Keyword Harvesting: Promote Winning Search Terms Without Paying Twice

Automatic campaigns are a cheap way to let Amazon show you what shoppers actually type. They are a poor place to keep running the winners, because you cannot set a bid for an individual search term there. Keyword harvesting is the routine that fixes this: you read the search term report, pick the terms with proof behind them, and move them into a manual campaign where you control the bid. The routine has two failure modes. One is promoting on thin evidence. The other is leaving the term running in several places, so your own campaigns compete for the same click.

What the search term report gives you

The Sponsored Products search term report lists the customer search terms that triggered your ads, which of your targets matched them, and the clicks, spend, orders and sales that followed. Some rows are product ASINs rather than words, which are shoppers arriving from another product's page. Treat those the same way, as candidates for a product targeting campaign.

Pull a window long enough for a term to have a fair chance of earning orders, and then combine the same term across campaigns in a spreadsheet so that you judge it on its full record, not on a weekly slice. Leave the most recent days out, because orders are credited for a while after the click and the newest rows always look worse than they will end up.

Remember too why the term exists in the report at all: Amazon chose to show your ad on it. That is evidence of demand and of relevance in Amazon's eyes, not of profit. Profit has to be checked, term by term.

The promotion rule: three tests

A term earns a place in a manual campaign when it passes all three of these tests.

• Relevant: a human reads the term and agrees it describes what you sell, in this size, material or pack count.
• Enough evidence: it has at least a set number of orders. Two is a reasonable minimum, and you can raise it for expensive products or when you have plenty of data.
• Profitable: its ACoS is at or below your target, with the target taken from the product's break-even ACoS.

Evidence deserves more thought than the order count alone. Suppose your product converts about 10% of its clicks, so you expect one order every ten clicks. A term with one order from 40 clicks is not a winner. It converts at 2.5%, far below your product's norm, and the single order was luck. A term with two orders from 20 clicks is converting in line with expectation and deserves attention, while one with no orders after 40 clicks is a negative keyword candidate, not a harvest candidate.

Match type and starting bid

Promote proven terms as exact match. The point of harvesting is control, and exact match is the type that gives the most of it. Phrase match can sit in a separate discovery campaign to catch variations of a winning term, as long as you treat its results as new evidence, not as proof carried over from the exact version.

For the starting bid, work out a ceiling from numbers you already have: price times conversion rate times target ACoS. For a hypothetical $40 product converting at 10% with a 25% target, that is $40 times 0.10 times 0.25, or $1.00. Compare it with the average cost per click the term paid in the automatic campaign. Start at the lower of the two. If the term loses impressions after the move, raise the bid in small steps. If it spends freely, lower it.

Two special cases need their own home. Your own brand terms usually convert at a very different rate from generic terms, so promote them into a brand campaign and report them separately. Left mixed in, they flatter the ACoS of the whole campaign. Product ASIN rows go into a product targeting campaign, where you bid on another product's page. Apply the same three tests there, plus a check that the other product is a fair neighbor on price and function.

Avoid paying against yourself

When two of your campaigns could serve the same search, Amazon chooses which one competes, and that choice is neither visible in advance nor something you control. You end up with split data, an unclear picture of what any one bid is doing and, sometimes, two of your own ads fighting over one placement. The fix is simple: every term has exactly one owner.

When you promote a term, add it as a negative exact keyword in the source campaign, and in any discovery campaign that could also catch it. This sends that exact search to the manual campaign. Close variations can still reach the automatic campaign, which is what you want, because the automatic campaign keeps scouting. Do this in the right order: create the manual keyword and confirm it is enabled and serving, then add the negative. If you negate first, the term has nowhere to go for a while.

Keep a harvest log with one row per promoted term: the date, the term, the source campaign, the orders and ACoS at promotion, the destination campaign and the starting bid. It doubles as your duplicate check, because before adding any keyword you can search the log and see whether another campaign owns it.

Check what you promoted

A term was picked because it did well, and things that were picked for doing well tend to do a little worse next time. This is regression to the mean, and it is worth expecting. Look at each promoted term again after a few weeks, and do not judge it on the first week. Keep it if its ACoS is near target. Adjust the bid if it is clearly too high or too low. Pause it and note why if it has gone cold. Your log makes this review quick, since the starting numbers are already written down.

Run the whole routine on a steady schedule, weekly or every second week, and use the same report for your negative keyword review. One pass through the report gives you both the terms to promote and the terms to block.

Where to start

• Pull a search term report long enough to give terms a fair chance, and leave the newest days out.
• Combine each term across campaigns before you judge it.
• Apply the three tests: relevant, enough orders, ACoS at or below target.
• Promote winners as exact match, with a starting bid at or below the ceiling.
• Add each promoted term as a negative exact in its source campaign once the new keyword is serving.
• Log every promotion and review each one a few weeks later.