# Survey questions that push people toward an answer > Leading stems, fused questions, charged words, absolute wording and lopsided scales change what respondents say. Here is how to spot each one and reword it. Published 2026-10-28 ยท https://aiskills402.com/blog/survey-question-bias-wording A survey question is a measuring instrument, and wording can bend it. The same issue, asked two slightly different ways, can produce quite different answers, so a questionnaire with careless wording reports the writer's opinion back to the writer. Five habits do most of the damage: a stem that hints at the right answer, two questions glued into one, a word that already passes judgement, an absolute word that nobody can honestly agree with, and a rating scale tilted to one side. This guide describes each, shows a neutral rewrite, and ends with a routine for proofing a draft, by hand or with a model, before it goes out. ## Why wording matters at all Pew Research Center, which publishes its survey method openly, reports that tiny changes in phrasing can move the results by a wide margin. Its own example is a question from early 2003 on military action in Iraq. Asked plainly, 68% favoured it and 25% opposed it. After a clause about possible heavy casualties was added, 43% favoured and 48% opposed ([Pew Research Center](https://www.pewresearch.org/writing-survey-questions/)). Nobody lied in either version; the second question simply described a different situation. That is the point to hold on to. Bias in wording is rarely a trick. It is usually a sentence written by someone who already knows what they hope to hear. ## Habit one: the opening that hints at the right reply A leading question carries its preferred answer inside it. Typical forms are a closing tag that invites a yes, a clause that makes an argument before the question arrives, and compliments woven into the sentence. - Biased: "Surely our redesigned checkout is much faster, isn't it?" - Neutral: "How does the new checkout compare with the old one in speed?" with options from much slower to much faster. A related trap is the agree/disagree format. Pew notes that it is among the most common formats and that some respondents lean toward agreeing with a statement whatever it says, which it calls acquiescence bias. Its suggested alternative is to offer a choice between alternative statements ([Pew Research Center](https://www.pewresearch.org/writing-survey-questions/)). The agree/disagree format is not always a flaw, but a statement that argues its own case is. ## Habit two: two questions in one A double-barreled question asks about more than one thing, so a single answer cannot fit. Pew's example asks how much trust respondents place in a president across both home affairs and international affairs, and says it is more effective to ask the two as separate questions. It adds that such questions are hard to answer and the responses hard to interpret ([Pew Research Center](https://www.pewresearch.org/writing-survey-questions/)). - Biased: "How satisfied are you with our prices and delivery speed?" - Neutral: two questions, one for prices and one for delivery speed. Not every "and" is a problem. A fixed name for a single item, such as a document called "terms and conditions", is a single subject. The test is simple: could a sensible person answer one half with yes and the other half with no? If so, split the question. ## Habit three: the word that judges first Some words carry a verdict. Pew advises considering whether certain words may seem biased or offensive to some respondents. It reports that in a 2005 survey 51% favoured a legal option for physicians to help dying patients end their own lives, while only 44% favoured the version that used the word suicide, although both versions asked about the same thing. It also notes that people back extra help for the poor far more readily than extra spending labelled welfare ([Pew Research Center](https://www.pewresearch.org/writing-survey-questions/)). In everyday questionnaires the loaded term is often praise or blame slipped into a noun phrase: "our brilliant support team", "the outdated portal", "unnecessary paperwork". The fix is to name the thing and let the respondent supply the adjective, perhaps through the answer options. ## Habit four: absolute words in the stem Words such as always, never, all, every and none turn a question into a trap. "Do you always read the instructions before starting?" is nearly impossible to answer yes to honestly, so the answers understate how often people read them. Pew's page does not discuss this habit; it is common practice among survey writers, not a finding from the source cited here. The remedy is to ask about frequency with a proper list of answer options, or to ask about a specific recent occasion. Absolute words are fine as answer options such as "every day", where they describe a choice rather than force one. ## Habit five: a scale with more of one side A rating scale should offer as many favourable points as unfavourable ones either side of a neutral middle. A list of "excellent, very good, good, fair, poor" gives three favourable steps and one unfavourable, so even a lukewarm group drifts to the positive end. The matched alternative is something like very good, good, fair, poor, very poor. Pew's page does not cover this point either, so treat it as general practice. What the page does say about options applies here too: closed-ended lists should include every reasonable response, and the categories should not overlap ([Pew Research Center](https://www.pewresearch.org/writing-survey-questions/)). It also explains that ordered categories are kept in order because the order itself carries information. ## A routine for proofing a draft 1. Number every question and write out its answer options beneath it. 2. Read each stem on its own and ask: does it argue, praise, blame, or ask for two things? 3. Look for absolute words in the stem, and decide whether each one is really an answer option. 4. Read each option list: matched sides, nothing missing, nothing overlapping. 5. Rewrite only what failed, in a single neutral sentence. 6. Test the check itself. Plant one deliberately leading question and one clean question in a copy of the draft; a reviewer that flags the clean one, or misses the planted one, is not yet trustworthy. Question order also changes answers, since earlier questions can give context to later ones, as Pew shows with a pair of questions where the share dissatisfied was 88% in one order and 78% in the other ([Pew Research Center](https://www.pewresearch.org/writing-survey-questions/)). A wording check cannot see that, so the order needs its own pass. ## When this does not apply Closed questions in a questionnaire, and how they are phrased, are the only subject here. Left out are who is asked, how the sample is drawn, the mode of the survey, or what the results prove afterwards; those need other checks. Some questions are loaded on purpose, for example in message-testing experiments where two wordings are compared, and there the difference is the point. Where the respondents are experts, a technical term may be neutral even though a layperson would find it charged. The Pew examples come from United States public-opinion polling and may not transfer to every language or culture. The statements about leading stems, absolute words, unbalanced scales and the split test for double-barreled questions rest on common practice, not on the single source cited, and this guide did not check them against a second source or run its own experiment. The routine above plants a clean question for a reason: the post on [proofreading with an LLM](/blog/llm-proofreading-changes-correct-text) shows a reviewer that changed text which was already right. ## Proofing a questionnaire with a skill Anyone who wants an automatic extra reader can use the [survey-question-bias-check](https://aiskills402.com/skills/survey-question-bias-check) skill. Give it your questions in order, each with its choices, and it replies with one line for every problem found, naming the question, the kind of bias, the reason and a calmer rewrite. All five habits above are covered, harmless look-alikes are skipped, and a clean draft gets the short reply "No findings." It reads wording only, never sampling or order.