usually the segments above. Ask the agent to count distinct conversation keys (partition_key).
• Line breaks split records at upload. See code C04 in Upload errors.

Fewer rows than you expected

• Only enriched conversations appear in a prompt’s table. Shares should be calculated over enriched conversations. If a whole upload is missing, it needs a backfill: see No Dimensions yet.
• Part of the file didn’t land. Compare the row count and date range with your file after every upload.
• A date range limited the backfill. See Troubleshooting enrichment.
• Rows with empty text were dropped.

A date is a day off

Two clocks are involved:

Clock	Time zone	What it affects
The platform’s date fields	US Pacific	Which day or month a conversation falls on. A record at 04:18 UTC on the 1st is still the last day of the previous month in Pacific.
The agent’s own clock	UTC	What “today”, “this week” or “last month” means when you don’t give dates. In the evening in the Americas, UTC is already tomorrow.

For month-end reporting, decide which time zone you report in, tell the agent, and give explicit dates: “Use March 1 to March 31, Pacific time.”

The agent’s numbers don’t match your BI tool

Work through these in order:

1. Same unit? Conversations, not messages. Ask the agent to state the base (the number of conversations) for every percentage.
2. Same days? The platform splits days in US Pacific time. Check which time zone your BI tool uses.
3. Same population? A prompt’s table holds only enriched conversations. Your BI tool may count every conversation.
4. Same pull? If your BI tool reads an Export API pull, compare the same integration ID and the same date range.
5. Same calculation? Ask the agent: “Show me the SQL you ran.”

Results differ between two runs

Two separate analyses can use slightly different populations or definitions. When you compare periods, use the same analysis plan for both, and ask the agent to state the base for every percentage.