Harvesting and negating search terms

A keyword is what you bid on. A search term is what a shopper actually typed. On exact match they are the same; everywhere else they are not, and the gap between them is where most of the money in an Amazon account is found or lost.

Broad keywords, phrase keywords and auto campaigns all match many search terms. The Search Term report shows you which ones, and how each performed. Two questions follow: which of these deserve their own keyword, and which should never match again.

Harvesting: give the winners their own line

When a search term converts repeatedly inside a broad or auto campaign, it is being priced by an average — the bid on the keyword that matched it, which is also serving a dozen other queries with different economics. That average is almost certainly wrong for it.

Promoting it to its own exact-match keyword lets you bid it deliberately. That is a harvest.

The bar should be orders and clicks, not orders alone. A single click that happened to convert produces spectacular apparent metrics — a 100% conversion rate and an implied bid ceiling that would bankrupt you. Requiring a minimum click count as well as a minimum order count filters these out.

Set the opening bid from what it actually cost. The term already told you what winning the order costs: its realized CPC over the period. Starting from Amazon's suggested bid, or from a round number, discards information you already have.

One adjustment. Generic terms — "cutting board" — tend to convert at low bids and behave badly when bid up. Starting them at a haircut off their earned CPC means they have to earn a raise rather than begin at the ceiling. Specific terms — "end grain walnut cutting board 18 inch" — can start closer to what they proved.

The half nobody does

Harvest a term and it now exists in two places: your new exact keyword, and the auto or broad campaign that discovered it. Both are eligible to serve. You are now bidding against yourself, and Amazon is happy to let you.

So a harvest is not one change, it is a pair: create the keyword, and block the term at its source. Approving half of that pair is worse than approving neither.

There is a sequencing subtlety. Block the source before the new keyword is actually serving and you have a window where the term is blocked in one place and not yet live in the other — you lose the traffic entirely. The block should wait for proof the new keyword is serving, and expire harmlessly if it never does.

Negating: the one you cannot undo

Everything else in Amazon advertising is reversible. Raise a bid, lower it back. Pause a campaign, resume it. Negating is different: the sales that term would have made simply do not happen, and you will never see them in a report to know what you lost.

That asymmetry should govern how you approach it.

Wait for enough clicks — measured against your own conversion rate. At a 10% conversion rate, five clicks with no orders is unremarkable; you would expect roughly one order per ten clicks, so five clicks of nothing is ordinary variance. Forty clicks with no orders is a real signal.

If each click independently had a 10% chance of an order, the probability of zero orders after ten clicks would be 0.9^10, about 34.9%. That is not the probability that the term is a dud. Traffic quality varies, attribution can lag, and a posterior claim needs additional assumptions. Use the calculation as context, not an automatic cutoff.

Use exact match for ordinary performance negatives. A negative phrase blocks every query containing it. Negate the phrase "cast iron" and you have also blocked "cast iron skillet," "pre seasoned cast iron pan," and everything else in that family. Configured irrelevant-term classes can propose phrase negatives using the wording and match type chosen for that class. Review the actual proposal's match type and scope, and any collision with a live positive, before approval or export. Do not turn an exact negative into a phrase negative simply to make a reusable library shorter.

The family check

This is the mistake that quietly damages accounts, and it is worth understanding properly.

"Cast iron skillet," "cast iron skillets," and "cast-iron skillet" are one thing to a shopper and three separate rows in your report. Each row has its own click and order counts. It is entirely normal for one variant to show twenty clicks and zero orders while another shows eight clicks and three orders.

Judge that first row in isolation and you will negate it — and because negatives at the ad group or campaign level apply broadly, you may have just cut off the whole family.

The rule that prevents this: before negating any term, assemble every variant of it across the entire account and check whether any of them has orders. If any does, do not negate any of them.

Not "flag it for review." Not "warn the user." Block it. A warning is something a person clicks through at 11pm while sweeping a queue of sixty rows.

Most tools in this category warn you about this hazard in their documentation and leave the judgment to you. Preventing it is better than describing it.

Junk is the safe category

Not all negatives carry equal risk. "Cast iron radiator" matching a skillet campaign is off-category — there is no plausible path to a sale, and no family to protect.

These are worth identifying separately, because they are the highest-confidence, lowest- risk negatives available. They also cluster: a sudden burst of them usually means an auto campaign's matching has loosened, which is worth knowing about on its own.

The family check still runs on them. There is just rarely anything for it to find.

The loop

Put together, the routine is:

1. Auto and broad campaigns discover what shoppers type.
2. Terms that convert repeatedly are harvested into exact, priced from what they

earned, and blocked at their source.

1. Terms that spend without converting — past an evidence bar tied to your economics, and

clear of the family check — are negated.

1. The discovery campaigns get cleaner, and next month's report contains new terms rather

than the same expensive mistakes.

An account that discovers but never harvests keeps paying average prices for above-average queries. An account that harvests but never negates keeps paying for the same duds forever. An account that does neither is running an expensive, permanent experiment it never reads the results of.

What AutoPPC does with this

Terms are classified generic or specific, and harvest bars differ accordingly. Harvests ship paired with their source block as a single approval unit. Opening bids come from realized CPC, clamped to your floor and cap, with a haircut on generics.

Negatives require the click bar derived from your conversion rate, default to exact, and pass the family check — which is an Invariant, not a setting. Brand terms and protected families are never eligible. Your own ASINs are excluded from harvesting unless you opt them in deliberately.

In AutoPPC, the baseline zero-order click bar uses target ACOS, AOV and CPC with a minimum floor; conversion-family and other checks also apply. See the audit checklist.