The Pause Rulebook's concepts, translated to search terms: when a query becomes a negative keyword, when it never can, and how the "usa map" problem is handled. A negative is more permanent than a pause, so the bar is higher and the review loop is stricter.
A paused product type keeps generating data around it, expires after 90 days, and reverts on its own. A negative keyword does none of that. The moment it is applied, the traffic stops and so does the data: there is no report showing what a negative blocked or what it would have earned. You cannot re test a negative without removing it. So three things change:
A search query can declare its intent in words, which a product never can. Intent routing happens before any performance math, in both directions.
Cheap high volume traffic makes conversion rate meaningless. Every verdict is ACoS: what it cost against what it returned.
Because a negative silences its own evidence, the whole list is re justified quarterly against a 12 month window, and negatives that no longer hold are removed.
Anyone typing "bulk" or "wholesale" is someone we want in front of, full stop. The interesting work is everything else.
| Lane | Signals | What happens | The evidence at WSD, 12 months |
|---|---|---|---|
| Declared for us | bulk, wholesale, case of, pallet, reseller, distributor, liquidation, piece counts | Protected. Never negated, at any performance. A qualified term that loses money is an Investigate item: a landing page, pricing, or bidding problem. Blocking declared B2B demand treats our conversion problem as the buyer's fault. | 8,658 terms · $49,864 57% of search spend 2.87% CVR · 15% ACoS |
| Declared against us | brands we do not stock, freebie intent (free, printable, template, diy, clipart, how to) | The fast lane. The only relevance based negatives. Eligible at a $5 per year floor because the query names the wrong buyer in its own words. Every candidate still passes the guardrails below. | WSD test run: 5 phrases · ~$7/mo 5 held by guardrails |
| Marketplace names | amazon, walmart, temu, dollar tree, five below, costco, etsy and the rest | The evidence gates, same as ambiguous. Moved out of the fast lane on the owner's call: a person searching a marketplace can still be our buyer comparison shopping, so marketplace names must prove they lose money before they are blocked. The first test run proved the point, see below. | WSD test run: 0 of 27 qualified to block dollar tree converts at 31% ACoS |
| Brands we stock | hanes, fruit of the loom, gildan, disney and the rest of the carried brands | The evidence gates, monthly. Stocked no longer means untouchable, on the owner's call: a brand we carry still has to earn its spend. Failing brands usually route to Investigate rather than a block, because a brand phrase would also silence its bulk and wholesale searches. | WSD test: hanes left alone at 25% ACoS fruit of the loom flagged at 57% |
| Ambiguous | everything else: "socks", "backpacks", "usa map", terms with no qualifier either way | The evidence lanes below. This is the pool where a general consumer and a business buyer type the same words. No single query here is ever judged alone. | 23,969 terms · $33,479 1.31% CVR · 14% ACoS |
The variations trap: some versions of a query convert, most do not. Here is what that looks like for real.
Judged one query at a time, nearly the entire map cluster looks dead and gets negated. Judged as what it actually is, one pool of near free traffic hitting a placemat product that schools buy by the case, it is the cheapest revenue in the account. The same trap runs in reverse: a phrase can look fine on its converting variations while the negative you write would also block them. Two rules follow:
A phrase negative blocks every query containing that word sequence. So the thing we evaluate is never a query, it is the entire pool of traffic the proposed negative would block, pulled from 12 months of terms plus the PMax export. If the pool earns its keep, there is no negative, no matter how many individual variations sit at zero.
The map cluster converts at 0.11%. It is also the best ROAS in the account, because the clicks cost a cent and the orders are $600+ case purchases. Conversion rate is not a verdict on anything. Cluster ACoS is the only performance test a negative candidate faces.
The exact match scalpel. There is a second way past "can't carve", because negative keywords do not expand to close variants the way normal keywords do: an exact match negative [backpacks] blocks only that literal query (plus misspellings), never "bulk backpacks", "backpacks wholesale", or even the singular "backpack". So when a phrase pool is protected by its converters, the bare query can still be judged on its own evidence and blocked alone if it fails. Plural and singular are separate negatives and are listed as a pair. The reach is limited by design: it removes one query's spend, not the pool's, and at this account the failing pools live almost entirely in the long tail. The bare query "school" is $5 of its $1,206 pool.
The carve out, translated. Products get carved out of a type exclusion; queries get carved out by narrowing the phrase. If "usa map" as a pool failed, we would not negate "usa map", we would look for the word that separates the failing traffic from the converting traffic ("printable usa map", "usa map tattoo") and negate only that. And if no word sequence separates them, there is no negative: traffic that cannot be carved cleanly is a bidding and landing page problem, not a blocking problem. Can't carve, can't cut.
Every candidate phrase, always evaluated as the full pool it would block. First exit wins.
| Rule | Threshold | Why |
|---|---|---|
| Fast lane floor | $5 / 12 months | Intent is already declared against us; the floor just keeps the list from filling with one click noise. |
| Evidence floor, ambiguous pool | spend ≥ 2x cost per order, plus one of: clicks ≥ 3 expected orders · spend ≥ 3x cost per order · 3+ orders landed | Three ways to earn a verdict. The click path suits cheap click pools; the spend path catches expensive click pools that would never reach the click count (a $3 per click pool with zero orders on $130+ is not luck); three landed orders make the pool's ACoS itself trustworthy. At WSD the numbers are $87, 287 clicks, $130. |
| Losing money | pool ACoS above 35% | Same gate as the pause rules. Judged on the whole pool the phrase would block. Conversion rate is never consulted. |
| Fine last year | LY ACoS ≤ 35% on $50+ | A pool that worked in this window last year is an Investigate item, not a negative. |
| Confirmation | two consecutive monthly runs | Because a negative cannot be re tested without removing it, nothing in the ambiguous pool is blocked on one reading. |
| Peak guard | no new negative within 60 days of the matched category's peak | Seasonal search demand gets the same protection seasonal products get. |
| Converter protection | AOV ≥ $100, or any bulk/wholesale qualifier | The standing owner rule: a phrase is rejected if it would block any converter meeting either test. |
"usa map" blocks every query containing those two words in that order, including ones that convert. Every candidate is tested pessimistically, plural and possessive aware, against 12 months of terms plus the PMax export, for everything it could block. Pessimistic about what it blocks, conservative about what it claims to save.
The feed decides which lane a brand takes, not whether it is safe. A brand we do not carry rides the fast lane; a brand we stock goes through the evidence gates monthly, because carrying Hanes is not a reason to fund Hanes searches that never buy. A stocked brand block carries an extra flag: blocking "hanes" would also block "hanes wholesale", so the review always shows the declared B2B sub pool a brand negative would silence. Failing stocked brands usually route to Investigate, and any brand negative that does ship is re argued at the quarterly review like every other. The feed check re runs monthly because stocking changes.
Some words mean two things: "staples" is a retailer and a thing both stores sell. A bare token like that is never negated; only a longer phrase that pins the meaning ("staples school supplies") can be. When the junk traffic comes from a badly worded product title, it goes to the title program, not the negatives list.
High volume on brief terms that convert poorly ("bulk t shirts", "backpacks") are demand with a landing page or bidding problem; blocking them is self harm. And suppressing retail sized orders is a job for order minimums and value rules, not keywords.
Run 1 Sep 2026 over Sep 2025 to Aug 2026: $86,380 of visible search terms, 354,803 clicks, 1,989 orders. Bars computed live: cost per order $43, so the spend floor is $87 and the click floor is 287.
Coverage, stated honestly: Google's API only reports search terms with cost for Shopping and Search campaigns, and Performance Max is 72% of this account's spend ($427K of $595K in the window). So this sweep saw about 15% of the account's traffic directly. PMax was checked through its category labels, which carry conversions but no cost and hide half the detail in a blank rollup. That check already pulled two proposals back (below). Every remaining proposal is provisional until it clears a PMax search terms export from the Ads interface, which carries the full data the API withholds. The 17 PMax campaigns at $5K+ cover 92% of PMax spend and are the export worth doing.
| Result | What the gates found | Value |
|---|---|---|
| Propose, fast lane | 7 phrases, all zero conversions on visible spend: goodfellow (Target's house brand, the biggest at $125), levis, george, calvin klein, burlington, how to, carhartt. Unstocked brands and freebie intent only. Provisional until the PMax export clears them. | $221 / 12mo · ~$18/mo |
| Guardrails held | nike and adidas stay: their pools contain $187 and $1,506 of protected conversions (bulk buyers search branded socks); "free" stays for the same reason. And the PMax category check pulled back two phrases the visible data had cleared: crocs converts in PMax ($3,905) and so does coloring ($1,432). Stocked brands are no longer auto protected; their verdicts are in the row below. | $7,146 protected |
| Stocked brands: measured, not shielded | 11 carried brands went through the gates instead of the old blanket protection. hanes earns its place on evidence: $673 at 25% ACoS, left alone. fruit of the loom ($754 at 57% ACoS) and gildan ($190 at 41%) cleared the floors and failed the money gate, but both are held by guardrails, so they go to Investigate (price, feed, bidding) rather than the block list. The rest sit under the evidence floors and watch. | $2,642 hanes revenue kept 2 brands to investigate |
| Marketplaces: zero blocks | 27 marketplace names checked through the evidence gates. None qualified. dollar tree cleared both floors and exits "left alone" at 31% ACoS with $354 of visible orders, plus another $3,498 in PMax; walmart cleared the floors at 88% ACoS but its pool contains protected converters, so it is held for narrowing rather than blocked; everything else is below the evidence floors and watches. | 0 of 27 blocked |
| Exact match lane: two flags, both excused | Only two bare queries in the whole account clear the evidence bars at over 35% ACoS, and neither becomes a negative: [school supplies] ($260 at 242%) converted a $100+ case order and ran at 12% last year, and [socks] ($256 at 104%) ran at 14% last year. Both are Investigate items: the head queries broke this year, they are not bad queries. If a head query fails again next year with its excuse gone, the exact lane blocks it alone while every qualified variant keeps serving. | 0 blocked · both to investigate |
| Ambiguous pool: one flag | Every one, two and three word pool in the unqualified queries was measured: 238 pools clear the spend floor and about 90 run above 35% ACoS, so the waste is real. But nearly every failing pool is a generic word (school, backpacks, wool, black, cheap) whose traffic also contains bulk and wholesale converters, and no word sequence separates the two. Can't carve, can't cut: that spend is a bidding and landing page problem, and it echoes the traffic quality finding from the pause runs. The clean exits: "nike tech" converted fine last year and goes to Investigate, and the school cluster ($1,206 at 512% ACoS, with "school supplies" at 584%) is an investigate grade collapse that matches the school supply kits finding in the September pause run. | 0 blocked · 2 investigations opened |
What the test says about this account: Wholesale Sock Deals has almost no negatable waste in its search terms. The gates' biggest output was protection, not blocks: the marketplace change alone kept dollar tree and walmart traffic alive that a relevance based sweep would have negated. The account's real search problem remains the traffic quality story in the pause runs, not keywords.
The answer to "how often": candidates monthly, blocks confirmed a month later, the whole list re argued quarterly.
Runs with the pause run, same trailing window. Two sweeps, both needed: the wrong buyer lexicon, and off catalog relevance. Ambiguous pools that fail gates 4 to 6 go on the confirmation list.
Fast lane phrases apply after your review. Ambiguous pool phrases apply only if this is their second consecutive failing run. Everything lands in the account level negative list, which covers Shopping and Performance Max together.
Every negative is re justified against a fresh 12 month window, and removed if it no longer holds. This is not hypothetical: the year view already rescued "five below" and "dollar tree", both of which converted outside the 90 day window that nominated them.
The September and October reviews specifically clear the list against holiday demand, the same 60 day peak guard the pause rules use.
Same standing rule as the pause runs: nothing is blocked before you have seen it. Each month's proposed negatives arrive as a short list with the pool each phrase would block, what it spent, and what it returned. The running All Time Trading list, 27 phrases worth about $151 a month, is built exactly this way and is still awaiting application; the Wholesale Sock Deals test sweep above proposes 5 provisional fast lane phrases and nothing else.