Amazon Search Term Report With AI: Find Wasted PPC Spend

Download the Sponsored Products search term report, total clicks, spend and orders per search term, and flag any term with 20 or more clicks and no orders. Give that list to an AI model to group by cause and draft negative keywords. You decide which to add, because a wrong negative removes sales silently.

The search term report shows what shoppers actually typed before your ad appeared, which is often not the keyword you bid on. That gap is where money disappears: broad matches, auto campaigns and product targeting pull in searches that look close and are not. The report is big, and reading ten thousand rows by eye is no one's idea of a good morning. AI is well suited to the sorting. It is poorly suited to the final call.

Follow me on Instagram@sagnikteaches

Getting the report out

Menu names move, so treat these as signposts. In the Amazon Ads console, guides place reports under Measurement and Reporting, then Sponsored ads reports, then Create report. Choose Sponsored Products as the ad type, Search term as the report type, a date range and a daily or summary time unit. Some guides place the same function in Seller Central under Reports, then Advertising Reports. Sponsored Brands has its own search term report. Amazon's reporting documentation adds two details worth knowing: the report only includes searches that produced at least one click, and where an ad ran on a product page with no search behind it, the search term shows as an asterisk.

Connect on LinkedInSagnik Bhattacharya

The look-back is short. Amazon's Ads reporting documentation lists 65 days of search term data for Sponsored Products and 60 days for Sponsored Brands, and a single report can cover at most 31 days. Pull the longest window your console offers, and then start saving a copy every month, so that you have a year of data when you need to look at a seasonal product.

Subscribe on YouTube@sagnikteaches

The fields you need, whatever Amazon calls them in your export: campaign, ad group, targeting (your keyword or product target), match type, customer search term, impressions, clicks, spend, orders and sales (note which attribution window your export uses), and ACOS or ROAS. Drop the rest.

Set your own threshold before you look

The 20-click line in this tutorial is a convention, not a rule. Better to derive yours. Say a cleaning company sells a $24 carpet-cleaning solution with a typical conversion rate of 1 order per 12 clicks, and an average cost per click of $0.65. A normal order therefore costs about $7.80 in ads (12 clicks at $0.65). Its margin before ads is $9.50 a unit. The ad cost of an average order is below margin, so the account makes about $1.70 a unit from ads. A search term that has taken 24 clicks with no order has already used $15.60, double what a typical order costs, and the chance of a sale is falling. That is the point at which to look. If instead the product converts 1 in 40 clicks, 24 clicks of no orders is still within normal luck, and the threshold should be around 60.

A rule that scales: set the click threshold at roughly two to three times the number of clicks you normally need for an order, or at the point where the spend on the term exceeds your profit per unit, whichever comes first. Write that rule in the sheet and let it set the number, not habit. One caution on small samples: with only a few clicks, any rate can look terrible by chance, which is why a click floor exists.

Do the first cut in the spreadsheet

A spreadsheet pivot is faster and safer than a chatbot for the arithmetic. Make a pivot with customer search term as rows and, as values, the sum of clicks, spend and orders. Sort by spend. Add a column for "flag": clicks at or above your threshold and orders equal to zero. Hide the rest. If you are not sure how to do this, the data-cleaning tutorial walks through tidying a messy export.

Two checks before you send anything to AI. The same search term may appear in several campaigns, so total across campaigns first; otherwise a term with 12 clicks in each of three campaigns looks innocent. And look at orders carefully: an attribution window means a click today may still become an order next week, so leave the most recent few days out of the file.

What the flagged list looks like

A trimmed example for the carpet solution (illustrative figures).

You can already see the problem and the temptation. The first line is a hire search, wrong intent. The second names a machine your solution may not suit. The third is a different product. The fourth wants something free. The fifth and sixth might be good terms that did not convert, perhaps because the price or the listing is wrong. Sorting these by cause is the model's job.

A prompt that groups and drafts, without deciding

I sell a 1-litre carpet cleaning solution for use in standard household machines, priced at $24. Below is a list of search terms from my Amazon Sponsored Products search term report that each had [20 or more] clicks and no orders.

For each term, classify the likely reason into exactly one of: WRONG PRODUCT (shopper wants something I do not sell), WRONG INTENT (hire, free, repair, DIY recipe), WRONG FIT (a specific machine or brand I may not suit), POSSIBLY GOOD (relevant but may need a better listing, price or bid), UNSURE.
Then draft negative keywords, using negative exact for single terms and negative phrase only when several terms share a phrase that is clearly wrong. For each negative, show the match type and give a one-line reason.
Do not include any POSSIBLY GOOD or UNSURE term in the negative list. List those separately with one suggested check each.
Do not guess figures. Do not claim what Amazon will do.

TERMS (term | clicks | spend | orders):
[paste]

The separation of "possibly good" is the point of the prompt. A blunt "negate everything with no orders" strategy loses terms that sell on another day.

A sample result, and what to fix

NEGATIVE EXACT: carpet cleaning machine hire (WRONG INTENT: shopper wants to rent a machine)
NEGATIVE EXACT: upholstery spot remover pen (WRONG PRODUCT)
NEGATIVE PHRASE: free carpet cleaner (WRONG INTENT: free)
NEGATIVE EXACT: carpet cleaner solution for hoover brand x (WRONG FIT: check compatibility first)
POSSIBLY GOOD: cheap carpet shampoo concentrate; check price against rivals and the listing's first bullet
POSSIBLY GOOD: carpet cleaner solution pet; check whether the listing says it is pet-safe, and whether that is true

Now fix it. The fourth line is where the model flagged a doubt in its own reason, and a negative here would be a mistake if your solution does work in that machine. Check, and either leave it out or keep it. The phrase negative "free carpet cleaner" would also block a legitimate search like "fragrance free carpet cleaner", because a negative phrase blocks any query that contains the complete phrase or a close variation of it. Narrow it to exact "free carpet cleaner sample". And look at the pet term: 22 clicks and no orders for a product with pet appeal is a listing question before it is a keyword question. A pest-control or garage seller reading this will see the same split in their own report: wrong intent goes to negatives; plausible terms go to the listing and the bid.

Match types in plain words

• Negative exact: blocks searches that match the exact phrase or a close variation of it. Up to ten words and 80 characters. Safest. Use it first.
• Negative phrase: blocks searches that contain the complete phrase or a close variation. Up to four words and 80 characters. One entry can block hundreds of searches. Powerful and risky.
• Where they go: in campaign manager, select the campaign and ad group, click Negative targeting, then Add negative keywords. A campaign-level negative reaches every ad group in the campaign (Amazon spells this out for negative product targets), so start at ad group level if different groups target different products.

Two more details from Amazon's help page. You cannot change the match type of an existing negative keyword, so delete it and add a new one. And negatives are not instant: Amazon gives a buffer of 72 hours for negative keywords and 96 hours for negative products, so judge a change after a week, not the next morning. Bulk Operations takes a whole list at once if you have dozens.

A cheaper trick: move good terms out

Not every finding is a negative. The terms with orders are the real prize. Look at the same pivot sorted by orders and pick the search terms that convert at a better-than-average rate. Move them to an exact-match keyword in a manual campaign, with their own bid, then add them as negatives in the auto or broad campaign that found them, so the two campaigns do not fight over the same click. A car repair garage selling a dash-cam might find "dash cam night vision 2 channel" converts at 1 in 8 inside an auto campaign. Promoting it to an exact keyword gives control of the bid; negating it in the auto campaign avoids paying twice. AI can build the list of promotion candidates with the same prompt style: orders of at least 2 and a conversion rate above your average.

A second walked-through case: a furniture maker with a long tail

A furniture maker runs one auto campaign for a folding desk, spend $420 in 60 days (two 30-day reports combined), 1,900 clicks, 31 orders. The flagged list has 74 terms. Sorted by cause: 22 are WRONG PRODUCT (standing desks, gaming desks, desk lamps), 15 are WRONG INTENT (assembly manual, desk plans, free), 9 are WRONG FIT (sizes the desk does not come in), 18 are POSSIBLY GOOD and 10 UNSURE. The 46 clearly wrong terms account for about $118 of spend, 28 per cent of the campaign. If the maker negates those, and the spend moves to better terms at the same conversion rate, about $118 is freed in each 60-day period (it is not saved, since the budget may simply find the next-worst term). The honest expectation is smaller than the headline, which is why you re-check after two weeks.

Where the AI misleads here

• It over-negates. Models like tidy lists. Ask for the doubtful terms separately, as the prompt does.
• It misreads shorthand and brand names. A search term that is a model code may be a perfect match for your product. Give it a list of your product's model and compatible names.
• It cannot see your listing. Terms that fail because of price, main image or a missing size are listing problems. Paste the title and first bullet so it can notice a mismatch.
• It does arithmetic carelessly. Sum the spend yourself.
• It does not know your margins. Add them, or leave the money judgement to yourself.

Product-targeting rows and look-alike ASINs

Your report will also contain rows where the "search term" is a competitor's product code, which appears when product targeting or an auto campaign placed your ad on another listing. Treat these as a separate group. If a code points to a product that is a poor match, the fix is a negative product target, added from the same Negative targeting view with Add negative product targets, where you can exclude individual ASINs or whole brands. If it points to a close competitor and the clicks convert, you have found a placement worth a manual campaign. Ask the AI to keep these rows apart from keyword searches, and tell it which products you consider true rivals.

A weekly before and after, with numbers

Take the carpet solution again. In the 30 days before the clean-up, the auto campaign spent $310 for 26 orders and $3,900 in sales across the campaign group, a cost of $11.92 per order. The first pass added 31 negatives that had absorbed $96 of spend with no orders. In the next 30 days the campaign spent $268 for 25 orders, $10.72 per order. Spend fell $42, orders fell by one, and the cost per order dropped about 10 per cent. The expected saving from removing $96 of waste was larger than what happened, because some of the budget moved to the next-worst terms. These figures are illustrative, but the pattern is the usual one: a real improvement that is smaller than the headline suggests.

A second prompt: finding the terms that deserve more

The same report has the good news, and a second prompt finds it:

From the table below (term | clicks | orders | spend | sales), list every term with at least 2 orders. For each, compute conversion rate (orders divided by clicks) and cost per order (spend divided by orders). Rank by conversion rate. Mark any term whose conversion rate is above my average of [X]% as PROMOTE and any whose cost per order is above [Y] as WATCH. Show your arithmetic for the top five. Do not suggest bids.

Check the arithmetic on the top five before promoting anything. A term with 2 orders from 6 clicks has a rate of 33 per cent, which is excitement from a tiny sample. Require a floor, such as 10 clicks, before treating a rate as real.

Mistakes that cost sales, not just money

One typical slip: a seller asked a model to "clean the report", and it dropped terms with no orders in the window, including a brand-name term that sold well through other channels and was now bringing branded traffic. Negating it cut the visibility of the seller's own name. Another: a phrase negative for "cheap" removed a whole group of buyers who simply searched "cheap" before a premium purchase. Both are reasons to keep exact match as your default and to review a phrase negative against a list of your converting terms before adding it. A third slip: negating terms in a launch campaign two weeks old. New products need clicks to learn; give them time, then review.

The decision rule you apply last

For each proposed negative, ask three questions. Would I ever want an order from this search? Is there a plausible reason it did not convert that I can fix instead? Does the term hide a good sibling, so that a phrase negative would block it? Only a clear no, no and no goes in. Everything else is either left running a little longer with a lower bid, or listed for a listing fix. That is the person's decision, and no model replaces it.

Adding the negatives and recording what you did

Open the campaign and ad group, click Negative targeting, then Add negative keywords, and add the list in batches by match type; Bulk Operations handles longer lists. Save a dated copy of the list in your spreadsheet, with a column for the reason and one for the reviewer's initials. Two weeks later, compare: did the campaign's spend per order fall, did orders hold, and did impressions on the good terms rise? If orders fell more than spend, a negative was wrong; remove the likely culprits one at a time.

Do not treat the first pass as the last. A campaign that has run for six months yields fewer new wasted terms each time, and the work shifts to new campaigns and new products. For a view of whether paid software that automates this is worth the fee at your size, see the break-even tutorial on AI PPC software.

What to do with the terms you decide to keep

Terms in the POSSIBLY GOOD group deserve a short investigation before they are either bid up or cut. Open your own listing as the shopper would after searching that exact phrase, and ask what a stranger would see in the first five seconds. Is the price in line with the other results? Does the main image show the thing they typed? Does the title contain the words they used? A pet-solution search that lands on a listing which never mentions pets will lose to a rival that does, however good your product is.

Give each kept term one of three treatments. Lower the bid by a quarter and watch for two weeks, if the term is plausible but expensive. Leave it and fix the listing, if the mismatch is yours. Or cap the spend by moving it to its own campaign with a small daily budget, so you can see its true cost per order without it eating the main campaign. Record the choice in the sheet, because the point of keeping a log is to know next month what you tried.

A pest-control firm selling a rodent trap might use the same method on a term such as "humane mouse trap": relevant, but if the trap is a snap model the listing mismatch is real and the right fix is a negative, not a better bid. The model can suggest this only if you tell it what the product is, which is why the first line of the prompt describes the product in a sentence.

Fitting it into a weekly rhythm

1. Monday: download the report for the last 30 days, with the most recent three days removed (10 minutes).
2. Pivot by search term; flag by your threshold (10 minutes).
3. Run the grouping prompt (5 minutes) and review the output line by line (15 minutes).
4. Add the approved negatives; promote one or two converting terms (10 minutes).
5. Log the changes and the date (5 minutes).

Under an hour a week for a handful of campaigns. The weekly reports review tutorial places this beside the other numbers you check.

How to check it worked

Compare two equal windows, before and after, for the campaigns you changed. Look at spend per order, total orders and total sales. A good result is lower spend per order with orders flat or up. If total orders fell sharply, you cut something good. Remember outside causes: a price change, a competitor's promotion, a seasonal shift or a stock-out can move the same numbers, so write down anything else that changed. And keep your expectation modest: removing waste improves efficiency, but it does not by itself make a weak product sell.

If your results do not change at all, look at the opposite problem: you may have too few clicks to show waste, and the better use of your time is on the listing. The review analysis tutorial shows how to find the words and images that raise conversion, which cuts wasted spend from the other end.

Questions about the search term report and AI

How far back can I download the report?

Amazon's Ads reporting documentation gives 65 days of search term data for Sponsored Products and 60 for Sponsored Brands, with at most 31 days in one report. Check the date picker in your own console too. Download the report on a schedule and keep the files, so your history does not depend on what Amazon keeps.

Is 20 clicks with no orders always wasteful?

No. It is a starting line, not a verdict. A term may be new, seasonal or a good one with a bad landing listing. Set the number from your own conversion rate and price: if you usually sell one unit per 10 clicks, 20 clicks with none is a signal. If you sell one per 40, it is not yet.

Should I use exact or phrase negatives?

Start with negative exact for single bad terms and negative phrase when a whole family of searches is wrong, such as every query containing a competitor's model name you do not match. Phrase negatives block more, so check that they do not catch good terms. Amazon allows up to four words in a negative phrase and ten in a negative exact.

Can AI add the negatives for me?

Technically a connected tool might, but I would keep adding them yourself or at least approving each batch. A model can misread intent, and a wrong negative quietly removes sales you would have had. A person reviews the list before it goes in.

How often should I do this?

Weekly for campaigns spending real money, monthly for small ones. Newly launched campaigns and auto campaigns collect the most irrelevant terms in the first weeks, so look at those soonest.

Further reads

• AI Max Negative Keywords: How PPC Agencies Control Search Term Spend — See how negatives work in another ad platform.
• AI Tool for Amazon Listing Copy: Bullets, Search Terms and Checks — Fix the listing when good terms do not convert.
• Amazon Seller Assistant Explained: Check Its Advice First — Check what Amazon's own assistant suggests for ads.
• Amazon FBA vs FBM Cost: Model Your Own Products With AI — Know your real margin before setting click thresholds.
• Amazon A+ Content With AI: The Built-In Generator and Checks — How to use the AI Ready modules in A+ Content Manager, and a pre-submission review for claims, image text and facts.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Amazon Ads help page on adding negative keywords or negative products (updated 20 October 2026) and Amazon Ads API reporting documentation for the search term report, both read 5 November 2026; menu names vary, so confirm them in your own console.