Negative Keywords List: How to Build One From Your Own Data Instead of Borrowing Someone Else's

2026-09-23

A negative keywords list is a set of terms you exclude so your ads stop appearing against them, saved once and applied across campaigns. That mechanical definition is covered thoroughly across the guides that come up for this search. The part that decides whether the list helps is the judgement underneath: each term you apply is a hypothesis that serving ads against that query is not worth paying for under this offer, these economics, and this campaign's job. Some of those hypotheses are easy to justify. Others are inherited assumptions wearing the costume of best practice. Paste 1,500 generic terms into your account and you have just made 1,500 targeting decisions on the advice of someone who has never seen it.

This article is about which exclusions are actually yours, how to tell whether the data you are building from can support the decision, and how to read the result afterwards.

First, check that traffic composition is the constraint

Excluding search terms is a traffic-stage move. It changes who arrives. It cannot fix an offer nobody understands, a page that does not load on a phone, or a lead record that never reaches your CRM.

When there is no obvious break, work the stages in order and stop at the first one that fails:

1. Viability — does a qualified, reachable, willing market exist at economics this business can support?
2. Measurement readiness — is the data real enough to make decisions on?
3. Offer, audience, and message — does the right person understand what this is and why it is for them?
4. Traffic — is the right person arriving, in enough volume to learn from?
5. Page and funnel — does the path from arrival to purchase hold together?
6. Email and lifecycle — is the value of people who did not buy immediately being captured?
7. Economics and scale — do the unit economics support spending more, and can fulfilment absorb it?

Negatives live at stage 4. The order matters because a failure at stage 2 makes any conclusion you draw at stage 4 unreliable. If the conversion event is not firing, a term flagged as wasteful because it produced no conversions may only have produced no recorded conversions.

One useful test at this stage: when engagement is strong and downstream conversion is weak, investigate who is arriving before rewriting or excluding anything. If click-through looks healthy and opt-ins or purchases do not, traffic composition is worth examining. If engagement and downstream conversion moved together, that particular mismatch is not pointing at traffic composition — which is not the same as clearing your traffic. A query that plainly cannot describe a buyer still deserves excluding on its own merits.

There is also a boundary worth naming. If your ads are already producing relevant enquiries and the stall sits between enquiry and sales conversation, work that unresolved middle rather than the top of the funnel. Tightening the search terms will not fix a stage nobody has built yet.

Make the search terms report decision-ready before you build from it

You are about to remove whole categories of query from your campaigns on the strength of a report. Check the report first. Each of these is a separate check with its own finishing condition — clearing one does not clear the rest.

• Date range, timezone, filters, campaign status, and event configuration. Run this check before you interpret anything. A report that quietly excludes a segment, or spans a period when the offer changed, can nominate terms for exclusion that never deserved it. If you have already run this check on this data set and nothing has changed since, skip it — over-verifying is its own waste.
• The whole count chain. Line up impressions, clicks, page views, sessions, actions, and leads for the same period, and investigate the largest unexplained gap. If most of the loss sits between clicks and page views, the problem you are attributing to bad queries may sit somewhere else entirely. This check is done when the gaps for this period are explained.
• What the conversion column actually counts. Write down the behaviour being counted and the population it is divided by. If "conversion" means a thank-you page view in one place and a booked call in another, the terms you flag as non-converting are non-converting against a definition nobody has stated. This check is done when every metric in play has a written definition with an explicit denominator.
• Whether the cohort has matured. If buying takes weeks, link each sale back to the enquiry that started it and report recent cohorts separately from matured ones. Judge a term on a window shorter than your sales cycle and you are reading a half-finished result as a verdict. Skip this only where purchase is effectively immediate.
• Whether one number is hiding two populations. Blend two campaign objectives, two offer versions, or two destinations into one figure and that figure cannot tell you which one moved. Split by source and version first. Skip this only where there genuinely was one source, one version, and one offer in the period.

The same discipline applies to any inherited number in an account — it is the logic behind making a cost per lead figure decision-ready before you optimise it.

Every exclusion is a hypothesis — write down the reason

Treat each targeting filter as a hypothesis, and either produce the evidence for it or label it as untested. This is foundational work in building a negative list, because it turns an invisible assumption into something you can check later.

Keep three columns for every term you add:

• The term.
• The reason. Spend with no qualified outcome across a countable number of events, a query that cannot describe anything you sell, a geography you cannot serve, a price point your margin cannot support.
• Observed, inferred, or untested. Keep what you observed separate from what you concluded from it, and never let the second quietly become the first. "Twelve clicks, no leads" is an observation. "This query does not pay for itself at our current offer" is an inference. "Anyone searching 'cheap' is a bad customer" is untested.

Match the evidence to the size of the decision. You can delete a negative keyword in seconds; what does not come back in seconds is the traffic it blocked and the history you would need to judge it again. Exclude a high-volume term and you put far more traffic behind that single judgement — so if the call is wrong, the opportunity you forgo and the evidence you never collect are both larger. Decisions of that size earn more evidence before you act. Count the actual events rather than the percentages, and ask what one more conversion would do to the number.

Model the rest of your targeting the same way — on how customers actually behave, with every additional restriction justifying itself. If targeting already reflects observed customer behaviour and each restriction has a stated reason, this work is done and the constraint is elsewhere.

What to do with the giant downloadable lists

Those lists work as a prompt and fail as a default. The rule for anything you inherit: when you encounter a specific number, setting, or tactic, record where it came from and what produced it — then verify it against your own current data, or do not use it.

Sort the terms by how confidently you can justify them

Some exclusions are definitional for your business: you do not sell software, so activation keys and torrents cannot describe what you offer. Those are close to free. Others encode an assumption about intent — "free", "cheap", "review", "how to", "course" — and that assumption may or may not hold for your offer. Put those in the untested column and decide deliberately rather than pasting the block.

Check the sector caveats against your own model, not the category label

The employment, education, and low-ticket blocks in these lists come with "unless you're in X" warnings. The warning is right and too coarse. The question is not which industry you are in; it is whether someone searching that way meets every condition required to buy from you at a cost your margin supports. If nobody has ever produced that list of conditions, that gap sits upstream of the negative keyword list and is worth closing first.

Treat match type behaviour as something to verify, not recall

Negative match types do not behave identically to their positive counterparts. Check the current documentation for the platform you are on rather than an explanation you read once. Any setting or figure of unknown vintage that is shaping your exclusions should be re-verified before you rely on it.

Apply the list so the result is readable

Negatives get edited into live campaigns, which makes them easy to apply and hard to attribute. Two habits keep the result interpretable.

Keep something to return to. Before the change, record the exact list, the date, and the campaigns it was applied to. If performance drops afterwards and there is no recorded baseline, you have lost the reference point as well as the traffic. Where you can duplicate rather than edit in place, do.

Change one thing. Reduce the change to one primary uncertainty and hold the rest on known controls. Add a large exclusion list on the same day you launch new creative and a new landing page, and whatever happens next belongs to all three. If you must do both, write down in advance which effect you will be unable to attribute.

Then split the reporting before and after the change date rather than averaging across it. A period containing two different targeting configurations holds two populations in one number, and that combined figure is for operations only — not as a test result.

Judge it on economics, not on cheaper clicks

Exclude a set of terms and you change what your traffic is made of. That can move click-through, cost per click, and cost per lead in either direction, depending on what the removed queries were costing and how budget reallocates across the queries that remain. None of those movements tells you whether you made more money.

Decide to keep, extend, or roll back the exclusions on contribution and trend against what you can afford to pay for a customer — never on the cost of clicks or leads alone. If cost per lead fell while qualified pipeline fell with it, that pattern is consistent with the list removing buyers alongside waste, and it is a reason to investigate rather than a finding.

When several stages all look improvable, estimate what a plausible improvement at each would do to qualified outcomes, and work the one with the largest effect rather than the one that irritates you most. Watching irrelevant queries burn budget is genuinely irritating. That is not, by itself, evidence that it is where the money is.

Reading the result when nothing obvious happened

There is a clean version of this result: if the excluded queries would never have produced a qualified outcome, and the budget they were consuming does not simply move elsewhere, qualified outcomes hold while spend drops. Real accounts rarely hand you that version, so read the patterns below as directions to look, not as answers.

• Spend fell and qualified outcomes fell in step. Consistent with excluding buyers along with waste — and also with an immature cohort, a thin sample, seasonality, or another change made in the same window. Check those before rolling anything back.
• Spend fell and nothing downstream moved. The removed queries may not have been where the loss was. They may also have been too small a share of spend, or the window too short, for any downstream effect to show.
• Spend did not fall at all. Budget can reallocate to other eligible queries. Look at what it moved to before deciding the exclusions did nothing.

Where to go next depends on what you actually verified earlier. If you reached the traffic stage without confirming measurement readiness or that the right person understands the offer on arrival, go back and verify those before interpreting anything about the exclusions — a stage-2 failure makes a stage-4 reading unreliable. If those stages were verified and held, the next move is either another traffic hypothesis or the next stage in the sequence, page and funnel. It is not automatically a return to the beginning.

If the arrivals still do not qualify with the list in place, the question becomes which stage produced them — a different investigation from search term exclusion, and the one covered in working out which stage actually produced your unqualified leads.

None of these checks prove anything by elimination. Insufficient volume, timing, execution, and interactions between concurrent changes can survive every one of them. If you work through the list and still cannot isolate a cause, the honest reading is that the evidence in front of you is not sufficient to diagnose it, and the next step is more first-party investigation rather than another round of exclusions.

If you have tightened targeting repeatedly and revenue has not followed, that pattern is itself diagnostic — GrowthHackr's free Campaign Leak Audit walks the same sequence against your own numbers to find the first stage that fails or that your current evidence cannot support.