Back
Matrix Questions in Surveys: When to Use Them (and When Not To)
SmartInterview Team
A matrix question asks several items in one grid, with the rows as the things you are rating and the columns as a single shared response scale. Use one when every row genuinely belongs on the same scale and you have a short, tight battery. Avoid one when the list is long, the survey is mobile-first, or the items do not share a scale.
The quick version:
Use a matrix for a compact battery of attitude, agreement or brand-attribute items that all share one scale.
Keep it small. As a working rule, stay under about 5 to 7 rows and 5 to 7 columns.
Expect straightlining. The longer the grid, the more respondents pick the same column down the whole set.
Mobile is the constraint. Most respondents answer on a phone, where a wide grid means horizontal scrolling and tiny tap targets.
Split, do not shrink the font. If a matrix does not fit, break it into two shorter batteries or switch to one item per screen.
Matrix questions are a layout choice, not a question type. The underlying question is usually a Likert scale repeated several times. That distinction matters, because it means you can always unpack a matrix back into individual questions without losing any data.
Matrix vs the alternatives
Four ways to ask the same battery of rated items. The right choice depends on how long your list is and how many respondents are on a phone.
Format | Mobile friendliness | Straightlining risk | Completion speed | Best use |
|---|---|---|---|---|
Matrix / grid | Poor to fair. Needs horizontal scroll or a stacked fallback above about 5 columns. | High. All options are visible at once, so repeating a column costs nothing. | Fastest per item. | Short batteries on one shared scale, desktop-heavy or panel audiences. |
One item per screen | Excellent. Full-width tap targets, no horizontal scroll. | Low. Each item is a separate decision. | Slowest. One tap plus one page load per item. | High-stakes items, mobile-first surveys, short lists you need clean data on. |
Card sort / accordion | Good. Vertical stacking, one item expanded at a time. | Low to moderate. Sequential exposure breaks the visual pattern. | Moderate. | Medium-length attribute lists where you still want a compact page. |
Slider battery | Fair. Drag targets are awkward on small screens and need a visible default handling rule. | Moderate. Easier to differentiate, but untouched defaults are a real risk. | Moderate. | Continuous ratings where fine gradation matters more than comparability. |
Fair. Drag-and-drop is fiddly; numbered entry is safer. | None by design. Forced differentiation. | Slow. High cognitive load. | When you need priority order, not absolute levels. Use a rank order question instead. |
Note the trade-off in the middle two columns. The formats that are fastest for the respondent are also the ones that make low-effort answering easiest. Speed and data quality pull in opposite directions here, and a matrix sits firmly on the speed end.
What a matrix question actually is
A matrix question, also called a grid question or a battery, presents multiple sub-questions in a table. Each row is an item you want rated. Each column is a point on a response scale that all rows share.
A typical example: the row labels are "Delivery speed", "Product quality", "Value for money" and "Customer support", and the columns are "Very dissatisfied" through "Very satisfied". The respondent works down the rows, picking one column per row.
The anatomy
Rows are items. Each row is an independent question that will produce its own variable in your dataset. A five-row matrix is five questions, not one.
Columns are the scale. Usually ordered points such as agreement, satisfaction, frequency or importance.
One scale for all rows. This is the hard requirement. If one row would need a different scale, it does not belong in that matrix.
One selection per row in a standard matrix. Multi-select grids exist but are much harder to read and to analyze.
The shared-scale rule is where most bad matrices come from. If you find yourself writing a column header like "Not applicable to me" purely to make one awkward row fit, that row should be its own single-select question instead.
When matrix questions genuinely work
Matrices are not a bad question type. They are a good question type used far outside their range. They work well when:
You have a true battery. A set of attitude statements measured on one agreement scale, designed to be analyzed together as a construct or index.
You are rating one brand on several attributes. Brand image and attribute batteries are the classic legitimate use, because the comparison across rows is the point.
You want comparability. Seeing all items on one screen makes relative judgments easier. Respondents calibrate their answers against each other, which is exactly what you want when your analysis is about which attribute scores highest.
You want to cut switching cost. Reading a new scale for every question is work. Reusing one scale across a short battery lowers that load and shortens the questionnaire.
Your audience is on desktop. Employee surveys, B2B panels and moderated research often skew to larger screens, where the layout constraint mostly disappears.
The documented downsides
Straightlining and non-differentiation
Straightlining is when a respondent picks the same column for every row, producing a vertical line down the grid. It is the signature failure mode of matrix questions, and it gets worse as row counts grow.
The reason is structural. In a grid, the cost of thinking about each row individually is high, while the cost of repeating the previous answer is a single tap in a visually obvious place. The layout itself rewards not differentiating. A milder version, near-straightlining, is harder to spot: the respondent varies between two adjacent columns without ever really engaging.
Satisficing
Satisficing is answering well enough to get through rather than accurately. Matrices invite it because the respondent can see how much work is left and can see a shortcut. The effect compounds through a survey. A matrix at question 3 will produce better data than the identical matrix at question 25.
Fatigue and drop-off
A large grid looks like work before any work has been done. That visual weight is a common abandonment point, and it is one of the mechanisms behind the broader decline in survey response rates. If you track drop-off by question, oversized matrices tend to show up as a visible cliff.
Mobile usability
On a phone, a five-column grid either shrinks the columns until the tap targets are smaller than a fingertip or forces horizontal scrolling. Both are bad. Horizontal scrolling is worse, because respondents can select the wrong column without realizing the header they see is not the one they scrolled from.
Order and position bias
Two biases show up specifically in grids:
Item-order effects. The first row anchors the rest. Respondents rate later rows relative to how they rated the first one. Randomizing row order across respondents spreads this out rather than eliminating it.
Left-side selection bias. In left-to-right languages, respondents skew toward the left-hand columns. If your scale runs from negative on the left to positive on the right, that is not the same bias as if it ran the other way. Keep scale direction consistent across your whole questionnaire, and note it when comparing to a study that used the opposite direction.
Practical row and column limits
These are design recommendations, not thresholds from a specific study. Treat them as defaults to deviate from deliberately.
Columns: about 5 to 7. Beyond that, labels get truncated and the grid stops fitting any screen. A 5-point scale is the safest default for a matrix.
Rows: about 5 to 7 per grid. Some methodologists accept up to 10 for a desktop audience with a well-motivated sample. Very few defend 15.
Total cells: keep it under roughly 35. Rows times columns is a decent proxy for perceived effort.
Row labels: short. If a row label wraps to three lines on a phone, rewrite it. Long statements belong on their own screen.
One matrix per screen, and ideally not two matrices back to back. Consecutive grids are where straightlining accelerates.
Mobile and accessibility
Matrix questions are one of the hardest question types to make both mobile-friendly and accessible. If you only fix one thing about your survey design, fix this.
Mobile-first alternatives
One item per screen. The cleanest data you can get from a battery. Each item gets full-width buttons and undivided attention. The cost is length, so reserve it for the items that matter.
Stacked / responsive collapse. The grid renders as a table on desktop and as a vertical list of items with the scale under each on mobile. This is the minimum bar for any modern survey tool, and it is worth testing on an actual phone rather than a browser resize.
Accordion or card patterns. Show one item expanded, collapse it once answered, reveal the next. Keeps the page compact while forcing sequential attention.
Carousel matrix. One row per swipe with a progress indicator. Feels fast, but hides how many items remain, which some respondents find worse, not better.
Slider batteries. Good for fine gradation, risky for defaults. Always require an explicit interaction before a slider counts as answered, or you will record the starting position as data.
Accessibility requirements
A matrix is a data table being used as a form. That combination is exactly where screen reader support gets fragile.
Real table semantics. The grid must use proper table markup with header cells, not a stack of styled divs. A screen reader user needs to hear "Delivery speed, Satisfied, radio button 4 of 5" rather than an unlabeled radio button.
Header association. Each input needs to be programmatically tied to both its row header and its column header. Without that association, the grid is unusable non-visually even though it looks fine.
Keyboard navigation. Arrow keys should move within a row's options, tab should move between rows. Test that you can complete the whole grid without a mouse.
Do not rely on color alone. Red-to-green column shading is a common pattern that conveys nothing to a colorblind respondent unless the labels carry the same meaning.
Contrast and target size. Shrinking a grid to fit almost always pushes tap targets and label contrast below usable levels. That is a signal to split the question, not to shrink further.
If you cannot get a grid to pass these checks in your tool, the safe fallback is always one item per screen. It has no accessibility problems that a normal radio-button question does not also have.
How to fix an oversized matrix
Say you have a 14-row, 7-column attribute grid that a stakeholder insists on. Here is the order to work through.
Cut rows against a decision. For each row, ask what you would do differently depending on the answer. Rows with no answer to that question are the first to go. This usually removes more than you expect.
Group the survivors into themes. Two grids of six rows, placed apart in the questionnaire with other question types between them, produce better data than one grid of twelve.
Shorten the scale. Move from 7 points to 5. You rarely lose analytical power, and you gain a grid that fits a phone.
Randomize row order across respondents, and keep the randomization seed in your data so you can check for order effects later.
Pull the critical items out. Whatever two or three items your headline metric depends on should be their own full-screen questions, not rows in a grid.
Replace depth with a follow-up. Rating 14 attributes on a 7-point scale is a lot of precision on questions you did not ask the respondent to explain.
That last point is the one most teams skip. A long attribute battery is often an attempt to find out why people feel a certain way by measuring what they feel in more and more detail. It rarely works. Asking six well-chosen rated items and then probing the low scores in the respondent's own words gets you closer, faster. This is the case for mixing methods rather than adding rows, and it is covered in more depth in qualitative vs quantitative research. Platforms like SmartInterview take this route directly: a short rated battery, then an AI voice follow-up that asks about the items scored lowest, with the open responses coded automatically so you still get countable output.
Detecting straightlining after fielding
You cannot fix a grid retroactively, but you can find out how much of your data to trust. Run these checks before analysis.
Response variance across the row set
For each respondent, compute the standard deviation of their answers across the rows of a given matrix. A standard deviation of zero means every row got the same column, which is a literal straight line. Near-zero values catch the respondents who alternated between two adjacent points.
Interpret with care. Zero variance is not automatically cheating. If someone is genuinely satisfied with everything, a flat row is the honest answer. Treat it as a flag to combine with other signals, not a delete rule on its own. Two things make it more convincing: reverse-coded items in the battery that a real respondent should answer in the opposite direction, and the same flat pattern appearing in more than one matrix.
Speeder checks
Record time-on-page per question and compare each respondent to the median for that question. Someone who completed a 10-row grid in a few seconds did not read the row labels. Combining a speeder flag with a zero-variance flag is a far stronger disqualification signal than either alone.
Other signals worth logging
Diagonal or zigzag patterns, which indicate deliberate pattern-filling rather than inattention.
Consistency between related items that should correlate. Contradictory pairs suggest the grid was not read.
Open-response quality elsewhere in the survey. Empty or nonsense verbatims alongside a flat grid is a clear pattern.
Straightlining rate by question, tracked over time. If one grid straightlines far more than the others, that grid is the problem, not the sample. Redesign it before the next wave.
If your straightlining rate is high across the board, the format is doing the damage. That is usually the point to test a different mode entirely, such as the voice survey approach, where there is no grid to straightline through.
Frequently Asked Questions
What is a matrix question in a survey?
A matrix question presents several related items in a grid. The rows are the items being rated and the columns are a single response scale shared by every row. Respondents pick one column per row. It is a layout for asking multiple scale questions at once, not a distinct question type.
How many rows should a matrix question have?
As a working guideline, keep it to about 5 to 7 rows. Up to 10 can be acceptable for a desktop or highly engaged audience. Past that, straightlining and drop-off climb sharply, and the grid stops fitting a phone screen. If your list is longer, split it into two thematic grids placed apart in the questionnaire.
Are matrix questions bad for mobile?
By default, yes. A wide grid on a small screen forces either horizontal scrolling or tap targets too small to hit reliably. Good survey tools collapse the grid into a vertical list of items on mobile, with the scale shown under each item. Test on a real phone, not a resized browser window.
What is straightlining and how do I detect it?
Straightlining is selecting the same column for every row of a matrix. Detect it by computing the variance of each respondent's answers across the rows of a grid. Zero or near-zero variance is your flag. Combine it with time-on-page to separate genuine uniform opinions from inattention, and use reverse-coded items to catch respondents who are not reading.
What can I use instead of a matrix question?
One item per screen produces the cleanest data. Accordion and card patterns keep pages compact while forcing sequential attention. Slider batteries suit continuous ratings but need an explicit-interaction rule so untouched defaults are not recorded as answers. Rank order works when you need priority rather than absolute levels.
Can a matrix question use different scales per row?
No. Every row must share the same column scale. If one item needs a different scale, or needs a "not applicable" option the others do not, it belongs outside the grid as its own question. Mixed scales inside one matrix confuse respondents and make the resulting variables non-comparable.
