First, today, open Meta Ads Manager and pause these two ads by name:
| Ad name | Campaign | Live since | Spent | Returned |
|---|---|---|---|---|
Catalog_ShopAll_EssentialTees_Single_2026-CW36 | 30_CBO_CATALOG_SHOPALL_US_2026-CW36 | 2026-08-30 | $300.41 | $113.15 |
Catalog_NewArrivals_Home_Single_2026-CW36 | 30_CBO_CATALOG_NEWARRIVALS_US_2026-CW36 | 2026-08-30 | $300.66 | $108.17 |
Together: $601.07 spent, 6 purchases, $221.32 back — a 0.37 return. The Essential Tees one has made zero sales in its last seven days on $167.51. Your own break-even, written in your own plan, is a 1.87 return; these are running at one fifth of it. At 52.3% gross margin that is about $485 of gross profit destroyed in thirteen days, and they are still spending roughly $45–60 a day as of 2026-09-11.
Second, this weekend, do the read that was never done. Your ten-cell destination test (Catalog_A1..E2_..._CW33) ran 2026-08-12 to 2026-08-31, spent $2,881.00 and returned $2,431.36 on 47 purchases. It cleared both stopping conditions your own plan set — 19 elapsed days against a 14-day minimum, and 47 orders against a ~40-order floor. Apply the plan's own decision rule (contribution per purchase = average order value × 52.3% − cost per purchase, break-even ≈ $21) and the verdict is not "which destination won." It is all five lost money:
| Destination | Spend | Purchases | Revenue | Return | Cost per purchase | Contribution per purchase |
|---|---|---|---|---|---|---|
| Homepage | $605.84 | 16 | $805.07 | 1.33 | $37.87 | −$11.55 |
| Essential Tees collection | $763.18 | 14 | $732.95 | 0.96 | $54.51 | −$27.13 |
| Product pages (deep links) | $765.15 | 13 | $639.01 | 0.84 | $58.86 | −$33.15 |
| Top Picks collection | $373.99 | 2 | $155.31 | 0.42 | $187.00 | −$146.38 |
| Bundle & Save collection | $372.84 | 2 | $99.02 | 0.27 | $186.42 | −$160.53 |
Third, write two sentences into wescale-kb/members/tim-dobyns.md under WeScale Scorecard Health: "Catalog destination test CW33 completed 2026-08-31. No destination cleared break-even; the two collection-page destinations lost roughly six times as much per purchase as the homepage, so collection pages are out for catalog ads until something changes." Then amend section I of planning/igl-cro-audit-spec-for-iris-2026-06-18.md, line 116, which currently reads "Catalog ads must land on collections/all-products, NOT a thin 'best sellers' page" — that rule points at the worst-performing destination class you have measured.
You ran a complete, properly designed experiment, it finished twelve days ago, it says every destination you tested loses money — and in the three days after it ended you launched two more catalog ads that have since burned $601 without anyone reading the result.
Tobi Lutke's account of what nearly killed Shopify is not that people did bad work. It is that work happened where his attention was not. A Toronto team built supermarket features for months: "it just turned out that like, you know, in Toronto office, there was a fairly big project to build the features to add to Shopify so you could use it for running a supermarket." He explains why he never saw it: "People learned pretty quickly what of the projects that I'm interested in — talk me about both and nothing else. People learned how to direct my attention to the things that they wanted to have my attention." The fix was brute force: "The moment I went through absolutely every project I did myself, took like 60-hour days. I canceled probably 60% of the projects."
IGL's attention is directed at design count, because that is the number that goes into a cell every Tuesday. The destination test was not neglected out of laziness — the opposite. There is a careful reconciliation document dated 2026-08-27 that joined Meta to Shopify session by session, verified 20 completed checkouts against 20 Shopify orders with zero delta, and paused seven arms on the evidence. It is good work. But it read the test mid-flight, on two partial windows (2026-08-17 to 08-23, and 08-22 to 08-26), and it said so plainly: "This is an operational pruning decision, not a guaranteed statistical winner or scaling verdict." It kept two arms, D2 and SA1, as the survivors.
Then both survivors stopped spending on 2026-08-30 — and on that same day two brand-new catalog ads started. Nobody wrote the final read. The attention had moved to the next wave before the last one reported.
That is the non-obvious part, and it is worse than simple neglect. The mid-flight prune looked like the test had been handled. A test that gets partially read is more dangerous than one nobody touches, because the partial read consumes the feeling of having decided.
Two specifics make the full read worth the hour. First, the bottom of the table replicates and the top does not. Pool all four structured catalog waves (CW33, CW34, CW35, CW36 — $4,134 of spend) and the homepage advantage evaporates: Essential Tees 0.87, deep links 0.84, homepage 0.83, a three-way tie. What holds is the floor — Top Picks collection 0.29 and Bundle & Save 0.27 across every wave. The confident finding is which destinations to stop using, not which to scale. Second, that floor contradicts your own most recent audit. The 2026-08-09 ads and conversion audit called Bundle & Save "the best destination in the store" at a 6.09% conversion rate and recommended routing to it. /collections/bundle-save is still the single largest landing page in the store — 13,845 sessions in the last 90 days, nearly six times the biggest product page. For catalog ads specifically, it was the worst of the five.
No single knowledge base could produce this. The mastermind shelf holds the catalog doctrine the test was built from; the Skool classroom holds the argument for stopping; the founder shelf holds the mechanism that explains the failure. The result itself lives in an operations database outside all thirteen bases, which is exactly why it went twelve days unread.
Honest limits. These are Meta-attributed purchases and revenue, not the Shopify-verified, utm-matched orders your plan names as the judgment standard — Meta usually reports generously, so the real picture is likely worse, not better. Per-cell counts are small (0 to 13 purchases), so single-cell reads are directional. The two losing destinations also got about half the spend of the others and stopped earliest, so their ranking is partly a consequence of having been paused for losing. And the format half of the test is genuinely split: carousel beat single image in four of five destinations (0.99 versus 0.71 overall), but the single best cell in the grid was Home/Single at 1.46, and single image produced more purchases (25) on lower revenue.
All ten cells, ranked by return, 2026-08-12 to 2026-08-31:
| Cell | Destination | Format | Spend | Purchases | Revenue | Return |
|---|---|---|---|---|---|---|
| D2 | Homepage | single | $397.74 | 13 | $578.72 | 1.46 |
| E1 | Essential Tees | carousel | $381.75 | 9 | $490.22 | 1.28 |
| D1 | Homepage | carousel | $208.10 | 3 | $226.35 | 1.09 |
| A1 | Deep links | carousel | $399.00 | 7 | $421.22 | 1.06 |
| B1 | Top Picks | carousel | $188.38 | 1 | $119.80 | 0.64 |
| E2 | Essential Tees | single | $381.43 | 5 | $242.73 | 0.64 |
| A2 | Deep links | single | $366.15 | 6 | $217.79 | 0.59 |
| C1 | Bundle & Save | carousel | $188.57 | 2 | $99.02 | 0.53 |
| B2 | Top Picks | single | $185.61 | 1 | $35.51 | 0.19 |
| C2 | Bundle & Save | single | $184.27 | 0 | $0.00 | 0.00 |
The shape that actually works, for when you rebuild: Catalog_New Arrivals - Single Image - Top Picks Landing Page — New Arrivals product set, single image, pointed at the Top Picks landing page. Over the last 90 days (2026-06-14 to 2026-09-11) it spent $733.94 for 34 purchases and $1,498.46, a 2.04 return — the highest-revenue ad in the account over that window, and second by return only to Life's Too Short (2.67 on $507.45) among ads spending $300 or more. Account-wide over the same 90 days: $27,438.92 spent, 499 purchases, $25,265.60 returned, a 0.92 return across 395 ads. It stopped spending on 2026-08-31. That is the one catalog ad worth rebuilding, and note what it is: the Top Picks landing page, which is a different page from the /collections/top-picks-v2 destination that placed fourth in the grid.
Where to check any of this yourself: ```sql -- the completed test, one row per cell SELECT ad_name, MIN(date), MAX(date), ROUND(SUM(spend),2), SUM(purchases), ROUND(SUM(revenue),2) FROM fact_meta_daily WHERE level='ad' AND ad_name LIKE 'Catalog_%CW33%' GROUP BY ad_name ORDER BY SUM(revenue)/SUM(spend) DESC;
-- the two ads still running SELECT ad_name, date, spend, purchases, revenue FROM fact_meta_daily WHERE level='ad' AND ad_name LIKE 'Catalog_%CW36%' AND date >= '2026-09-05' ORDER BY date; `` Database: G:/iris/igl-ops-sage/data/igl.db`, synced through 2026-09-11.
founders-kb/interviews/009-tobi-lutke-transcript — primary. The seed, and the mechanism. The Toronto supermarket project he never knew about (line 34–36), the attention-routing admission and the fix (line 40). Raw transcript of the David Senra interview show, published 2026-01-18.wescale-kb/calls/call-28 — primary. Jakob & Petra's four-step catalog pipeline, the doctrine this test was built on: a low-budget tag-based catalog auto-runs on every launch, designs with 10+ clicks graduate out, and "always remove promoted designs from the testing catalog." It is the rotation discipline that the CW36 pair skipped. It also sets the depth gate at 10–15 purchases in 7 days — a bar no cell in the grid came close to.stress-free-scaling-kb/lessons/most-ad-accounts-need-two-things-not-twenty — primary. Skool classroom lesson, captured 2026-08-24: "unless you're consistently spending $100k+ per month toward a problem and persona, most of what you need to do in your ad account is remarkably simple." The argument for reading wave two before launching wave four.wescale-kb/themes/catalog-ads-playbook — theme-aggregate, the one permitted. Establishes that the cohort treats catalog ads as the discovery engine and statics as the scaler, and records the standing IGL prescription — which is what makes the unread result the gap, rather than the general advice.G:/iris/igl-ops-sage/data/igl.db — primary, outside all thirteen knowledge bases. fact_meta_daily: every figure in this card, re-derived by the operator. fact_landing_page_daily: /collections/bundle-save at 13,845 sessions over 90 days.planning/igl-cro-audit-spec-for-iris-2026-06-18.md — primary. Dated 2026-06-18, a spec, not a status record. Section I, line 116: catalog ads must land on collections/all-products. Line 119 hedges on format — "Catalog = carousel (more products visible), though test single-image for bestseller catalogs" — which the test vindicates.G:/iris-03a/deliverables/ads-test-implementation-plan-2026-08-10.md — primary. The test design and, crucially, its own read rule: collapse format twins into five destination buckets; judge on contribution per purchase (AOV × 52.3% − cost per purchase, break-even ≈ $21); minimum 14 elapsed days and a ~40-order floor before a verdict.G:/iris-03a/deliverables/catalog-14-arm-shopify-reconciliation-and-action-plan-2026-08-27.md — primary. The careful mid-flight prune: 14-arm registry, 20 checkouts reconciled to 20 Shopify orders with zero delta, seven arms paused, D2 and SA1 kept. Explicitly "not a guaranteed statistical winner or scaling verdict."G:/iris-03a/deliverables/sor-ads-cro-audit-2026-08-09.md — primary. Calls Bundle & Save "the best destination in the store" at 6.09% conversion (72 Shopify orders / 1,183 sessions, 30 days) — the claim the catalog test contradicts for catalog traffic specifically.Catalog_ShopAll_EssentialTees_Single_2026-CW36 | 30_CBO_CATALOG_SHOPALL_US_2026-CW36 | 2026-08-30 | $300.41 | $113.15 | | Catalog_NewArrivals_Home_Single_2026-CW36 | 30_CBO_CATALOG_NEWARRIVALS_US_2026-CW36 | 2026-08-30 | $300.66 | $108.17 | Together: $601.07 spent, 6 purchases, $221.32 back — a 0.37 return. The Essential Tees one has made zero sales in its last seven days on $167.51. Your own break-even, written in your own plan, is a 1.87 return; these are running at one fifth of it. At 52.3% gross margin that is about $485 of gross profit destroyed in thirteen days, and they are still spending roughly $45–60 a day as of 2026-09-11. Second, this weekend, do the read that was never done. Your ten-cell destination test (Catalog_A1..E2_..._CW33) ran 2026-08-12 to 2026-08-31, spent $2,881.00 and returned $2,431.36 on 47 purchases. It cleared both stopping conditions your own plan set — 19 elapsed days against a 14-day minimum, and 47 orders against a ~40-order floor. Apply the plan's own decision rule (contribution per purchase = average order value × 52.3% − cost per purchase, break-even ≈ $21) and the verdict is not "which destination won." It is all five lost money: | Destination | Spend | Purchases | Revenue | Return | Cost per purchase | Contribution per purchase | |---|---:|---:|---:|---:|---:|---:| | Homepage | $605.84 | 16 | $805.07 | 1.33 | $37.87 | −$11.55 | | Essential Tees collection | $763.18 | 14 | $732.95 | 0.96 | $54.51 | −$27.13 | | Product pages (deep links) | $765.15 | 13 | $639.01 | 0.84 | $58.86 | −$33.15 | | Top Picks collection | $373.99 | 2 | $155.31 | 0.42 | $187.00 | −$146.38 | | Bundle & Save collection | $372.84 | 2 | $99.02 | 0.27 | $186.42 | −$160.53 | Third, write two sentences into wescale-kb/members/tim-dobyns.md under WeScale Scorecard Health: "Catalog destination test CW33 completed 2026-08-31. No destination cleared break-even; the two collection-page destinations lost roughly six times as much per purchase as the homepage, so collection pages are out for catalog ads until something changes." Then amend section I of planning/igl-cro-audit-spec-for-iris-2026-06-18.md, line 116, which currently reads "Catalog ads must land on collections/all-products, NOT a thin 'best sellers' page" — that rule points at the worst-performing destination class you have measured. ## Why (the one-liner) You ran a complete, properly designed experiment, it finished twelve days ago, it says every destination you tested loses money — and in the three days after it ended you launched two more catalog ads that have since burned $601 without anyone reading the result. ## The insight Tobi Lutke's account of what nearly killed Shopify is not that people did bad work. It is that work happened where his attention was not. A Toronto team built supermarket features for months: "it just turned out that like, you know, in Toronto office, there was a fairly big project to build the features to add to Shopify so you could use it for running a supermarket." He explains why he never saw it: "People learned pretty quickly what of the projects that I'm interested in — talk me about both and nothing else. People learned how to direct my attention to the things that they wanted to have my attention." The fix was brute force: "The moment I went through absolutely every project I did myself, took like 60-hour days. I canceled probably 60% of the projects." IGL's attention is directed at design count, because that is the number that goes into a cell every Tuesday. The destination test was not neglected out of laziness — the opposite. There is a careful reconciliation document dated 2026-08-27 that joined Meta to Shopify session by session, verified 20 completed checkouts against 20 Shopify orders with zero delta, and paused seven arms on the evidence. It is good work. But it read the test mid-flight, on two partial windows (2026-08-17 to 08-23, and 08-22 to 08-26), and it said so plainly: "This is an operational pruning decision, not a guaranteed statistical winner or scaling verdict." It kept two arms, D2 and SA1, as the survivors. Then both survivors stopped spending on 2026-08-30 — and on that same day two brand-new catalog ads started. Nobody wrote the final read. The attention had moved to the next wave before the last one reported. That is the non-obvious part, and it is worse than simple neglect. The mid-flight prune looked like the test had been handled. A test that gets partially read is more dangerous than one nobody touches, because the partial read consumes the feeling of having decided. Two specifics make the full read worth the hour. First, the bottom of the table replicates and the top does not. Pool all four structured catalog waves (CW33, CW34, CW35, CW36 — $4,134 of spend) and the homepage advantage evaporates: Essential Tees 0.87, deep links 0.84, homepage 0.83, a three-way tie. What holds is the floor — Top Picks collection 0.29 and Bundle & Save 0.27 across every wave. The confident finding is which destinations to stop using, not which to scale. Second, that floor contradicts your own most recent audit. The 2026-08-09 ads and conversion audit called Bundle & Save "the best destination in the store" at a 6.09% conversion rate and recommended routing to it. /collections/bundle-save is still the single largest landing page in the store — 13,845 sessions in the last 90 days, nearly six times the biggest product page. For catalog ads specifically, it was the worst of the five. No single knowledge base could produce this. The mastermind shelf holds the catalog doctrine the test was built from; the Skool classroom holds the argument for stopping; the founder shelf holds the mechanism that explains the failure. The result itself lives in an operations database outside all thirteen bases, which is exactly why it went twelve days unread. Honest limits. These are Meta-attributed purchases and revenue, not the Shopify-verified, utm-matched orders your plan names as the judgment standard — Meta usually reports generously, so the real picture is likely worse, not better. Per-cell counts are small (0 to 13 purchases), so single-cell reads are directional. The two losing destinations also got about half the spend of the others and stopped earliest, so their ranking is partly a consequence of having been paused for losing. And the format half of the test is genuinely split: carousel beat single image in four of five destinations (0.99 versus 0.71 overall), but the single best cell in the grid was Home/Single at 1.46, and single image produced more purchases (25) on lower revenue. ## Worked examples All ten cells, ranked by return, 2026-08-12 to 2026-08-31: | Cell | Destination | Format | Spend | Purchases | Revenue | Return | |---|---|---|---:|---:|---:|---:| | D2 | Homepage | single | $397.74 | 13 | $578.72 | 1.46 | | E1 | Essential Tees | carousel | $381.75 | 9 | $490.22 | 1.28 | | D1 | Homepage | carousel | $208.10 | 3 | $226.35 | 1.09 | | A1 | Deep links | carousel | $399.00 | 7 | $421.22 | 1.06 | | B1 | Top Picks | carousel | $188.38 | 1 | $119.80 | 0.64 | | E2 | Essential Tees | single | $381.43 | 5 | $242.73 | 0.64 | | A2 | Deep links | single | $366.15 | 6 | $217.79 | 0.59 | | C1 | Bundle & Save | carousel | $188.57 | 2 | $99.02 | 0.53 | | B2 | Top Picks | single | $185.61 | 1 | $35.51 | 0.19 | | C2 | Bundle & Save | single | $184.27 | 0 | $0.00 | 0.00 | The shape that actually works, for when you rebuild: Catalog_New Arrivals - Single Image - Top Picks Landing Page — New Arrivals product set, single image, pointed at the Top Picks landing page. Over the last 90 days (2026-06-14 to 2026-09-11) it spent $733.94 for 34 purchases and $1,498.46, a 2.04 return — the highest-revenue ad in the account over that window, and second by return only to Life's Too Short (2.67 on $507.45) among ads spending $300 or more. Account-wide over the same 90 days: $27,438.92 spent, 499 purchases, $25,265.60 returned, a 0.92 return across 395 ads. It stopped spending on 2026-08-31. That is the one catalog ad worth rebuilding, and note what it is: the Top Picks landing page, which is a different page from the /collections/top-picks-v2 destination that placed fourth in the grid. Where to check any of this yourself: ``sql -- the completed test, one row per cell SELECT ad_name, MIN(date), MAX(date), ROUND(SUM(spend),2), SUM(purchases), ROUND(SUM(revenue),2) FROM fact_meta_daily WHERE level='ad' AND ad_name LIKE 'Catalog_%CW33%' GROUP BY ad_name ORDER BY SUM(revenue)/SUM(spend) DESC; -- the two ads still running SELECT ad_name, date, spend, purchases, revenue FROM fact_meta_daily WHERE level='ad' AND ad_name LIKE 'Catalog_%CW36%' AND date >= '2026-09-05' ORDER BY date; ` Database: G:/iris/igl-ops-sage/data/igl.db, synced through 2026-09-11. ## Evidence trail - founders-kb/interviews/009-tobi-lutke-transcript — primary. The seed, and the mechanism. The Toronto supermarket project he never knew about (line 34–36), the attention-routing admission and the fix (line 40). Raw transcript of the David Senra interview show, published 2026-01-18. - wescale-kb/calls/call-28 — primary. Jakob & Petra's four-step catalog pipeline, the doctrine this test was built on: a low-budget tag-based catalog auto-runs on every launch, designs with 10+ clicks graduate out, and "always remove promoted designs from the testing catalog." It is the rotation discipline that the CW36 pair skipped. It also sets the depth gate at 10–15 purchases in 7 days — a bar no cell in the grid came close to. - stress-free-scaling-kb/lessons/most-ad-accounts-need-two-things-not-twenty — primary. Skool classroom lesson, captured 2026-08-24: "unless you're consistently spending $100k+ per month toward a problem and persona, most of what you need to do in your ad account is remarkably simple." The argument for reading wave two before launching wave four. - wescale-kb/themes/catalog-ads-playbook — theme-aggregate, the one permitted. Establishes that the cohort treats catalog ads as the discovery engine and statics as the scaler, and records the standing IGL prescription — which is what makes the unread result the gap, rather than the general advice. - G:/iris/igl-ops-sage/data/igl.db — primary, outside all thirteen knowledge bases. fact_meta_daily: every figure in this card, re-derived by the operator. fact_landing_page_daily: /collections/bundle-save at 13,845 sessions over 90 days. - planning/igl-cro-audit-spec-for-iris-2026-06-18.md — primary. Dated 2026-06-18, a spec, not a status record. Section I, line 116: catalog ads must land on collections/all-products. Line 119 hedges on format — "Catalog = carousel (more products visible), though test single-image for bestseller catalogs" — which the test vindicates. - G:/iris-03a/deliverables/ads-test-implementation-plan-2026-08-10.md — primary. The test design and, crucially, its own read rule: collapse format twins into five destination buckets; judge on contribution per purchase (AOV × 52.3% − cost per purchase, break-even ≈ $21); minimum 14 elapsed days and a ~40-order floor before a verdict. - G:/iris-03a/deliverables/catalog-14-arm-shopify-reconciliation-and-action-plan-2026-08-27.md — primary. The careful mid-flight prune: 14-arm registry, 20 checkouts reconciled to 20 Shopify orders with zero delta, seven arms paused, D2 and SA1 kept. Explicitly "not a guaranteed statistical winner or scaling verdict." - G:/iris-03a/deliverables/sor-ads-cro-audit-2026-08-09.md — primary. Calls Bundle & Save "the best destination in the store" at 6.09% conversion (72 Shopify orders / 1,183 sessions, 30 days) — the claim the catalog test contradicts for catalog traffic specifically. ## Assay verdicts - Novelty — PASS (4/5). No prior art in any of the thirteen knowledge bases: a grep for CW33, TopPicks_Home and "destination verdict" across every .md file returns zero hits, and none of the 37 prior nuggets reads this test. Operator amendment: the judge's phrasing "nothing has read it" was too strong and is corrected in this card — the 2026-08-27 reconciliation read it mid-flight and pruned on it. What is new is the completed-test verdict, the contribution-per-purchase read, the pooled four-wave floor, and the two ads still running. - Grounding — PASS (5/5). Every figure re-derived independently by the operator from igl.db; all document quotations checked verbatim. Four corrections applied before publication (below). - Actionability — PASS (5/5). Two ads named exactly, a rule amended at a named file and line, a decision recorded in a named block. One sitting, no new spend, no staff. ## Operator corrections applied before publication 1. "The homepage won" — removed. True within CW33 alone (1.33). Pooling all four structured waves ($4,134) drops the homepage to 0.83, third of five and tied with the other two. What replicates is the floor, not the peak. 2. "All ten cells… the homepage won and the collection pages lost" reframed to "all five lost money." By the plan's own decision metric every destination came back negative; the best was −$11.55 per purchase. 3. "Nothing in the knowledge bases has read it" narrowed. True of the knowledge bases, false of the workspace: the 2026-08-27 reconciliation document pruned seven arms on partial windows. 4. "The single best-performing ad in your account" corrected to "the highest-revenue ad." Life's Too Short returned 2.67 against its 2.04. 5. "Instead of running a third wave blind" corrected. Three further waves (CW34, CW35, CW36) launched after CW33 began; the one still running is the fourth. 6. "More than six times any product page" corrected to "nearly six times" — 13,845 against 2,325 is 5.96×. 7. Move's format instruction de-contradicted. The hunt prescribed carousel while naming a single-image ad as the shape to copy. The card now reports the split honestly rather than resolving it by assertion. Thread: Lutke on work happening where attention is not → the mastermind's catalog pipeline doctrine the test was built from → the catalog-ads playbook's standing IGL prescription → the ten CW33 cells in fact_meta_daily`, finished and unread → the June landing-page rule the result contradicts → the Skool lesson arguing against launching the next wave first.