US Non-Brand Search — NEGATIVE KEYWORD LIST (4wk: Jun 22 – Jul 19, 2026)

Built from search_term_view, all Search_USA_Non_Brand_* Search campaigns, account 9105682234.
Pulled 2026-07-27. Judged on L2 subs (metrics.conversions == L2 subs on non-brand).

Coverage caveat — read first

Google suppresses low-volume search terms, so ~35% of spend has no term attribution. Every dollar
figure below is a floor — scale up by ~1.5× for the true blocked spend.

The shape of the waste: it is a long tail, not a few big terms

Only 12 terms spent >$50 with zero subs. A term-by-term exact-negative list is therefore the wrong
instrument — it would block $994 and leave the rest untouched. The list below is built as phrase-match
thematic negatives that sweep the tail, plus a short exact list for the individual offenders.

⚠️ Don't over-read "$11,342 on zero-sub terms." Most of those terms spent $2–5, i.e. too little to have
expected a sub at a $151 blended cost/sub. The buckets below are the ones that are statistically
dead or provably expensive, not merely small.

Intent buckets — what actually converts

⚠️ Two standing content-rule assumptions are FALSIFIED by this window

The brief's rule was "free / test / quiz / IQ / kids / puzzle intent piles up signups but ~0 subs."
Over these 4 weeks, half of that is wrong:

• "Free" DOES subscribe: $1,128 → 10 subs → $113/sub, better than the $151 blended average and
  better than 9 of the 16 campaigns. Converting free terms include free brain games (2 subs, $120),
  cognitive speed training games online free (1 sub, $102), brain training apps free,
  free brain training games online, free cognitive training exercises. A blanket free negative
  would cut a top-decile bucket — it is the freemium funnel working as designed.
• "Puzzle" DOES subscribe: $402 → 4 subs → $101/sub, the best bucket in the table.
• Test / quiz / IQ is confirmed dead ($354 / 1 sub), and kids is confirmed dead but trivially
  small ($13).

The real waste pattern this window is INFORMATIONAL QUERY SYNTAX, not topic. Somebody searching
"how to improve my memory" wants an article; somebody searching "brain games" wants a product —
even when they append "free". Token-level evidence:

your at $606/sub over 211 terms is the single strongest negative signal in the dataset.

Where it lands, by campaign

The two Broad Exercises campaigns carry 60% of it ($1,399 of $2,329) — and 100% of the competitor,
brand-misspell and medical leakage. Confirms the brief's hypothesis for those buckets. But note the
howto bucket is account-wide ($1,147 spread across 9 campaigns, only 41% in Broad) — Memory_Exact_Mobile
alone leaks $214 of it, and even ⭐Brain_Training_Exact_Desktop leaks $146. Informational negatives
belong on a shared account-level list, not just on Broad.

THE LIST

A. Account-level shared negative list — NB_Informational_Intent

Apply to all Search_USA_Non_Brand_* Search campaigns. Blocks ~$1,147/4wk @ $255/sub.

⚠️ Do NOT add bare your as a negative despite its $606/sub — it collides with legitimate
product queries (train your brain, games to keep your mind sharp). The phrase negatives above
capture most your spend via the how-to/ways-to constructions that carry it. Re-measure your
after 4 weeks and only then consider tightening.

B. Account-level shared negative list — NB_Competitor

Apply to all non-brand Search campaigns. Blocks ~$517/4wk @ $172/sub.
Flagged as missing since 2026-06-25 and still unapplied.

C. Account-level shared negative list — NB_Test_Quiz_IQ

Apply to all non-brand Search campaigns except Tests_Quizzes_* (which should just be paused —
see D). Blocks ~$354/4wk @ $354/sub. This bucket is confirmed dead per the account content rule.

D. Account-level shared negative list — NB_Medical

Apply to all non-brand Search campaigns. Blocks ~$223/4wk @ $223/sub.
Also an FTC-risk hygiene win — we should not be buying dementia-prevention intent given the 2016
settlement (see openai-ads-setup.md guardrails).

E. Account-level shared negative list — NB_Junk

Apply to all non-brand Search campaigns. Blocks ~$37/4wk. Small, but zero-sub and permanent.

F. Brand-misspell negatives — non-brand campaigns only

Blocks $51/4wk. These are brand queries being bought at non-brand CPCs (~$2.50–3.00 vs brand's
~$0.20). Add as negatives to non-brand and confirm the brand campaigns' misspell coverage catches them.

G. Exact negatives — individual offenders

Terms with real money and zero subs that the phrase lists above do not catch. Add as exact
negatives to the named campaign only.

Section G subtotal: $930/4wk. Note 6 of 11 sit in Brain_Games_Exact_Mobile / Brain_Games_Adults
— that ad group is the account's biggest single spender and needs a keyword-level prune, not just
negatives.

Expected impact

Reallocating that $2,329 into the campaigns that work — Memory_Exact_Desktop ($79/sub),
Brain_Training_Exact_Mobile ($91), Brain_Training_Exact_Desktop ($105) — buys roughly 24–29 subs
instead of 9.5, at unchanged spend. Blended cost/sub moves from $151 → ~$137, cash ROAS from
0.74× → ~0.81×.

That is a real gain but it does not reach the ~$100 profit line on its own. The binding constraint
is the $250 ad-group tCPA (CPC +49% over 4 weeks against −5% impressions). Negatives are necessary
and cheap; the tCPA reset is the actual fix.

BUDGET REALLOCATION — 5 actionables

Ranked by confidence × size. All figures are the 4-week aggregate (Jun 22 – Jul 19), never a
single week.

⚠️ SUPERSEDED IN PART — updated 2026-07-27 after pulling W9 (Jul 20–26)

Actionable #1 below is moot: the two Broad Exercises campaigns collapsed on their own starting
Jul 16 (Mobile −98%, Desktop −88% by W9) — still ENABLED, budget-lost IS = 0, rank-lost IS
77%→90% ⇒ Smart Bidding withdrew at the $250 tCPA. Replace "cap them" with "pause or rebuild
them" (they will otherwise ramp and collapse a third time). Knock-on effects for this document:

• ~60% of the negative-list spend ($1,399/4wk) sat in those two campaigns and has already stopped
  flowing. The realistic live prize is now ~$930/4wk, not $2,329.
• List A (informational syntax) is now the main prize — it was only 41% Broad, and all nine
  campaigns that leak it are still spending.
• Lists B (competitor), D (medical) and F (brand-misspell) were almost entirely Broad-fed — still
  worth adding as permanent hygiene, but expect little immediate saving.
• Section G (exact offenders) is unaffected — none of it is in Broad; 6 of 11 are in
  Brain_Games_Adults, now 48% of all non-brand spend.

See weekly/week-09_2026-07-20.md. The intent-bucket economics and the
content-rule corrections above are unchanged.

1. 🛑 Cap the two Broad Exercises campaigns to ~50% → PAUSE OR REBUILD — self-collapsed Jul 16

Exercises_BrainHealth_Broad_v2_Mobile ($5,521 / 23.5 subs / $235) and _Broad_Desktop
($2,405 / 11 subs / $219) are 32% of non-brand spend at 1.5× the account average and ~2.3× the
profit line. They also carry 100% of competitor, brand-misspell and medical leakage. Broad is
structurally the problem: Broad $228/sub vs Exact $140/sub across the window.
Action: halve budgets, apply lists A–F, and re-read in 2 weeks. Do not pause outright — W8 showed
$144/$174, so there is a workable core once the leakage is negated.

2. ⭐ Scale Brain_Training — both devices — currently 8.8% of spend

Brain_Training_Exact_Desktop $1,470 / 14 subs / $105 and Brain_Training_Exact_Mobile
$730 / 8 subs / $91. Mobile posted 7 subs at $60/sub in W8 alone off a $423 spend — the
strongest single campaign-week in the set, from a campaign that had produced 1 sub in the prior three
weeks combined. Both sit near the profit line with rank-lost headroom.
Action: +50–100% budget on both, funded from #1. Caveat: Brain_Training_Exact_Mobile's 4wk base
is 8 subs — treat the W8 breakout as promising, not proven. Re-read after 2 weeks before doubling again.

3. ✅ Scale Memory_Exact_Desktop — $713/4wk at $79/sub, 2.8% of spend

The best cost/sub in the account and the only core campaign never raised to the $250 tCPA (still
$150 — per the 2026-07-12 finding, that is precisely why it works). It is starved.
Action: +100% budget. Do not raise its tCPA — that setting is the reason it converts.
Its mobile sibling (Memory_Exact_Mobile, $166/sub, $250 tCPA) is the counter-example.

4. 🛑 Pause the 0-sub burners — $1,130/4wk, 0–4.5 subs

• Tests_Quizzes_Exact_Desktop — $206 / 0 subs, still live. Its mobile sibling was flagged for
  pause at $836/sub back on 2026-06-25 and this one was never actioned. Pause.
• Exercises_BrainHealth_Exact_Desktop — $916 / 4.5 subs / $204, and $312 / 0 subs in W8.
  Degrading (8wk was $168). $121 of its spend is howto-informational.
• Trivial-spend zombies to clean up: Puzzles_Word_Math_Mobile, Apps_Broad_Mobile,
  Exercises_Exact_Mobile, BrainHealth_Generic_Exact_Mobile (~$16 combined).

5. ⚠️ Prune Brain_Games_Exact_Mobile / Brain_Games_Adults rather than cutting the campaign

It is the biggest campaign ($7,385 / 29% of spend) at a fair 4wk $134/sub — so do not slash
it on the W8 view ($206), exactly as the brief warned. But its CPC rose $2.01 → $3.56 (+77%) across
the window and it holds 6 of the 11 exact negatives in section G ($486/4wk of 0-sub head terms).
Action: apply the section-G exact negatives and cut bids on the surviving 0-sub keywords in the
Brain_Games_Adults ad group; leave the budget alone.

Root cause behind all five

The $250 ad-group tCPA raised in early July + the shared $1,180/day "V2 Campaigns" budget.
Four weeks of CPC +49% against impressions −5% is the anti-scale signature: we bought dearer clicks in
the same auctions, not more auctions. Cash ROAS fell 0.92 → 0.90 → 0.71 → 0.62 in lockstep.
Recommend reverting core ad groups to ~$150 (the level Memory_Exact_Desktop still runs at, and the
only campaign hitting the profit line). Negatives + reallocation are worth ~$14/sub; the tCPA is worth
the rest.

Sources: search_term_view 4wk pull 2026-07-27 (2,981 rows, 2,527 distinct terms) + campaign weekly
grid in weekly/week-08_2026-07-13.md. Methodology per
../conversion-events-reference.md.