Studio Lead Engine — what happened today

Last full day Tue 25 Aug · America/New_York

25 Aug, hour by hour all times New York

Everything happened between 12:00 and 16:00. Nothing was sent before noon and nothing since 16:00. Replies peak an hour after the biggest send, exactly as you would expect.

25 Aug by city where the 4,140 texts went

The single biggest group today has no city tag at all — a direct consequence of the enrichment break on 21 Aug. Those leads still route correctly in the CRM (it reads the address itself), but you cannot tell which supplier originally paid for them.

What was actually said script mix, 25 Aug afternoon sample

Six scripts are in rotation. Two thirds of today’s volume is the opener and the first bump. The CMN script — the one that asks permission to ring them — went to just 75 people, and it is the only script with a zero failure rate.

Half the people who said yes got no answer the worst fault on this page

Across the full day of 25 Aug, 547 real replies came in. The engine answered 150 of them.

Ignoring the negatives is correct. But of the people who said yes or sent a photo, it is a coin flip — 114 answered, 97 ignored.

When it does answer, the routing is right. Every follow-up sent after a reply was the CMN script, from every starting point — so the logic is sound, it just fires on about half of what reaches it. One defect visible here: 24 people were sent CMN twice, while 78 others got nothing.

Which places each script went to 25 Aug afternoon sample, 3,496 texts

Two things stand out. RE-ASK is almost entirely the no-city group — 666 of its 724 sends (92%), and 80% of everything that group received; they got one opener between them, so the bare leads are on a separate script track from every real city. And Orlando received no openers at all — 97% of its sends were BUMP 1, a message saying “just checking my last message came through” to people who were never sent a first message today.

The no-city leads have no marketing code at all

834 of today’s texts (23.9%) went to leads with no city tag. I pulled 20 of them and checked every field: 0 of 20 carry a marketing code — the sle:source tag is simply absent, not wrong. So for roughly a quarter of today’s sends you cannot say which supplier originally paid for the lead.

But the CRM has every one of them. I walked 30 months of CRM lead history (Mar 2024 → Aug 2026) and matched all 20 by phone — 46 records for 20 people. The original code below is the earliest CRM record for that person, not the code the engine assigned later. So the attribution hole is only in GHL, and it is fully recoverable.

13 of the 20 have been through the CRM more than once — Shauntel Clark six times, Nigeria Mincey and Wanda García four each. The “codes since” column is the recycling trail: #ORL11F → #ORLCMN → #ORL6AB → #ORLN1N is one person re-entering under a new code each time. Two of them first came in back in 2024. Quoting the latest code as the source double-counts the same lead against whichever supplier happens to be last.

What changed this afternoon re-swept 16:31 New York

Mark said he had changed some things. Two sweeps 90 minutes apart show exactly what: the import feed came back on at 14:52 and a new path went live. The published workflows were not touched — both are unchanged since 17:28 UTC — so this was a data and intake change, not a logic change.

The engine is re-ingesting its own output. Five new source tags appeared between the sweeps. Three of them — chin1n, fln1n, orln1n — mean leads whose original code is an N1N code: leads this engine already pushed into the CRM as N1N are coming back in as fresh nurture leads. Boston, Dallas, Houston, Nashville and New York already did this; Chicago, Florida and Orlando just joined. Worth deciding whether that loop is deliberate.

sle:source:chin1n   sle:source:fln1n   sle:source:orln1n   sle:source:dal3cx1fr   sle:source:orl1fx

Yesterday, for comparison Mon 24 Aug

The people who said yes and never arrived today + yesterday

Thirteen yes-sayers are absent from the CRM, but five of those opted out (dnc + remove-requested) — withholding them is correct behaviour, not a fault. That leaves 8 genuine misses. Three of them explicitly asked to be called now. Note that 5 of the 8 should have gone to a 6AB code — yes-without-a-photo is the branch that leaks.

Correctly withheld — these opted out, so the engine was right not to push them:

Did they go to the right place? accuracy record, 24–26 Aug

Every engaged lead in the last three days, scored against the rule {CITY}{FAMILY}. 174 leads. Mismatches are split by type rather than lumped, because they are not equally bad.

The 15 “CRM ahead” cases are leads coded CMN whose GHL tag still says N1N or 6AB — the tag lagging the CRM, not a misroute. I have not counted them as errors, but it does mean the GHL tags are not a reliable record of what actually happened.

The 23 that went wrong named

The 8 photo disagreements are the telling ones. GHL has them tagged no-photo (6AB) but the CRM coded them N1N — so a photo was recognised at push time and the tag was never written. Same failure as the 329 unactioned photos. Two of the three “wrong city” cases are Florida → Orlando, which may well be legitimate since FL codes cover both; only Miyana Bailey (Houston → Dallas) is clearly wrong. And three people who had opted out were pushed into the CRM as callable leads — that is a compliance problem, not an efficiency one.

Every fault found, worst first

Background — how it got here 13–25 Aug

Two dates matter. 13 Aug — bulk imports switch on, contact creation goes from about 5/day to 761/day. 21 Aug — the old estate is retired: 16 legacy workflows set to draft in one day, leaving three published (SLE - 120 Day Nurture v1, SLE - API 1 Pilot - Reply Router v1, Assign State). Everything before 13 Aug is the old manual world and is excluded. Weekends are quiet by design, so a Saturday or Sunday report reading zero is not itself a fault.

Basis note: this history table counts the sle:start20day cohort only (7,629 conversations pulled in full), so its 25 Aug row reads 2,160 where the Today section reads 3,496 — Today covers every conversation in the account, not just that cohort. Both are correct on their own basis; do not add them together.

17 Aug shows 568 "replies" that are DnD enabled by customer system events (message type 25), not texts — real inbound that day was 79. Any report counting raw inbound will be wrong by 8x on days like that.

What this does not cover

What to do next

1. Check the restarted import holds tomorrow, and ask Mark which script the 61 new 400k leads get when they have no message-path. 2. Push the 8 by hand, starting with the three who asked for a call: Tycell Rufus, Alejandro Vasquez, Brianna Knowles. 3. Get tokens for the other 22 accounts — 93% of the flow is invisible. 4. Fix the 6AB branch; it accounts for 5 of the 8 losses. 5. Fix enrichment so sle:source is written — 24% of leads have no supplier code. 6. Make sle_automation_error actually write, so the next break shows on day one. 7. Ask Mark to filter message type 25 out of the reply count.

Static findings page — a snapshot taken 25 Aug 2026, not a live pull. Sources: GHL API (contacts, conversations, 19,849 messages, tags, custom fields, workflows, pipelines) and CRM leads-in for 23–26 Aug 2026. The CRM pull ran on a stale form-ID cache, so a small number of "missing" leads could sit under a campaign the cache does not know about.