Search terms · 16 min read · Interactive · Video

Which search terms should I block, and which deserve a chance? The Wilson score interval

It is review day. The search terms report has four hundred rows, and a dozen of them are painted red: each one spent real money and converted nothing. My cursor is already on "Add as negative keyword". For most of those rows, that click is a coin toss.

Why the obvious click is a coin toss

Every day your campaign runs, the search engine matches your keywords to searches you never bid on. Some are gold, a phrase describing exactly what you sell, found for free. Some are junk. This stream is where both the opportunity and the waste live.

The rule most account managers run in their heads is simple: the moment a new term has spent money without converting, add it as a negative. It feels tidy. It is also how good terms quietly vanish from accounts, because it trusts a conversion count far too small to say anything.

The idea, in plain words

A term's conversion rate is a share: conversions out of clicks. With a handful of clicks, that share is barely known. You would expect no conversions from three clicks even on an excellent term.

So stop reading the rate the report shows and start reading the range of rates still plausible. A term with no conversions in 30 clicks has a range from 0% up to about 11%. Your benchmark sits inside it. The data cannot yet tell this term from an average one, so there is nothing to decide.

Give the same term ninety-three clicks, still without a sale, and the range shrinks to just under 4%. Now the whole range sits below your benchmark, and blocking it is a judgement the data supports rather than a hunch. That is the entire trick: the range narrows as clicks pile up, and you act when it clears the line you set, not when your patience runs out.

And that line is yours to choose. Ninety-three clicks is where the term becomes worse than average at 95% confidence, which is not the same as useless. Its true rate could still be three per cent, and most people would keep a three per cent term running because it still pays. So set the benchmark at the rate below which a term stops paying for itself, usually lower than the account average, and the same arithmetic hands you the click count for that line.

The formula, written out

Here are the two pieces of maths in this post. The first gives the range. The second turns it into the number you want: how many clicks a term needs before it can be judged.

The range. For x conversions in n clicks, the Wilson score interval is

None of this is mine. Edwin Wilson published this answer in 1927, about survey answers and coin tosses, decades before anyone had a search terms report [1]. A click that either converts or does not is the same shape of thing.

The clicks a fair chance costs. Set x = 0 and the formula collapses. With no conversions at all, the top of the range is simply z²⁄(n + z²). Ask when that drops below your benchmark p and you get

At 95% confidence and a 4% benchmark that is 93 clicks. At a 1% benchmark it is 381. The lower your ordinary conversion rate, the longer every term has to wait, which runs against the instinct that low-converting accounts can least afford to.

How sure is the 93 itself? Once you fix p and z it is exact arithmetic, so its only uncertainty is borrowed from your benchmark. This shop’s 4% rests on 400 sales in 10,000 clicks, which puts the benchmark between 3.6% and 4.4%, and so the click count runs from 84 to 102. You are as sure of the click count as you are of your benchmark, and no surer.

There is a back-of-the-envelope version, the rule of three: with no conversions in n clicks, the top of the range is about 3⁄n, which puts the 4% crossing at 75 clicks. It is the dotted line in the chart above.

Why not the interval from school? The one most people were taught is the observed rate plus or minus two standard errors. At zero conversions it gives a range of zero width, a claim that the term's true rate is exactly 0% with certainty. That is nonsense, and it is the very case you most want to judge. Agresti and Coull showed that the textbook version, which they call the Wald interval, performs badly unless the sample is large [3]. Brown, Cai and DasGupta compared the alternatives and concluded: "we recommend the Wilson interval or the equal-tailed Jeffreys prior interval for small n" [2].

A worked example synthetic data

Picture a running-shoe shop buying search ads only (synthetic, built so the right answer is known in advance). The same report exists on Google, Microsoft and any other search engine. The notebook that comes with this post builds the shop under one fixed seed, so every number here regenerates exactly.

In one month, 600 new search terms appear, take 6,631 euros of clicks, and produce 217 conversions. Behind the scenes, visible to me and not to the advertiser, 256 of those terms are simply irrelevant and 83 are on topic but not shopping. Together they take 57% of the month's spend.

Where the 4% benchmark comes from. It is an input the shop brings: the rate its established keywords convert at, 400 sales in 10,000 clicks. It is not the rate of the 600 new terms. Those convert at 2.7% between them, dragged down by the junk among them, and judging a term against a rate its own junk lowered would be circular. Take your benchmark from the keywords you already trust.

The four kinds of term. Each one carries a hidden true conversion rate the shop cannot see.

That bottom row keeps the rest honest. Irrelevant is not never. The band runs up to 0.4%, four sales in a thousand clicks, and this month those terms produced six. Step one blocks all 256 of them on day one, and those six sales go with them.

Are these numbers plausible? Site-wide fashion ecommerce converted at 1.81% in July 2026 [6], and apparel paid-search clicks at 4.50% in WordStream’s 2026 sample [7], so the shop sits inside published ranges. The 6 to 10% band for the best terms is not a market claim, only the usual pattern that high-intent terms beat an account's average.

The clicks spread out the way real search terms do: the median term gets 6 clicks in the month, 90% get 30 or fewer, and a handful get hundreds.

That chart is the whole problem. Almost every term a marketer blocks in a monthly review is blocked on data that cannot support the decision.

So the two steps split the work. Step one costs nothing and does most of it. Checking each term against a short list of what the shop sells blocks all 256 irrelevant terms at once, with no statistics involved. Step two protects what is left. Terms with enough clicks are judged against the 4% benchmark, and the rest keep running under an honest "not yet known".

Here is the month run three ways, over 28 days at €0.83 a click (spend over clicks), which works out at €30.56 per conversion if you block nothing.

Follow the same terms for a quarter and the trade comes into focus. Over 90 days, block-on-sight leaves €8,016 of waste running against the two-step rule’s €2,974. The cost in good terms is nearly the same: 98 lost conversions against 107. Same damage, under two-fifths of the waste left behind.

Is any of it worth doing? That depends on what a conversion is worth to you, which the notebook cannot know. Counting all the spend each rule saves, block-on-sight breaks even at €62 a conversion over the month, and the two-step rule, run with the shop’s own word list, at €94. If a conversion is worth more than that, let the terms run.

But look at what you are paying for. Every good term the two-step rule killed was killed by step one’s word list. The interval killed none of them (27 by the list, against 27 in total). Add the missing words and the same rule kills 0 good terms, leaves €872 of waste and breaks even at €483 a conversion. The method was never the expensive part. An incomplete list was, and a list costs nothing to fix.

Even that version is not free: it still costs 6 sales over the month, from irrelevant terms blocked on day one, because “almost never converts” is not “never converts”.

The same trade in shop money. A block keeps saving its term's spend and keeps losing its term's sales, at the same pace, for as long as it stands, so weigh one quarter against one quarter. Say a pair of shoes sells for €50. Whether a lost sale costs you more than the spend saved depends on your margin, so the last column is the number to hold against what a conversion is worth to you.

Every quarter the blocks stand, the two steps save more than block-on-sight and lose only a little more. The complete list is what changes the trade: it keeps nearly all the saving and loses less than a third of the sales. The gap between the last two rows is three words in a text file: "sneakers", "trainers" and "jogging". The busiest casualty of their absence was carbon plate trainers for running, with 247 clicks, 18 sales and €206 of spend in the month, blocked on day one and never measured.

How sure should you be?

The confidence level is a dial, not a law. Turn it down and terms are judged sooner, on less evidence.

On this campaign the dial buys very little. Dropping from 95% to 80% cuts the clicks a term needs by 53 and saves €61 of waste across the month, for 2 more good terms blocked and 1 more conversion lost. The trade is small because step one does most of the work. With a weaker relevance list, the dial matters much more. Reddit, for what it is worth, runs this interval at 80% to sort its comments [4].

Who actually uses this

Two public cases, and only two, because I keep to what can be sourced.

Reddit sorts comments with it. Its open-source code holds a function its authors named the confidence sort, which computes exactly the lower end of the interval above. It runs at 80% confidence, a constant sitting in the code as z = 1.281551565545 [4]. A comment with one upvote and nothing else should not outrank a comment with forty, and the raw rate cannot say so. The lower end of the range can.

Stack Overflow looked at the same method and turned it down. Jeff Atwood wrote it up in 2009, granted it was a clear improvement for Reddit, and concluded: "I am not sure it's as clearly useful for Stack Overflow" [5]. That is the honest shape of a method's adoption: fitted to one problem, declined for a neighbouring one, by people who understood it.

Run it on your own data

Every download runs entirely on your own computer, and your data never leaves it. They take one search terms report, exported as CSV, plus a short list of what you sell.

Column names are matched loosely, so "Query" is read as the search term and "Conv." as conversions. If your report repeats a term across campaigns, ad groups or days, the notebook adds it up. Google's export opens with two title lines and ends with "Total" rows. The notebook skips both, so export it and leave it alone. It also refuses to give a confident answer it cannot support. With too few terms it says the benchmark is a guess, with zero conversions in the file it says step two cannot run at all, and with a list so broad that every term matches it tells you step one is doing nothing.

Getting your data out of the platform

Pull the whole life of each term. Clicks pile up, and a monthly review reads a running total rather than starting again each month. Export as far back as the terms have existed, not one billing period. At this shop’s pace the median term needs about 14 months to reach 93 clicks, which is why “not enough data yet” is the usual state.

In Google Ads, pick the campaign, then Insights and reports, then Search terms, with Impressions, Clicks, Cost and Conversions among the columns, over at least the last 90 days. In Microsoft Advertising, go to Reports, then the search terms report. The notebook also carries the query for the Google Ads API.

Two things people get wrong. Export the conversion column your campaign is actually optimised for, not "all conversions", because view-through and secondary actions make dead terms look alive. And keep it to one campaign, or the benchmark blends campaigns that convert at different rates.

You also need a list of what you sell, in the words a customer would type, not your internal product names. Two or three phrases of several words are enough to start. Single common words match everything and switch step one off entirely, which the notebook warns you about.

This part touches a spreadsheet or a light script, but you do not have to do it alone. Download the give-it-to-your-AI brief and hand it to ChatGPT, Claude, or any capable assistant. It carries everything the assistant needs to pull your report, shape it, and run the notebook, one step at a time, even if you have never written a line of code.

Common questions

From Stochastic Strata. I build marketing measurement that shows its working: every number carries its range, and when the data cannot answer a question, I say so and tell you what would. The worked example here is synthetic. The method is the real thing, and you can run it on your own data with the notebook or spreadsheet above. Nothing in this post is financial advice.