> For the complete documentation index, see [llms.txt](https://boundaryai.gitbook.io/boundaryai-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://boundaryai.gitbook.io/boundaryai-docs/analysing-your-feedback/analysing-results/segmenting.md). # Segmenting your feedback Segmentation is the BAI Analytics tool for breaking an analysis apart by who answered. Pick one or more **structured questions** as segmentation axes (country, customer tier, NPS bucket, ticket priority, source platform) and the rest of the page splits into per-segment views you can navigate between, so you can see how the picture differs from one cohort to the next. It's a *breakdown* surface, not a filter. Selecting *Country* doesn't narrow the page to "just Germany"; it produces one segment per country and lets you browse them (*Germany*, *France*, *US*, etc.) each with its own metrics, themes, and Custom Monitoring breakdowns. For the surface that hosts segmentation, see [Analysing results](/boundaryai-docs/analysing-your-feedback/analysing-results.md). For where the underlying axes come from, see [Creating a Survey](/boundaryai-docs/bringing-in-your-feedback/surveys/editor.md), [Uploading an Existing Dataset](/boundaryai-docs/bringing-in-your-feedback/uploading-an-existing-dataset.md), [Connecting your tools](/boundaryai-docs/bringing-in-your-feedback/connect-to-your-existing-systems.md), and [Social Listening](/boundaryai-docs/bringing-in-your-feedback/social-listening.md). *** ### How segmentation works The unit of segmentation in BAI Analytics is a **question**, not a single answer value. When you add *Country* as a segmentation axis, BAI Analytics computes a segment for every distinct value the question contains (*Country = Germany*, *Country = France*, *Country = US*, etc.) and the page exposes a navigator (tabs or a dropdown, depending on the layout) for switching between them. Pick **two questions** and segments are computed for every **combination** of values: *(Germany × Pro)*, *(Germany × Free)*, *(France × Pro)*, *(France × Free)*, and so on. Add a third question and the breakdown deepens further. Up to **four** segmentation questions can be active at once. More than that and the resulting segments become too thin to draw conclusions from in most datasets. *** ### What can be a segmentation axis The segmentation picker offers any question that produces a bounded set of values: * **Single Choice**: every answer option becomes a segment (one segment per country, one per customer tier, one per role). * **Multiple Choice**: each option is its own segment; respondents who selected several options appear in *each* of the corresponding segments. * **Metadata**: auto-collected fields like UTM source, device type, browser, country / region, plus any explicit metadata questions you've added to the survey, and the Context Fields of an uploaded dataset. These three types apply uniformly across surveys, uploads, connector data, and scraper output, so a connector's *priority* metadata field, a scraper's *platform* field, and a survey's *role* question all behave identically as segmentation axes. A dataset needs both an axis and something to break down. A dataset made only of open-ended questions plus metadata columns (a batch of interview transcripts with an age and a tenure column, say) is segmentable: the metadata fields are the axes and the open-ended analysis is what gets split. A dataset with metadata columns and nothing else has nothing to break down, so the button stays hidden. #### Question types that can't be segmentation axes * **Short Answer / Long Answer**: open-ended text isn't bounded; there are no discrete buckets to split into. * **Linear Scale** and **NPS**: not exposed in the main picker, because their numeric ranges would generate too many segments. They appear instead as **targets** (the metric you're looking at) and as **per-theme breakdown axes** in the qualitative view (see Per-theme segmentation below). * **Information / Acknowledgment**: these don't capture answer values. #### The 35-option cap A question is only offered as a segmentation axis if it has **35 or fewer distinct values** in the dataset. The cap keeps segment counts manageable (35 × 35 = already 1,225 cells with two axes) and keeps the picker usable. If a metadata field has hundreds of unique values (a free-text "company name" column, for example) it's silently excluded from the picker; collapse it into a coarser dimension upstream if you want to slice on it. *** ### Adding and removing segments Click **Add Segmentation** on the results page to open the segmentation picker. You see the list of eligible questions with their option count and type badge. Tick the questions you want to slice by and confirm. Choice question cards offer a shortcut: **Add to Filters** in the card footer adds that question as an axis in one click, and turns into **Active Filter** while it is on. Click it again to remove the axis. Selected questions appear as **chips** at the top of the page, each labelled with the question name. Click the X on a chip to remove it; the page recomputes immediately. Chart bars, theme cards, and metadata cells are not click-to-segment shortcuts: interactions on those elements scope to their own surface and don't change the active segmentation. *** ### Where segmentation propagates Adding or removing a segmentation question recomputes the page in real time. The surfaces affected are: | Surface | Updates with segmentation | | ------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------- | | **Overview tab** | No: the Overview is intentionally global. It always shows the dataset's full picture so you have a baseline to compare segments against. | | **Single Choice / Multiple Choice charts** | Yes: distributions split per segment. | | **Linear Scale charts** | Yes: distribution, mean, and median recompute per segment. | | **NPS module** | Yes: score, promoter / passive / detractor split, and trend recompute per segment. | | **Thematics tab** | Yes: theme counts and representative comments reflect the active segment. | | **Custom Monitoring tab** | Yes: matched-comment counts recompute per segment. | | **Raw Data tab** | Yes: the comment list narrows to the active segment. | The navigation pattern depends on how many axes are active: * **One segmentation question**: segments appear as **tabs** at the top of each affected component (*Overall* / *Germany* / *France* / *US*). * **Two or more segmentation questions**: segments appear as **dropdowns** so the cross-product fits in a small space. In both layouts, *Overall* is always the leading entry, so you can return to the un-segmented view with one click without removing any tags. *** ### How respondents are placed into segments A respondent appears in any segment whose criteria they satisfy. * For **Single Choice** axes, each respondent lands in exactly one bucket. * For **Multiple Choice** axes, a respondent who picked *Email*, *Phone*, and *Chat* lands in all three corresponding segments. Their feedback contributes to each. * For combined axes (say *Country* and *Tier*) a respondent must have answered *both* questions to appear in any cross-product segment that involves both. If a respondent answered only one of the two, they're still included in segments that involve only that one question; they're simply absent from cross-product cells they don't have data for. This last point matters for sparse metadata: optional UTM fields, browser / OS detection that didn't fire, or a metadata question added partway through a survey's lifetime. BAI Analytics keeps partial respondents in the segments they qualify for instead of dropping them entirely. Percentages inside a segment are always shares of **that segment's respondents**, so a Single Choice column sums to 100% and a Multiple Choice column can legitimately sum past it. *** ### Comparing segments side by side Flipping between segment tabs works for a quick look; when you need the numbers next to each other, use **Compare segments**, in the 3-dots menu of any Single Choice or Multiple Choice card. The entry is greyed out with a reason until segmentation is on and yields at least two segments. The popup lets you tick **2 to 6 segments** and pick a **Baseline**. With two segments you get columns A, B and Δ; with more, every column is compared against the baseline. On Single Choice questions *Overall* can be one of the columns; on Multiple Choice it isn't offered, because the question-level total counts selections rather than respondents and wouldn't be comparable. For each option and segment the table shows the **count** and the **share** of that segment's respondents, and the Δ columns show the difference in **percentage points** against the baseline. Rows are the union of options across the selected segments (an option absent from a segment shows 0), sorted by largest absolute difference by default; click a column header to sort by it instead. Every column header carries its **n**, with a **Low sample** marker under 30 respondents, because small differences there may be noise. **Export CSV** reproduces exactly what's on screen: labels, n's, low-sample markers and the current sort. *** ### Statistical significance Once segmentation is active, a choice question's card footer gains a link to its **details page**, which adds inference to the breakdown: * A **Statistical significance** summary table listing, for every segmentation question, the test used, the adjusted p-value and the verdict (**Significant** or **n.s.**). Choice questions use a chi-square test of independence; p-values are adjusted for multiple comparisons (Benjamini-Hochberg, α = 0.05). * A **Cross-tab** per segmentation question, with counts, within-segment percentages and **95% confidence intervals** in parentheses. * Where a reliable test isn't possible, the badge says why rather than overstating: **Insufficient data** (too few responses per cell), **Too many options**, or **No variation**. Numeric questions get the same treatment. On Linear Scale and NPS cards, and on the NPS **More Details** page, a panel shows each segment group's **mean with its 95% confidence interval and n**, plus a significance badge per segmentation question (Welch's t-test for two groups, one-way ANOVA for more, same adjustment). See also [Benchmarking your NPS](/boundaryai-docs/analysing-your-feedback/analysing-results/benchmarking-your-nps-against-industry-standards.md). The tested counts are always the same counts you see on screen. *** ### Per-theme segmentation In addition to the page-wide picker, the Thematics tab offers a **per-theme segmentation panel**. Inside any theme's drill-down view, expand the segmentation panel and pick a single axis to break that theme's mentions down by: country, customer tier, NPS bucket, scale rating. Two things make this surface different from the page-wide picker: * **It's per-theme, not page-wide.** The selection only affects the theme you're inside; the rest of the page stays at the active page-wide segmentation. * **It accepts Linear Scale and NPS** as axes, because the panel buckets them automatically: * **NPS** is split into **Detractors (0 to 6)**, **Passives (7 to 8)**, **Promoters (9 to 10)**. * **Linear Scale** is split into **Low**, **Medium**, **High** thirds of the configured range. Use the per-theme panel when you want to know *whether the people complaining about price are the same people giving low NPS scores*, or *whether the "checkout flow" theme is concentrated in a specific country*. The page-wide picker can't answer that elegantly because it would split every chart simultaneously. *** ### Sharing a segmented view Your active segments are saved in the page link. Share that link with a colleague and they will open exactly the same segmented view. If a segment refers to a question that no longer exists (for example, a renamed survey or a deleted question) it is simply skipped, and the page still opens normally. *** ### Speed Segmented views are saved as you open them, so switching between segments you have already viewed is instant. The first time you open a new combination, very large datasets may take a couple of seconds to calculate. Views refresh automatically when you re-run the analysis or the underlying data changes. *** ### Running in-depth analysis per segment For the deeper in-depth view on open-ended questions, each segment's analysis is run on demand rather than all at once: select a segment and click **Run analysis** (or **Retry analysis** if a prior attempt failed). Use the navigator's **Analyze all segments** to kick off every segment in one action instead of clicking through them one by one; the navigator shows progress and how many combinations were queued, skipped (no respondents) or failed. Re-running a segment that was already analysed asks for confirmation, because it overwrites the saved result and uses AI credits. *** ### Weighting is the exception Segmentation and [sample weighting](/boundaryai-docs/analysing-your-feedback/analysing-results.md#weighting-your-results-redressement) can't be active at the same time on the same view; turning one on pauses the other. If you need both a corrected overall picture and a cohort breakdown, run them as separate passes rather than together. *** ### Limits and edge cases * **Maximum simultaneous tags**: 4. * **Maximum options per axis**: 35 (questions above this threshold are excluded from the picker). * **Compare segments**: 2 to 6 columns at a time; the cap is stated in the popup rather than silently truncating. * **Empty segments**: if a segment matches zero respondents in the current data, it still appears in the navigator but its content area shows an empty-state message, useful for confirming a slice really has no data versus the segmentation simply not being applied. * **Deleted questions**: removing a question from a survey while a tag is active simply drops the tag from the active set; URL params for missing questions are skipped on restore. * **Datasets with no segmentable axes**: if a dataset has no Single Choice / Multiple Choice / Metadata fields with bounded values, the *Add Segmentation* button is hidden; there's nothing to slice by. *** ### Best practices * **Start with one axis.** A single segmentation question already splits every chart on the page; a second axis multiplies the segments. Add the second only when the first surfaces an asymmetry you want to drill further into. * **Avoid stacking thin axes.** Splitting a 200-response dataset by *(Country × Tier × Role)* gives you cells with two or three responses each, too thin to draw conclusions from. * **Use Multiple Choice axes deliberately.** A respondent who picked five options will contribute to five segments, which inflates totals if you compare them to a Single Choice axis. The reported segment counts are accurate, but per-segment sizes don't sum to the dataset total. * **Compare against the *****Overall***** segment.** It's always the first navigator entry. Click it to confirm whether a per-segment finding actually deviates from the dataset's baseline or just looks like it does. * **Check significance before you present a gap.** A 6-point difference between two segments of 20 respondents is rarely worth a slide; the details page tells you whether it holds up. * **Share segmented URLs in stakeholder updates.** A colleague opening *"NPS Promoters in EU, last 30 days"* sees the same view you do, without having to recreate the segmentation by hand.