MUD\WTR
Attribution deep dive · 18 weeks

CAC — Two Lenses

The weekly pulse compares click-attributed CAC against survey-implied CAC on about 900 responses. At that sample only Meta is readable. This is the same comparison pooled over 18 weeks — Mar 30 – Aug 2, 2026 — on 21,473 responses, which is enough to price every channel, test whether each divergence is stable or noise, and include the $622,947 of media that has no Northbeam platform at all.

Window
18 wks
Mar 30 – Aug 2, 2026
Survey responses
21,473
40.7% of new customers
New customers
52,763
Shopify D2C, tracker
Media spend
$5,970,441
99.5% mapped to a lens
Blended CAC
$113
all spend ÷ all NCs
Paid-only CAC
$152
excl. 13,390 earned NCs
Called to your attention
  1. Fund YouTube first: its marginal customer costs $145 against $316 for podcast.

    Both channels get the same test. YouTube's high-versus-low weeks ($16,456 → $31,832/wk) price its marginal customer at $145, near the $113 blended average, with no diminishing returns and the tightest dose-response on the page (r=0.887). Podcast, re-priced across every answer its budget actually moves (see below), comes in at $316 — better than the $726 its own survey row implies, but still 3.5× the $90 target and 2.2× YouTube's. YouTube got cut 44% between May and July; that is the reversal to make first, and it does not require touching podcast at all.

  2. We did the opposite between May and July: YouTube got cut 44%.

    YouTube ran at $36,617/week in mid-May and $20,523/week across the last four weeks. Self-reported YouTube discovery followed it down, 12.6% → 8.9% of new customers. Across all 18 weeks the correlation between YouTube spend and YouTube discovery share is r=0.887 — the tightest dose-response on this page, and the reason to treat that decline as caused rather than coincidental.

  3. Google's $64 CAC is the most misleading number in the account.

    Click attribution prices Search at $64, the cheapest paid channel we run. The survey prices the same $634,714 at $178 — and it does so in 18 of 18 weeks without a single exception. The survey answer "Online Search" also catches organic search, which inflates the denominator and makes $178 the optimistic end of the range. Do not read Search efficiency as headroom to scale.

  4. A quarter of new customers are earned, so paid media actually costs $152 per customer, not $113.

    25.4% of respondents name a source we do not buy — word of mouth 10.5%, retail discovery 2.4%, a healthcare professional 1.0%, plus 9.4% who pick "Other". That is roughly 13,390 of 52,763 customers over 18 weeks arriving without a media dollar attached. Blended CAC divides all spend by all customers and therefore flatters paid performance by about 34%.

  5. Podcast is $540,187 that has never appeared in either lens — and “podcast” is not a channel you can act on.

    Northbeam has no Podcast platform, so 9.0% of media — the third-largest line in the company — has been invisible to every weekly two-lens table we have published. But nobody buys “podcast”; we buy shows, one insertion order at a time, and the accounting system already knows which (GL 60201 lists Modern Wisdom, Dear Media, Aliza Pressman, dr will cole, The Blonde Files and others as separate vendors). An aggregate $316 marginal CAC across that portfolio almost certainly hides shows well under the $90 target and shows several times over it. Nothing on this page should be used to cut or keep podcast as a block. To price it per show we need two things we do not currently have wired up: monthly GL 60201 spend by vendor, and the mapping of which promo codes belong to which show — Shopify discount codes are already queryable and creator-style codes are visible in the data, we just cannot tell which are podcast reads and which are CX or ambassador codes.

The table

Two lenses, 18 weeks

Click CAC is tracker spend ÷ Northbeam first-time customers (Clicks Only, cash). Survey CAC is the same spend ÷ the customers the post-purchase survey attributes to that channel. Ratio is click ÷ survey: above 1 means clicks under-credit the channel, below 1 means they over-credit it. Stability counts how many individual weeks point the same way as the pooled figure — the column that separates a finding from noise.

ChannelSpendClick CAC Survey CAC95% bandRatio StabilitySurvey shareRead
Meta Meta → “Facebook or Instagram Ad” $3,715,42562.6% $16922,039 NCs $150$148–$152 1.13×prior 1.2× 17/18weeks agree 47.06%±0.67pt TRUST CLICKS
Search Google, Microsoft → “Online Search (i.e. Google)” $634,71410.7% $649,986 NCs $178$169–$187 0.36×prior 0.37× 18/18no exceptions 6.77%±0.34pt OVER-CREDITED
Podcast NEW click-invisible Podcast → “Podcast” $540,1879.1% no platform $428only 37% credited here one lens only 2.39%±0.2pt RE-PRICE PER SHOW
YouTube YouTube, YouTube Partnerships → “YouTube” $434,5897.3% $2241,943 NCs $78$75–$81 2.86×prior 2.17× 18/18no exceptions 10.53%±0.41pt SCALE
Creator / affiliate NEW Impact, Partnership Ad Spend → “Someone I Follow On Social Media” $266,3124.5% $1192,231 NCs $222$204–$243 0.54× 16/18weeks agree 2.28%±0.2pt NO SIGNAL
Amazon Amazon → “Amazon” $149,5902.5% no click NCs $483$413–$582 one lens only 0.59%±0.1pt CAN’T READ
AppLovin (games) AppLovin → “Video Game or Mobile Game” $105,7761.8% $185572 NCs $246$215–$288 0.75×prior 0.67× 10/14weeks agree 0.81%±0.12pt WATCH
Native / listicle NEW click-invisible GeistM + First Media → “An Article” $82,7601.4% no platform $91$83–$101 one lens only 1.73%±0.17pt NOT PROVEN
TikTok TikTok → “TikTok” $6,4600.1% $1,0636 NCs $5$5–$6 197.20× too thin weekly 2.27%±0.2pt NOISE
Snapchat Snap → “Snapchat” $3,8820.1% $13529 NCs $38$29–$56 3.50× 2/3weeks agree 0.19%±0.06pt NOISE

Ratios for Meta (1.13×), Search (0.36×) and YouTube (2.86×) can be compared against the growth playbook's 12-month priors of 1.2×, 0.37× and 2.17×. Meta and Search land almost exactly on their priors; YouTube's divergence is wider than the playbook assumes. Those priors are static text in growth-playbook.md — worth updating to these figures.

What to do about each

Meta TRUST CLICKS
$3,715,425 · 62.6% of media · confidence high

At 47.1% of all responses this channel is the mix, so the share-based marginal test does not apply to it — its spend moves total new customers, which is the very thing the test holds fixed. The two lenses agree within 13% in 17/18 weeks, so manage it on click CAC and creative-level data.

Cleanest mapping in the account — one buy, one unambiguous answer.

Search OVER-CREDITED
$634,714 · 10.7% of media · confidence high

Clicks claim 2.78× more customers than the survey supports, in 18/18 weeks without exception. Its $64 click CAC is a harvesting artifact, not headroom, and its survey share barely responds to spend (r=0.374), which is what brand-demand harvesting looks like. Prove it with a holdout before either scaling or cutting.

Survey answer also catches organic search, so its survey CAC reads slightly cheap — which makes the over-crediting finding conservative.

Podcast RE-PRICE PER SHOW
$540,187 · 9.1% of media · confidence medium

Its budget provably moves three different survey answers — its own row +0.4pt, YouTube +0.42pt, Creator / affiliate +0.26pt per $10K/week — so only 37% of the discovery it drives is credited under its own name. Re-priced across all three, the marginal customer costs $316, not the $726 its own row implies. Still 3.5× the $90 target — but this is a portfolio of individual shows bought separately, and an aggregate number cannot tell you which ones to keep.

NEW to this analysis. No Northbeam platform exists, so podcast has never appeared in either lens despite being the 3rd-largest spend line. Insertion orders are booked lumpy, so judge the pooled number, not the weekly line.

YouTube SCALE
$434,589 · 7.3% of media · confidence high

The survey prices it at $78 against a $113 blended average, spend and share move together tightly (r=0.887), and even the marginal customer costs $124 — no diminishing returns visible up to $31,832/week.

Includes Agentio creator placements on the Northbeam side.

Creator / affiliate NO SIGNAL
$266,312 · 4.5% of media · confidence low

Weekly spend of $21,512 against $8,078 produced no additional customers once the falling customer base is controlled for, and spend and share are uncorrelated (r=-0.055). The regime difference is not statistically significant (p=0.2544), so read this as “we cannot detect what this money buys”, not as a proven zero. Either way nothing here justifies the current level.

NEW. The survey answer catches unpaid creator mentions too, so this row is a ceiling on performance, not a measurement.

Amazon CAN’T READ
$149,590 · 2.5% of media · confidence low

A D2C-only survey plus a broken Northbeam Amazon feed. This row prices Amazon-discovery→D2C-purchase, nothing more.

Northbeam has reported $0 Amazon ad spend since Jul 22 2026 (broken integration), and the survey audience is D2C-only buyers, so this row prices Amazon-discovery→D2C purchase, not Amazon itself.

AppLovin (games) WATCH
$105,776 · 1.8% of media · confidence medium

Spend and share do move together (r=0.85), but the marginal customer costs $387 — 3.4× the blended average. Hold at the current level; it is not a scale lever at this price.

Small spend; the survey answer is unusually specific, which helps.

Native / listicle NOT PROVEN
$82,760 · 1.4% of media · confidence medium

Spend stopped entirely for 11 weeks and self-reported share barely moved (1.9% → 1.58%, p=0.0705) — a difference this sample cannot separate from zero. The $91 survey CAC is mostly earned mentions, not media.

NEW. Also catches earned press. Cross-check against rise-2/listicle landing-page sessions.

The test that decides it

Dose–response and the marginal customer

Survey attribution alone can be argued with — people misremember. So for every channel this asks a harder question: when we spent more, did more customers say they came from there? Each channel's 18 weeks are split at its own median spend and the two regimes compared.

The critical control: total new customers fell 33% from the first four weeks to the last four. Comparing raw customer counts between an early and a late regime would credit or blame a channel for the whole account's decline, so implied customers are computed at a fixed base of 2,931 new customers per week throughout.

Channelrspend vs shareShare per +$10K/wk Low regimeHigh regimeΔ NCs/wk Marginal CACp
Metatest not applicable — this channel is the mix 0.738 +0.47pt $174,686 $238,139 n/a n/a
Search 0.374 +0.30pt $27,021191 NCs/wk $43,503207 NCs/wk +16.6 $995not significant 0.101
Podcast 0.720 +0.38pt $20,17258 NCs/wk $39,84985 NCs/wk +27.1 $726 <0.001
YouTube 0.887 +2.41pt $16,456241 NCs/wk $31,832365 NCs/wk +124.1 $124 <0.001
Creator / affiliate -0.055 -0.04pt $8,07870 NCs/wk $21,51264 NCs/wk -6.8 no gainnot significant 0.254
AppLovin (games) 0.850 +0.88pt $1,51511 NCs/wk $10,23834 NCs/wk +22.6 $387 <0.001

Marginal CAC = extra weekly spend ÷ extra weekly customers between the two regimes, at the fixed customer base. It is the number that should drive a budget change — an average CAC tells you what a channel has cost, a marginal CAC tells you what the next dollar buys. p tests whether the survey-share difference between regimes is real.

Why Meta is greyed out. Survey share is a composition — every channel plus earned demand sums to 100%. At 47.1% of responses Meta can only gain share by taking it from the others, and most of the movement in total new customers is Meta. Holding the customer base fixed, which is exactly what makes this test valid for a 10.5% channel, guts it for a 47.1% one. Meta's marginal read has to come from creative- and campaign-level data and from the fact that its two lenses agree (1.13× over 17/18 weeks), not from this table.

YouTube against Podcast, plotted

weekly spendsurvey share of new customers

YouTube's two lines track each other closely (r=0.887). Podcast's own row responds too (r=0.72), but at $726 per marginal customer — and that figure is wrong in podcast's favour, because its budget also moves two other answers; see credit leak, which re-prices it at $316. Native/listicle got a cleaner test still: spend stopped entirely for 11 weeks and self-reported "article" discovery moved from 1.9% to 1.58% (p=0.0705) — a difference the sample cannot distinguish from zero. Taken at face value the point estimate implies $1,030 per incremental customer, consistent with the March GeistM audit that priced it at $331–463.

Whose answer does each budget move?

The survey misassigns credit between channels

A survey answer is what the customer remembers, and people remember the surface, not the buy. A video podcast is watched on YouTube; a podcast host is also someone you follow. So the test: regress each answer's weekly share on all five major budgets at once, which asks whether podcast spend moves the YouTube answer while holding YouTube's own spend fixed.

Survey answer Meta budgetSearch budgetPodcast budgetYouTube budgetCreator budget
Meta +0.45p=0.002-0.55-0.27-0.56-0.56 0.731
Search -0.27p<0.001+0.03-0.17+0.34+0.43p=0.050 0.700
Podcast -0.07-0.32p=0.042+0.40p<0.001+0.21-0.06 0.653
YouTube -0.03+0.31+0.42p=0.046+2.36p<0.001+0.31 0.876
Creator / affiliate +0.06-0.06+0.26p=0.022-0.32-0.10 0.433

Each cell is percentage points of that answer's share bought per +$10K/week of that budget. Bold = p<0.05. The diagonal is a channel getting credit for itself; anything off the diagonal is credit landing on the wrong row. df=12.

What this changes
  1. Podcast is credited with only 37% of the discovery its budget drives.

    $10K/week of podcast spend buys +0.4pt of “Podcast”, +0.42pt of “YouTube”, +0.26pt of “Creator / affiliate” — 1.08pt of discovery in total, of which its own row carries 0.4pt. Re-priced across all three, podcast's marginal customer costs $316, not the $726 the single-row test produced. Both mechanisms are the obvious ones: the shows we buy publish on YouTube, and their hosts are creators people follow.

  2. The YouTube conclusion survives it.

    Podcast and YouTube budgets are essentially uncorrelated across the window (r=0.033), so podcast is not what produced the YouTube dose-response. Controlling for podcast spend, the partial correlation between YouTube spend and YouTube share is +0.919, against a raw 0.887 — it strengthens rather than weakens, and YouTube's own-effect marginal CAC re-prices from $124 to $145. Some of what the YouTube row shows is podcast-driven; most of it is not.

  3. Only leaks with a mechanism are applied.

    At 12 degrees of freedom some off-diagonal coefficients will clear p<0.05 by chance. Search's apparently negative effect on the podcast answer has no story behind it and is not used to re-price anything. A significant coefficient with no mechanism is a coincidence.

The quarter nobody buys

Earned demand

25.4% of new customers name a source we do not pay for. That is ~13,390 of 52,763 customers over 18 weeks. It is why blended CAC ($113) and the cost of a customer that paid media actually produced ($152) differ by 34%.

No CAC is shown for these rows on purpose. The referral program (Superfiliate, $13,685) touches only a sliver of word of mouth; dividing it by the whole word-of-mouth share would print a $2 CAC and mean nothing.

SourceShare95% bandNCsn
Word of mouthEarned. The single most valuable line on this page: paid media should not be charged for these customers. 10.53%±0.41pt 5,556 2,261
Other / unattributed9.4% of respondents pick "Other" — the ceiling on how precise any survey lens can be. 9.37%±0.39pt 4,944 2,013
Retail discoveryDiscovered in retail, bought on D2C — the retail halo, measurable here and nowhere else. 2.43%±0.21pt 1,282 522
Healthcare professionalEarned recommendation. 0.96%±0.13pt 507 207
TVNo TV line in the tracker for this window — likely brand/PR recall. 0.75%±0.12pt 396 160
Other social (X, Pinterest)Organic. 0.70%±0.11pt 369 151
Email / newsletterPaid newsletter placements are not broken out in the tracker. 0.63%±0.11pt 332 136

Word of mouth is the steadiest line in the dataset. It has held flat while paid volume fell — 11.6% across the last four weeks against 11.2% in the first four, a +0.4pt difference at p=0.5619 — inside the noise, so do not read it as a brand-demand trend either way. Holding share while the paid base shrinks does mean word of mouth is a growing proportion of a smaller total, not a growing number of customers. Retail discovery (2.4%, ~1,282 customers) is the retail-to-D2C halo, and this survey is the only place it is measurable at all.

Method

How this is built, and where it can be wrong

Spend comes from the Daily Spend Tracker, not Northbeam. This is the one deliberate departure from the weekly pulse's version of this table. Northbeam's platform export has no line for Podcast, GeistM/First Media or partnership spend — a Northbeam-sourced table silently drops 10.4% of media and cannot price the third-largest channel in the company. Where Northbeam does carry a platform the tracker mostly relays it verbatim: Search matches to the dollar, Meta differs about 1.5%.

Click CAC is Northbeam, Clicks Only, cash accounting — the standard house model. Survey CAC divides the same spend by (survey answer share × D2C new customers), so it assumes the 59.3% of customers who did not answer look like the 40.7% who did. That is the largest assumption on this page and it cannot be tested from inside the data.

The survey measures memory, not media — and memory attaches to the surface, not the buy, which is why the credit-leak section exists. "Online Search" catches organic search; "An Article" catches earned press; "Someone I Follow" catches unpaid creator mentions. Each of those inflates a denominator and makes that channel look cheaper than it is — which is why the two channels flagged CUT and NOT PROVEN are flagged on dose-response and on/off evidence rather than on their survey CAC alone. Rows carry a confidence rating for this reason.

Northbeam attributes on a 14-day window, so click NCs for the final week (2026-07-27) are not fully settled and that week's click CAC reads high. Pooled figures are barely affected. Amazon is unreadable here on two counts: Northbeam has reported $0 Amazon ad spend since 22 Jul 2026 (a broken integration, not a media pause), and the survey audience is D2C buyers only.

The composition trap. Survey shares sum to 100%, so a channel's share can move because a different channel changed. Every finding on this page that rests on the marginal test therefore applies only to minority channels; for the 47.1% channel the test is suppressed rather than reported. The same caution applies to reading any single week's shares against each other.

Aggregate channels are not decision units. Podcast is the clearest case — a portfolio of individually negotiated shows reported as one line — but it applies to Creator / affiliate too. Per-show economics need monthly GL 60201 spend by vendor (which exists in accounting, not in any sheet wired up here) plus a promo-code-to-show mapping. shopify_orders.discount_code is queryable in Omni today and creator-style codes are plainly visible in it; what is missing is knowing which codes are podcast reads rather than CX or ambassador codes. Promo codes will still under-count badly — most listeners do not use the code — so treat them as a relative ranking between shows, never as absolute volume.

What is not on this page: incrementality. Neither lens is a holdout test. Where the two disagree by more than 2× on real money — Search and YouTube — the right next step is a geo or audience holdout, not a bigger survey.

Every dollar of spend, and which lens saw it

Tracker channelSpendClusterIf not clustered, why
Meta $3,715,425 Meta
Google $626,348 Search
Podcast $540,187 Podcast
YouTube $434,589 YouTube
Impact $243,643 Creator / affiliate
Amazon $149,590 Amazon
AppLovin $105,776 AppLovin (games)
GeistM + First Media $82,760 Native / listicle
Partnership Ad Spend $22,669 Creator / affiliate
Shopify Shop $15,437 not clustered Shop app surface, no survey answer maps to it.
Superfiliate $13,685 not clustered Referral program — see the word-of-mouth note; no defensible denominator.
Microsoft $8,366 Search
TikTok $6,460 TikTok
Snap $3,882 Snapchat
Rokt $1,624 not clustered Post-purchase ad network — reaches people who already bought.

Sources: Daily Spend Tracker (spend_daily, customers_daily) · Northbeam platform export, 18 weekly pulls, Clicks Only + cash · Fairing question 2864 "What led you to purchase today?", new customers only, via fairing-history.json · rebuild with node scripts/pull-two-lens-history.mjs && node scripts/build-two-lens-report.mjs.