The worst 10% of leads, city by city

You cannot buy "not the bottom decile". The decile is a score this page computes; no supplier sells against it. So the first cut below is the ceiling — what a perfect filter would do — and the other two are things you could actually do on Monday. The gap between them is the real answer to "what happens if we take those leads out".

Cutting a campaign is also not the same as cutting its bad leads. Turn off PMax and you lose its 543 bookings too, not just its waste. That is why the PMax row saves the most money and gives up the most.

Green is better off, red is worse off. Net is spend saved less cash up front given up — the only column that answers whether the cut is worth making.

The bookers see 170,845 leads a quarter — 126,262 fresh and 44,583 recycled, about 56,948 a month. That total, not the fresh count, is the number a cut has to be measured against.

Recycled costs no new ad money, so anything it books is free. Blended as it is sent today it books at 2.90% — already 1.7× the 1.73% of the leads you would be dropping. But the families are nothing like each other, and the supply of the good ones is the real limit.

Read the last column before the fourth. CMN is the best recycled lead we have and there is nowhere near enough of it — filling the whole hole from CMN would need more than a quarter's entire supply. A realistic backfill is CMN and 5AB first, then 6AB, and N1N last.

People, not lead rows — 713,553 rows over 13 months are 475,481 people, 1.5 rows each. A person is in the pool if they have never booked and we have not touched them for that long.

So supply is not the constraint and never was. The constraint is response: these are people who have already not booked. That is why the family mix above matters more than the pool size, and why a fourth send to the same person books at almost nothing.

$179,108 freed by cutting the worst 10%, at each family's own real cost per lead. This is the option the numbers prefer, and it is not close.

The honest caveat, and it is a big one. Every row assumes the CPL holds as you scale, and it will not — Blue Rooms' cost per booking tripled between April and July as volume rose, and Hash has bid-capped the ad sets 20% below where they were coming in. Treat these as the size of the prize, not a forecast. The safe version is to move money in steps and watch the CPL after each one.

The two options solve different problems. Recycled fixes the bookers' volume for free. The reinvestment fixes the bookings. Nothing stops you doing both: cut the bad fresh spend, hold the bookers' desk volume with CMN and 5AB, and move the money into Blue Rooms KIDSX and Blue Rooms 1FX.

And if you only do one thing: Google PMax costs $656 a booking while Blue Rooms KIDSX costs $211. They cost almost exactly the same per lead. That is the whole argument on one line.

This table does not use the score at all. It is the plain whole-window record: a code, an age bracket, enough volume to mean something, and a booking rate under one percent. If you want a list to act on tomorrow without accepting a word of the modelling, this is it.

Neil, 16 Sep 2026: "we now have the actual lead cost for each of those. please do not use a standard lead cost across the board, as they are all very different." They are: Alan's DALFB3 runs $7.52 a lead and his HOU3CX runs $27.68 in the same quarter. Every lead here is priced by its own campaign code in its own month, from the supplier that invoiced it.

Three sources, none of them a blend. Alan — his own client-reporting portal at leadcodes.vercel.app, per code per day, the system he invoices from. Blue Rooms — invoiced spend per campaign per month from the USA Models portal's JSON API. Lead Pronto — Lauren Renshaw's own monthly report, in GBP, converted at 1.3198.

How fresh is this? lead_costs.json was built 2026-09-16. Two of its three sources are BANKED BROWSER PULLS, not live feeds — Alan's portal and the USA Models portal are both session-gated, so they are pulled in Neil's Chrome and written to data/. A nightly rebuild of this page re-reads those files; it does not re-pull them. Sources in use: data/alan_leadcodes_2026-09-16.txt (session-gated; pulled in Neil's logged-in Chrome) and data/usa_models_spend_2026-09-16.txt.

Neil: "anything that looks too far out of the normal range please flag as we found they were wrong a lot." These are the entries the builder refused to take at face value. A CPL under $2 or over $60 is a broken feed, not a cheap or expensive campaign, so those leads fall through to their code's own average instead.

At full code level (BOIG23F, DAL3CX2M …) his portal misaligns leads and spend across the tier and gender splits inside one channel: 8 code-months with leads and no spend, 16 with spend and no leads, and 37 at under $2 or over $60 a lead. Rolling up to the channel the money is actually bought on (BOIG2, DAL3CX) clears almost all of it. The remainder is in the flagged table above.

The worst of them: 2026-06 DAL3CX2F $91.42/lead (24 leads) · 2026-06 DAL3CX2M $107.82/lead (11 leads) · 2026-06 DAL3CX3F $130.38/lead (16 leads) · 2026-06 DAL3CX3M $81.46/lead (13 leads) · 2026-06 DALFB31F $1.98/lead (80 leads) · 2026-06 DALFB31M $0.63/lead (102 leads).

One genuine hole survives the roll-up. New York's NYFB3 carries 222 leads for $6 across the quarter. That is not a cheap channel, it is a channel Alan's portal is not costing at all — worth putting to him directly.

The decile is cut once across the whole window, so each month's share of it moves. A month above 10% contributed more than its share of the bad leads; below 10% and it contributed less. This is the column that says whether the problem is being fixed.

Each lead is scored on three things at once — its campaign code, its 3-digit ZIP and its age band — and the score is that lead's expected booking rate. Every city's leads are then sorted worst-first and cut at 10%. Nothing else is in the score: not the supplier, not the day, not the booker.

The three combine multiplicatively against the city's own mean, each shrunk toward it with a prior worth 250 leads. Shrinkage is not decoration: ranking raw rates ascending fills the worst list with twenty-lead cells at 0.0% and pushes the genuinely large problems off the page. With a prior this strong a segment has to carry real volume to move, so the decile fills with leads burned.

The model is fitted 3 times, each time leaving one calendar month out, and every lead is scored by the fit that never saw its own month (2026-06, 2026-07, 2026-08). This matters more than it sounds: in-sample, a bottom decile is guaranteed to look terrible — that is arithmetic, not a finding. Out of fold it is a prediction that either holds or does not.

It holds. The bottom decile books at 1.73% against 4.96% overall, on leads the model had never seen.

126,262 leads, 2026-06-01 to 2026-08-31 — every fresh lead on a known city code. Recycled codes are out (44,583 of them: they were never bought). Unknown-city codes are out (6) — including every #AT Atlanta lead, which no report can score until the Atlanta ZIP sheet exists.

Dupes, removed and rejected leads all stay in. That is deliberate and it is not the /lead-funnel "usable" figure: the question is what the money bought, and a lead the team had to throw away is exactly what makes a decile bad.

Bookings join on name + marketing code, never name + date — Date Distributed is not Date Lead In, and keying on the date matches under 8% of bookings and produces a plausible-looking 0.1% rate. 7,552 of 8,810 booking rows matched (85.7%); the remainder are mostly bookings against recycled and out-of-window leads. Booking, removal and rejection feeds are read past the window end so a late outcome still lands on the lead that earned it.

1. The money is real, but 2.8% of it is inferred. No flat CPL is used anywhere on this page — every lead is priced by its own code in its own month, from the supplier that invoiced it. 97.1% of leads come straight off an invoice. The rest fall back to the code's own average in another month (105 leads) or, where a code is costed nowhere, to that city's blended cost per fresh lead (3,518 leads). See the Cost basis tab.

2. Booking rate understates how bad the decile is. It stops at the booking. The same things that depress booking — distance and age — also depress the show rate (/distance-age: 64.6% at ≤25 mi down to 51.4% at 100–200 mi; 68.0% for under-18s down to 51.1% at 55+). A decile that books at a third of the average shows up at less than a third.

3. The decile still contains bookings. 219 of them. Cutting it is not free — that is the number to weigh the saving against.

4. The score cannot see interactions. The three dimensions multiply, so a code that is bad overall drags down every age band inside it even where one band is genuinely good. New York's GX codes do exactly that. Where the kill list and the decile disagree, the kill list is the plainer record.