| Studies | One Study workspace for in-depth Mind interviews, qualitative exploration at scale, questionnaires with directional quantitative readouts, concept and message tests, and segment comparisons. |
Included capabilities| Study formats | An in-depth interview with one Mind, a moderated discussion between several Minds, a question put to a whole Audience with every Mind answering in parallel, or a questionnaire with a research method. Formats can be combined in one Study. |
|---|
| Contextual Mind chat threads | Click Chat on a Mind response to continue from that exact question with its earlier answers and relevant Study files and images, without uploading them again. The saved thread reopens with its context and links back to the source Study; it excludes other respondents’ answers and later questions. Context-aware follow-up suggestions match the current Study language. |
|---|
| Claim-level citations in interviews | In an in-depth one-to-one Mind interview, knowledge-base, web-search, analyzed-link, and uploaded-document claims show the same inline source-logo hovers as Study answer cards: public sources open their exact URLs, while private files remain non-linking and never expose storage paths. |
|---|
| Choosing Minds and Audiences | Choose Minds from the main input to open the picker immediately; an open Minds panel collapses during the handoff. Recent Mind and Audience suggestions appear once each instead of repeating to fill empty slots. |
|---|
| Quick start with Audience discovery | For a request-driven Quick start, the keyboard-accessible picker keeps the stable loading state (Twinkle) visible while it resolves and analyzes the request plus attached websites, uploads, and supported integration assets before inferring the intended Audience. It jointly ranks relevant private-library and Marketplace matches and drafts useful new Audiences. Obvious workspace actions such as exporting, downloading, or reformatting an existing result are routed to action guidance instead of inventing analyst audiences. Discovery has a bounded wait: it preserves useful matches already found or offers retry instead of loading indefinitely. It can show several reusable and NEW exact or adjacent alternatives side by side, including a same-named new option when an existing Audience matches, while keeping every result unselected until the user chooses it. Existing Audiences keep their actual stored number of Minds, including larger manually curated Audiences; only NEW Quick drafts use the plan-safe 5–20 Mind range. Every Quick request requires explicit Audience confirmation before a Study starts, including an exact existing-Audience match. Confirmed NEW drafts enter the shared Audience creation flow with the analyzed sources preserved. |
|---|
| Project Studies | Projects can start a new Study from the project’s individual Minds and retain it with the project. Study messages stay explicit when an Audience has no available Minds or is still preparing them. |
|---|
| Progress and status card | A top-right status card keeps the current Study’s Audiences together with ongoing background Studies, and returns to a finished Study when a background result is ready. Inside a Study the card is always reachable and scopes to that Study, pairing its running progress with its Audiences as a management surface: ready Audiences show their real stored Mind count, a circle-minus control removes an Audience from the Study, and an Add Audience row opens the same staged select-Audiences picker used elsewhere in the workspace. On the overview the card continues to gather every saved draft, running Study, and Audience still being prepared. Other long-running features can publish their own progress rows into the same card, and clicking a row returns to the feature that published it. |
|---|
| Fast result history | Long-running Studies initially render the newest completed results so the transcript opens quickly. Earlier results remain available in predictable batches, and selecting any deferred result from research navigation reveals it before scrolling to the requested evidence. |
|---|
|
| Study Setup and Plans | Start a Study in Quick mode or through the six-step Custom setup. Every request becomes a versioned Study Plan with reviewed questions, response formats and method, confirmed before anything runs. |
Included capabilities| Quick mode and Custom setup | Start a Study through Quick mode or use the six-step, AI-guided Custom setup to define the research goal, add a website, files, or notes, select Audiences, review goal-matched questions, and confirm the complete setup before the Study runs. |
|---|
| Audience step with Marketplace picker | The Audience step in both Quick and Custom can open the full Marketplace as an external picker: confirmed Marketplace selections return to the plan, while cancelling restores the plan’s previous selection. Switching from the composer into Custom carries the current request plus attached website, upload, and supported integration context into the setup. Contextual Audience recommendations use the same ranked saved Audiences and audience drafts as Quick, with the inferred drafts as a relevance anchor so unrelated weak library matches are not presented as recommendations. NEW drafts are selected inside step 3; the New Audience tile accepts an additional description in the same step instead of leaving the wizard for the standalone creation slide-in. Every draft remains visibly marked for creation in step 6 and enters the shared grounded Quick Audience creation pipeline only after the user confirms the complete setup. |
|---|
| Resumable Study drafts | After the first planning choice, every new Quick or Custom setup is saved as an isolated Study draft whose sidebar subtitle is Draft. Closing, navigation, reload, or a failed start preserves the exact planning step, questions, method, Audience selection, and canonical uploaded context; several drafts can be resumed independently. Checkpoints remain isolated per draft when switching chats. Planner prompts opened later inside an existing Study are deliberately excluded. The same owner-scoped, revision-safe draft lifecycle is available through the v1 API and MCP so agents can create, revise, and resume a plan while execution still requires explicit confirmation. |
|---|
| Confirmation and launch | Confirmation creates and opens the durable Study immediately; uploads and NEW Audience preparation continue behind the existing in-Study loaders, and the resulting Audiences attach to that same Study before its confirmed questions begin. The draft is removed only after the server accepts that start: at queue acceptance, the matching draft is atomically consumed so one launch cannot appear as both a Study and a duplicate Draft; a rejected launch leaves the Draft resumable. Once confirmed, the running Study owns the Study state instead of reopening the setup on reload, and its result uses the concise confirmed research subject rather than the complete internal brief as its headline. |
|---|
| Plan capture and confirmation | The planner captures the objective, subject, primary question and main source, distinguishes user questions from agent suggestions, and returns explicit confirmation questions before anything runs. Users can revise the draft or keep the executable questionnaire path simple. The confirmation shows every question’s proposed response format—open text, choice, a standard numeric scale, or a custom scale—and lets users edit it before execution; the confirmed response contract stays attached through durable processing. After confirmation, Custom shows a concise planner-generated summary with localized question-count and method metadata in the originating message instead of repeating the full brief. Confirmation follows authorship rather than question count: a single question the planner wrote itself, including its reading of a short follow-up such as “wie viel”, is confirmed before it reaches the Audience, while a question the user wrote runs directly. |
|---|
| Study language | A confirmed Study persists the current UI language in its plan and uses it throughout processor execution; legacy plans without a study language retain the saved saved owner-language fallback. An explicit Study language governs server-added confirmation copy, questions, Mind responses, category labels, per-question and grouped findings, and later whole-Study summary refreshes even when a collaborator or API-key caller uses another locale. When a Study owner starts a questionnaire run after switching the request language, the run uses that current request language; collaborator-triggered runs retain the owner’s saved language so the shared Study stays consistent. |
|---|
| Allowance checks before and during a run | A confirmed Custom Study checks its Chat allowance before generated Audience work begins, so a blocked launch opens the existing contextual upgrade path immediately instead of spending minutes preparing an Audience first. Audience-response limits are enforced before launch and before each queued question: blocked launches open the upgrade path without starting, while mid-run limits preserve partial evidence, stop the remainder, notify the user, and remain explicitly plan-limited rather than completed. |
|---|
| Question origin block | Confirmed multi-question runs render one numbered origin block from the exact questions, while abandoned accepted/pending records without a queued job remain hidden. |
|---|
| Plans via API and MCP | Confirmed executable Studies keep an immutable plan snapshot and expose durable questionnaire progress, method calculations, and flexible semantic summary blocks through the v1 API and MCP; visual-asset plans preserve heatmap-oriented output intent for UI renderers. |
|---|
| Question detection and plan review | Extract and detect questions from prompts typed in a Study or uploaded as questionnaire files, let users review the interpreted goal and source, distinguish user-authored from suggested questions, and edit multi-question plans generated from broader research objectives before they run. Quick, Custom question generation, plan revision, v1, and MCP share the same research-planning policy and fast planner model. A confirmed available method keeps its exact method and pipeline contract through the normal stream or durable questionnaire processor; experimental and planned methods remain reviewable but cannot start. Processor-side document analysis uses authenticated internal upload bytes directly when available, so Custom and questionnaire workers do not depend on turning a relative upload proxy path into a public URL. |
|---|
| Question queue | Show every waiting item and its live progress in a question queue above the Study input, append further typed queries in submission order, and keep each sub-question answer grouped under its original request. Method-generated tasks replace the optimistic question list with the processor-owned executable queue, so progress remains aligned after expansions such as MaxDiff. When a queued follow-up starts, it leaves the queue immediately and appears in the Study with its loading state before the answer arrives. Follow-ups remain behind the active primary research run. In a saved Study, queued follow-ups run durably in the processor, so they survive a page reload or closing and reopening the Study and still complete. Pending items can be removed before they start. Removing the final visible row of a queued request cancels its complete durable job, including any tasks created by later classification; after the queue is empty and no answer is running, the next message starts immediately. |
|---|
| What counts as one question | A request is only separated where the user wrote a list: explicit markers, separate lines, or consecutive questions. A multi-part question stays whole, so its options, follow-up probes, and soft-wrapped lines remain attached to the question they belong to, and pasted reference material is never mined for the rhetorical questions inside it. Anything else is interpreted as one research request instead of being divided on punctuation. |
|---|
| Completion email | When a Study with more than one question finishes a longer background run, the submitter receives a single email notification linking back to the results, including when some questions could not be completed; short runs, cancelled runs, and runs stopped by the plan's response allowance never send an email. |
|---|
| Templates | Save question sets as templates and reuse them in Custom research, privately or shared with your team. |
|---|
|
| Research Methods | Eleven deterministic research methods run end to end inside a Study: MaxDiff, NPS, top/bottom box, key driver analysis, TURF, Gabor-Granger, Van Westendorp, Kano, ranked preferences, segment comparison and conjoint. Significance is claimed only where the sample supports it. |
Included capabilities| Eleven executable methods | Eleven methods execute end to end: MaxDiff uses a server-designed forced-choice workflow, and NPS, top/bottom box scoring, key driver analysis, TURF, Gabor-Granger, Van Westendorp, Kano, ranked preferences, and segment comparison run through the same deterministic adapter contract with overall and per-Audience calculations, including pairwise between-group significance testing for segment comparison, top/bottom box scoring, NPS, and ranked preferences, using exact and small-sample-correct tests with a multiple-comparison correction across each result. Segment comparison compares the Audience groups that actually answered rather than a declared segment list, and reports a single answering group with an explanation instead of no result. A method expands one planner question into one task per item, price point, or feature pair, up to 100 questions in a Study; a design trimmed by that ceiling keeps feature pairs whole and reports what it could not ask. Conjoint executes as well: the server builds a D-optimal choice design, asks respondents to pick between complete configurations, and fits a conditional logit for per-level part-worths and attribute importance, refusing to expand a design whose part-worths would not be separately estimable. Every generated method task snapshots its exact response space into the durable job, preventing a worker from reinterpreting forced-choice options. |
|---|
| Calculations as authoritative evidence | Deterministic calculation artifacts are authoritative evidence for the result summary: the model receives their registered method, scope, methodological meaning, and exact values, decides where they materially belong, and explains them without recomputing scores. A completed summary can receive a missing valid calculation artifact later without being regenerated. Question-level Alignment and run-level method calculations share one summary-style Analysis modal, so repeated Alignment values stay attached to their exact question instead of appearing in a separate calculator popover. The processor start time is persisted so the loading counter survives reopening or reloading a Study. |
|---|
| Segment comparison | Compare responses across different Audiences to see where segments agree, disagree, or need different messaging. |
|---|
| Honest significance testing | Significance is only claimed where the sample supports it: share comparisons use the Fisher exact test rather than a normal approximation whose validity these sample sizes cannot meet, mean comparisons use Welch t on Satterthwaite degrees of freedom so small groups widen their own uncertainty, every test in one result is corrected together with Benjamini-Hochberg so an item cannot look significant merely because dozens of comparisons were run beside it, groups that answered without any internal spread cannot produce a significant difference, and a comparison that could not run at all says so instead of leaving a silence that reads as no difference. Reported p-values are floored rather than rounded to zero, because no result makes a difference certain. |
|---|
| Method selector with explanations | Methods whose calculators cannot execute yet stay visible in the method selector but cannot be chosen, and a Study still waiting for its calculation shows a dedicated calculating step in the result stream. The selector explains each method rather than only naming it: every row carries a plain-language line describing what the method is for, and hovering opens a card with the full explanation, naming the underlying technique so a market researcher recognises it — best-worst scaling, Welch t-tests, Total Unduplicated Reach and Frequency — and stating the same standing limitation the results carry, so the choice is informed before the Study runs rather than only after. An "Advanced" tag is no longer shown, since it signalled difficulty rather than fit; "Not available yet" remains, because it changes what can be picked. The explanations are written natively in all nine interface languages, using each language's own words for an Audience and a Study rather than the English ones. |
|---|
|
| Audience Responses | Ask an Audience once and receive parallel synthetic responses from its Minds, grounded in the websites, files, images, video and audio you supply. |
Included capabilities| Grounded in supplied material | When a Study message includes a successfully loaded website, file, source, screenshot, or image, each retained Mind response must use a concrete detail from that supplied material instead of treating the asset as missing. |
|---|
| Video and audio analysis | For an uploaded video in a normal Study question, attachment analysis combines the available transcript and a bounded soundtrack timeline with a bounded, scene-aware set of representative frames. Ad-length videos analyze the full soundtrack for material music, sound effects, ambience, silence, vocal emotion, and sonic-logo moments; longer videos use at most twelve uniformly distributed eight-second audio clips, while any audio-analysis failure leaves transcript and visual processing available. Standalone MP3, WAV, M4A, OGG, FLAC, and audio-WebM attachments use the same transcript and bounded non-speech soundtrack analysis, with either evidence path remaining available when the other fails. Visual sampling preserves opening and closing-scene coverage, distributes selections across fast-cut videos, and falls back to three evenly spaced frames when reliable cuts cannot be detected. |
|---|
| Durable background processing | Larger Study turns continue durably in the background, keep live progress visible, and recover after a page reload. For large Audiences, response work is split into durable batches across available processing capacity; each Mind response is streamed as soon as it finishes without waiting for a whole batch, while bounded recovery retries only the Minds still missing. |
|---|
| Headline, summary and charts | Each completed result reveals its headline and synthesized research summary first. A single question adds one compact answer chart before its supporting Mind-level response rows in the collapsible Evidence section. Multi-question research adds one to four side-by-side key-finding overview charts, normally two, selected against the complete research request and overall conclusion rather than by percentage alone. Every chart is backed by verified answer counts: one focused percentage or score for a standalone result, or a two- or three-value comparison whenever the finding makes a comparative claim. The result gives every processed question its own foldout, answer chart, Audience distributions, and Mind-level evidence. |
|---|
| Attachment comparisons and asset assignment | Attachment-comparison results show original image creatives and first-frame video previews directly above their A/B/C columns, match variants by persisted filename and occurrence rather than raw attachment position, and leave historical votes and findings unchanged when older results are reopened. Multiple block-level assets require explicit question assignments across videos, images, documents and websites: each asset must be used by at least one question, and an empty assignment means no material. MCP, v1 and the study builder enforce the same execution contract. Result cards use the saved per-question mapping, keep each assigned video accessible by name with its own saved scene heatmap and Mind reactions, and expose document and website references without changing the response type or attachment-comparison layout. |
|---|
| Source citations on answers | When a Mind uses retrieved knowledge-base, audience-grounding, or web evidence, its hover card and Closer Look show claim-level source logos through the same inline citation system as Marketplace descriptions. Public URLs render as clickable source logos with domain tooltips; private knowledge files remain non-linking document sources and never expose storage paths. Generation-time attribution is stored with the answer, while historical answers can be lazily checked against the Mind’s current knowledge base and Audience grounding. The v1 result payload exposes optional citation markers, source identities, and provenance without returning the underlying private knowledge text. |
|---|
|
| Stimulus Testing | Test copy, landing pages, screenshots, decks, product concepts, and competitor material inside a Study. |
| Heatmap Testing | Run website, video or image heatmaps to capture simulated attention and interaction signals. |
Included capabilities| Evidence from the observed page | Visible website and media evidence is authoritative over an earlier unsupported assumption, and website summaries are derived from the observed page trace rather than a stale pre-browse answer. |
|---|
| Video and transcript handling | Uploaded media uses a configured backup speech service when the primary transcription provider is unavailable. Video reactions connect the full available transcript to exact visible elements in sampled frames and fall back to visual-only analysis when transcription remains unavailable. |
|---|
| Several images or videos per question | Each uploaded image is analyzed as an independent creative in one result, with lettered tabs that retain its filename. For questions with several assigned videos, selecting a video selects its own saved scenes, heatmap and Mind reactions. Missing analysis starts for that question and asset; switching assets preserves existing analyses and the overall question results. |
|---|
| Consistent across UI, API and MCP | UI, v1 and MCP use the same target validation and analysis pipeline. |
|---|
|
| Scheduled Studies | Schedule a Study to rerun daily, weekly, monthly, or quarterly. Runs continue reliably in the background, recover from worker interruptions, and preserve progress for historical review and export. |
| Research Credibility Controls | The evidence boundary stays visible from Audience creation through Study answers: modelled assumptions are never presented as measured facts, and every Study states its own completeness. |
Included capabilities| Modelled assumptions stay labelled | When source material is readable but no evidence-backed distribution can be applied, Minds retains every uploaded material for review and holds that Audience to modeled assumptions rather than measured customer facts: its Minds are instructed to treat persona details as synthetic scenario assumptions, so the boundary is enforced in the answers themselves rather than as a badge the reader has to notice and interpret. Synthetic persona generation and panel answers avoid unsupported exact employee counts, revenue, budgets, income, policy leakage, incident counts, and market share. |
|---|
| Briefs and sources travel with the plan | Attached study briefs feed the planner itself; every explicitly named criterion is preserved as a separate reviewable question, and the complete Study Plan survives the Audience handoff into confirmation instead of only the first question running. A Study also keeps the material its request carried. A website or file named in the message, attached in the composer, or added while reviewing the plan becomes the plan’s confirmed source and travels with every question of the run, so replacing the original wording with planner questions can no longer lose the page or document under review. A context file that cannot be read stops the Study instead of publishing findings against material nobody opened. |
|---|
| Completeness of a finished Study | A finished Study also states its own completeness. When a question fails, the Study concludes from the questions that did answer and reports how many those were, instead of discarding a run whose remaining answers are still valid; retrying the failed question rebuilds the overall conclusion against the fuller evidence. A Study that stopped responding says so with the number of answers it collected and can be concluded from them on request, rather than showing an open-ended progress state. A conclusion assembled from the individual question findings, without a full synthesis, is labelled as such. |
|---|
|
