Most sellers cannibalize their own ads when harvesting Amazon keywords from auto to exact. The 3-step funnel that avoids it: promote, negate, isolate.
Most harvesting guides will sell you the same workflow: run auto for a few weeks, promote winners to broad, then move broad winners to exact. That funnel sounds clean. It's also one of the more expensive ways to waste money on Amazon. I've been running my own seller account (Rowdy Rooster Woodworks) and managing PPC for other operators long enough to see the same pattern over and over: the textbook funnel gets you into broad-match purgatory and bleeds budget while you wait for "data."
The version that actually works is shorter, more boring, and most of the controversy in this article is in the details people skip. This is the workflow I use on my own account and what we built our automation around.
Forget the funnel. Use two discovery sources.
The "auto → broad → exact" funnel is the most-repeated bad advice in Amazon PPC. Broad match is a tool, not a tier. Used as a middle stage it's essentially a blank check to Amazon: you give them a seed term and they decide what to match against. That can include relevance you don't want, search variants that don't convert, and competitor brand traffic you're not equipped to handle. The cost of "exploring broadly" gets paid out of your daily budget.
Where broad does belong is narrow and intentional — most often brand defense. A broad match on your own brand name (for me, "Rowdy Rooster") catches misspellings, plural variations, and adjacent queries you want to own without manually listing every variant. That's broad's job. It's not a discovery layer for the rest of the catalog.
For everything else, you have two discovery sources running in parallel:
-
Auto campaigns — Amazon matches your product to search queries based on your listing content. You don't pick the keywords; Amazon does, across four sub-targetings (close match, loose match, substitutes, complements). The output is a steady drip of real shopper queries.
-
Category / product targeting (PAT) — manual product targets at the ASIN or category level. These don't surface "search terms" the same way auto does; they surface competitor ASINs and category placements that your product shows alongside. Same idea, different lens. The winners from PAT also get promoted to exact match (or, in some cases, isolated as exact ASIN targets).
-
Exact match campaigns are where winners go to scale. One keyword per ad group, controlled bid, focused budget. This is the destination for everything you've validated upstream.
No middle tier. No broad-as-funnel. Two discovery sources, one performance destination.
The workflow
-
Always have a small auto campaign running. Not "run for 2-4 weeks and then stop." Always-on. Language drifts. Amazon's matching shifts. Seasonal queries appear out of nowhere. The harvest loop is most valuable as a long-running listener, not a one-time data collection exercise. A modest budget ($5-15/day depending on product) is enough to keep the discovery engine alive.
-
Pull the search term report on a cadence you can actually keep. Weekly is the right tempo for most sellers. Every two weeks is acceptable. Monthly is too slow — winners go stale and losers keep spending.
-
Filter to the right window. Use 14+ days of data, not 7. Amazon's attribution has lag — a term that looks great over 7 days can have a return wave hiding behind it. Two weeks is the minimum where the conversion numbers are trustworthy.
-
Promote terms that clear the bar (see thresholds below). Drop them into a dedicated exact-match campaign — never glue them into an existing ad group that already has other keywords. You'll lose the ability to evaluate them individually.
-
Decide what to do in the source: negate, lower bid, or leave alone. This is the most-contested step in PPC and the textbook answer is wrong. There's a full section on it below.
-
Negate the obvious losers. High spend, zero orders, real impressions — these are bleeding money. Negative exact most of the time. Negative phrase only when you genuinely want to kill an entire term family (covered below).
The negation question: it depends, and that matters
Most articles tell you to immediately negate the harvested term in the source campaign to prevent cannibalization. I've seen cannibalization first-hand, so I take the risk seriously. But this is not a universal rule — it depends on whether placement multipliers are active on the source campaign.
If your auto campaign uses placement multipliers (top of search, product pages, rest of search), the auto and exact campaigns aren't competing for the same eyeballs the way a static bid suggests. The auto might be winning top-of-search at an inflated multiplier while exact picks up the rest. Reflexively negating the term in auto can throw off a balance you've already tuned. If you have no multipliers, the cannibalization risk is more direct and negating tends to be cleaner.
For most operators without advanced multiplier tuning, the safer default is: negate the exact term in the source campaign after promotion. You stop bidding against yourself and your data on the term consolidates into one place.
For sellers running placement multipliers or those treating auto as an ongoing discovery layer they don't want to dampen, don't negate. Leave the term in source and let the discovery keep producing variants. The downside is some bid-against-yourself spending; the upside is you don't choke a working pipeline.
For broad targets specifically used for discovery (not the brand-defense case above) — don't negate. The whole point of that target is to keep finding things. The right control isn't negation, it's a human review layer between "I found something interesting" and "I'm promoting it." That's the loop our automation runs by default: candidates surface, you approve, then the action ships.
When is a term actually ready to harvest?
"Promote at 2 orders" is the floor — and 2 is barely enough. The reason it's a floor instead of a real threshold is that you observe it after promotion in the exact-match setting. The exact match campaign gives you isolated data on that term so you can decide quickly whether it earns its bid or gets pulled. That's a different role than the auto report can play.
The four signals I actually look at:
-
Orders ≥ 2 (the floor, not the goal): Filters out one-off flukes. Two independent conversions is the minimum for "this is a real pattern" — and even then, treat the promotion as a test that the exact-match campaign will confirm or reject.
-
Conversion rate ≥ your category baseline: If your other PPC keywords convert at 8%, a new term needs to at least match that. A term doing 4% CVR in an 8% category is a weak signal worth more data.
-
ACoS ≤ break-even ACoS: Break-even = (margin after fees ÷ price) × 100. Promoting an unprofitable term to exact match doesn't fix the math — it just gives you more control over an unprofitable bid. If you can't get it under break-even with a lower bid in exact, it's not a winner.
-
Lookback ≥ 14 days: Two-week minimum for trustworthy attribution. Returns lag, attribution lags, Q4 surges hide regression.
On a typical mid-size catalog this filter leaves you with 5-15 terms a week to act on — not 100. That's the right volume. Manual harvesting that produces hundreds of candidates per cycle is a sign the criteria are too loose.
Setting the starting bid in exact: depends what you're testing
Most guides give you one answer ("match the auto CPC, adjust slightly up"). That's only right if your goal is aggressive testing. There are two valid approaches and they correspond to different intents:
-
Ease in (the conservative path): Set the exact bid at the campaign default. Let the term find its own price through normal bidding. You'll lose some impression share at first but you'll never overpay. This is the right call when you're harvesting many terms at once and don't want to babysit each one.
-
Aggressively test (the validation path): Set the exact bid above the auto CPC — sometimes meaningfully above. You're saying "I want to see real performance data on this term as fast as possible." You'll spend more per click but you'll get a verdict in days instead of weeks.
The wrong answer is to pick a number arbitrarily and never revisit it. Bids on harvested terms should be cycling — small upward nudges if the term is showing potential but not winning impression share, small downward nudges if it's converting but overpaying. RedHen automates this with configurable bid nudges that cycle keywords at a rate you set — so winners get found faster and losers get pulled back without you watching dashboards.
The three most common mistakes
-
Treating the harvesting workflow as a one-time setup. The most common version of this: a seller runs an auto for 6 weeks, harvests once, shuts auto down, and walks away. Three months later they wonder why their PPC is stale. Auto needs to be always-on — at a small budget if you must — because shopper language and Amazon's matching both change continuously.
-
Defaulting to negative phrase. Negative exact is the right tool for almost every negation you'll do. Negative phrase is a big move — it blocks every variation containing the phrase. Save it for terms you genuinely hate and never want to see again (irrelevant brand competitor names, terms that have proven unprofitable across dozens of variants). Using negative phrase reflexively is how sellers accidentally kill profitable longtail traffic they didn't know they had.
-
Promoting on a single order. One order on a search term means almost nothing — could be impulse, could be a wrong-item click, could be a return waiting to happen. Two is the floor. Even at two, treat the promotion as a test that the exact campaign will validate. If the term doesn't continue to perform once isolated, pull it back.
Why this is painful at scale
Run this on five products and it's an hour a week. Run it on 50 products and it's an unwinnable game. Each product has its own auto, its own PAT targets, its own thresholds — and a lot of decisions sit in the gray zone where you genuinely need a human to look. Most sellers do the loop diligently for a few months and quietly stop. The cost of stopping doesn't show up as a single bad day; it shows up as a slow drift toward stale keywords and quiet waste.
RedHen runs this loop continuously. Search terms that clear your thresholds get queued for your review (we don't auto-apply changes you haven't approved), the negation strategy follows your account's actual configuration including placement multipliers, and the bid-nudge layer cycles bids on harvested terms so winners surface and losers get pulled without you watching dashboards. Same workflow, no spreadsheet hours. Try it free for 14 days.
.png){kind=small}
Partner