# UX research report video evidence cards Market Research Report - Global

**Generated on:** 2026-10-16 16:16:56.264700  
**Industry:** UX research report video evidence cards  
**Geography:** Global  
**Details:** # Deep Research Prompt: Respondent Video Evidence Cards in Research Reports (VOXDEV-3708)

## Context

Voxpopme is a B2B video research platform used by enterprise insight and market research teams. Respondents answer questions on video, recorded on mobile (mostly portrait) or webcam (mostly landscape), so clip aspect ratios are mixed.

Findings reports and Share Pages, often generated by Compass (Voxpopme's agentic AI research assistant), present findings backed by evidence "cards". Each card pairs a respondent video clip with a quote that supports a finding.

Current problems:
- Videos inside cards are heavily cropped, so faces and context are lost, especially for portrait clips in landscape slots and vice versa.
- In Compass output, quotes are detected in generated text and the respondent video is injected inline. This breaks the flow of paragraphs and can make the video look unrelated to the claim it sits next to.
- The current component works acceptably in lists and as a single hero statement, but not inside running prose.

Direction already agreed internally:
- A new findings report design (Brand 2026) places the video side by side with the quote rather than above or behind it.
- Two card types are proposed: a large "Hero" video panel for headline evidence, and a smaller side-by-side video plus quote card for supporting evidence, so a full report does not feel too spaced out.
- Engineering suggests that inside paragraphs, quotes with attribution may work better than embedded video, and that any inline video needs careful UX thought.

Goal: define a robust, evidence-based pattern for showing respondent video with quotes across reports, Share Pages and AI-generated output, that a designer can specify and an engineer can build directly.

## Research sections

### 1. Aspect ratios and cropping
- What are the established techniques for displaying mixed portrait and landscape video in a fixed card (object-fit cover vs contain, letterboxing, pillarboxing, blurred or colour-sampled background fill, adaptive containers)?
- How do leading products (social platforms, news sites, video research tools) handle portrait video in landscape layouts and the reverse?
- Is face-aware or saliency-based smart cropping viable for respondent video? What are the accuracy, cost and trust risks?
- Should card dimensions adapt to the clip's native aspect ratio, or should the clip adapt to a fixed card? What are the trade-offs for layout consistency?
- What poster frame or thumbnail strategy best represents a respondent before play (face detection, frame at quote start, custom selection)?

### 2. Card taxonomy and decision rules
- What evidence card variants are justified: Hero video panel, compact side-by-side video plus quote, quote only with attribution, grouped or carousel evidence, others?
- What clear rules should decide which variant is used (importance of finding, position in report, number of supporting clips, clip length, quote length, aspect ratio)?
- How many Hero panels per section or per report keep emphasis meaningful without visual fatigue?
- What is the optimal ratio of video to quote width in a side-by-side card, and how should long quotes be handled (truncation, expand, scroll)?
- How should multiple respondents supporting the same finding be shown?

### 3. Video evidence inside running text vs lists
- What patterns exist for attaching media evidence to prose without interrupting reading (inline citations, footnote-style markers, margin or sidenote evidence, hover or tap previews, expandable evidence, end-of-paragraph evidence blocks)?
- What does research on reading comprehension and flow say about interrupting text with embedded media?
- How do long-form journalism, academic publishing and qualitative research tools link claims to source evidence?
- When should a quote in prose become a video card, and when should it stay as text with attribution and an optional "watch clip" affordance?
- How should list-based findings differ from narrative findings in evidence treatment?

### 4. Trust, relevance and quote-to-video alignment
- How can the design make it obvious which claim a video supports, so evidence never looks unrelated?
- Should clips start at the exact quote timestamp, and how should surrounding context be made available without misrepresenting the respondent?
- What attribution details build credibility for enterprise stakeholders (respondent descriptor, segment, date, study name) while protecting respondent privacy?
- How should AI-selected evidence be signposted so users can verify and challenge it?

### 5. Rhythm and scale across a full multi-page report
- How should Hero and compact cards be distributed to create visual rhythm across many pages?
- What density benchmarks exist for evidence per finding in professional research reports?
- How should the pattern behave in different report lengths (one-page summary vs 30-page findings report)?
- How do layout grids, section openers and summaries interact with video evidence?

### 6. Responsive behaviour
- How should Hero and side-by-side cards adapt at mobile, tablet and desktop widths (stacking order, video above or below quote, max heights)?
- Are container queries preferable to viewport breakpoints for cards that appear in varied layouts?
- How should portrait clips behave on mobile, where they may dominate the viewport?
- What maximum and minimum video sizes keep faces legible and avoid excessive scrolling?

### 7. Accessibility and compliance
- What does WCAG 2.2 AA require for prerecorded video in this context (captions, transcripts, audio description applicability)?
- Best practice for keyboard operation, focus order and visible focus across cards and inline evidence triggers.
- Screen reader patterns: how should a card announce the quote, the respondent and the video controls without duplication?
- Autoplay, motion and prefers-reduced-motion: what defaults and behaviours are recommended?
- Cognitive accessibility: how to avoid overload when many videos sit on one page.
- Relevant obligations under the European Accessibility Act, UK and US standards for enterprise software.

### 8. Technical implementation
- How to store and expose clip aspect ratio metadata so layout can be decided before video loads (preventing layout shift).
- Recommended CSS approaches (aspect-ratio, object-fit, grid, container queries) with example specifications.
- Performance: lazy loading, poster images, adaptive streaming, limiting simultaneous players on long pages.
- How the pattern should render in exports (PDF, slides, print) where video cannot play.
- Component API suggestions for a design system (props, variants, states).

### 9. Rules for AI-generated output (Compass)
- How should an AI agent decide where evidence goes and which card type to use? Propose a decision framework or schema.
- What structured output format would let Compass place evidence reliably rather than injecting video wherever a quote is detected?
- How do other AI research or analysis tools present source evidence alongside generated summaries?
- What guardrails prevent misleading pairings of generated claims and respondent clips?

### 10. Competitive and adjacent landscape
- How do Dovetail, Marvin, Condens, UserTesting, Medallia LivingLens, Discuss.io, Qualtrics, Remesh, Grain and Gong present video evidence with quotes in reports or highlight reels?
- What patterns from news (for example interactive long-form articles), social platforms and documentation tools are transferable?
- Where are competitors weak, and what could Voxpopme do that is genuinely differentiating?

### 11. Edge cases and failure modes
- Very short clips, very long quotes, clips with no visible face, poor lighting or low resolution.
- Missing or inaccurate captions and transcripts, non-English clips and translated quotes.
- Respondents who must be anonymised or blurred.
- Share Pages viewed by senior stakeholders on mobile, in email previews or with slow connections.
- Pages with many clips, failed video loads and fallback states.

### 12. Validation and success measures
- What metrics show the new pattern works (play rate, completion, time on report, scroll depth, stakeholder trust, sharing)?
- What usability test tasks and comparisons would validate Hero vs compact vs quote-only treatments?
- What qualitative signals from enterprise clients should be captured?

## Output instructions
- Provide specifications a designer and developer can implement directly, using specific values where possible (sizes, ratios, breakpoints, counts), not just ranges.
- Cite sources for every factual or research-based claim, and distinguish evidence from opinion.
- Flag trade-offs explicitly, and state a clear recommendation for each section.
- Highlight anything surprising or that contradicts the direction already agreed internally.
- End with a summary decision table: content situation, recommended card type, and reason.
- Use UK English.

---

# Video Evidence Cards That Preserve Context and Trust

## Executive Summary

**Recommendation for VOXDEV-3708:** retain the proposed Hero and compact side-by-side designs, but add a first-class **attributed quote with an explicit evidence link**. Treat evidence as a structured relationship between a finding and a source span, not as media inserted wherever Compass detects quotation marks.

- **Preserve The Whole Respondent**: `object-fit: contain` preserves the source frame; `cover` clips it when source and container ratios differ. YouTube itself adapts player dimensions and sometimes adds padding. -> Default to contained footage, with orientation-aware media slots. [6][3]
- **Two Cards Are Not Enough**: Google's source-citation interface separates readable answers from optional inspection of quoted context. -> Use quote-only evidence links in running prose; reserve visual cards for deliberate evidence blocks. [43]
- **Keep Related Evidence Close, Not Intrusive**: Spatial-contiguity research supports proximity between related words and pictures, while seductive-details research warns against interesting but irrelevant additions. Neither establishes an optimal enterprise video-card layout. -> Attach evidence to a specific claim without forcing playback into its sentence. [48][9]
- **Expose The Provenance Already Promised**: Voxpopme publicly says Compass validates quotes against original video timings and surfaces sample sizes. -> Make source spans, context and respondent counts inspectable in the interface. [13]
- **Source Links Are Competitive Baseline**: Dovetail documents source-linked insights, Discuss links takeaways to quotes and clips, and Grain offers cited reports. -> Differentiate through faithful framing, claim-level alignment, accessible playback and controlled sharing, not merely an AI-citation badge. [23][19][30]
- **Smart Cropping Adds Operational Risk**: Cloudinary documents analysis taking seconds or minutes and transformation-count conditions; its face modes can focus on the largest face or all detected faces. -> Make reframing optional and reviewable, never the default evidentiary rendering. [12]
- **Accessibility Is Content Plus Interaction**: Captions are required for ordinary prerecorded synchronised video; transcripts do not replace that requirement. Simple talking-head footage may not need additional audio description when important visual information is already conveyed appropriately. -> Build an accessibility publishing checklist into evidence creation. [14][36]
- **Optimise Decision Quality Before Plays**: The original HEART paper explicitly warns that engagement can be less meaningful in enterprise settings. -> Make correct claim-to-source identification, comprehension and verification effort primary measures; treat plays and time on report as diagnostic signals. [58]

**Scope and interpretation.** This is a global product-pattern and competitive assessment of evidence presentation within research platforms, not a revenue forecast for a separately defined software market. Competitor statements below describe documented capabilities, not independently measured performance. **All proposed dimensions, counts, thresholds and acceptance targets are product specifications to validate, not established industry optima.**

## 1. Preserve Mixed-Ratio Footage Without Destabilising Reports

### YouTube demonstrates adaptation, not universal cropping

YouTube specifies **16:9** as its standard computer-player ratio, but says the player adapts to the video and device. For some combinations, including vertical video on desktop, it may add padding. It explicitly advises against baking padding into the uploaded video because that obstructs dynamic adaptation. [3]

The transferable decision is to separate the **source frame**, **playback slot** and **outer evidence card**. Voxpopme can keep a consistent report grid without forcing every recording to fill an identical rectangle. The recommendation is an outer card governed by report layout, an orientation-aware slot, and uncropped video inside it.

| Technique | What it does | Trade-off | Recommended use |
|---|---|---|---|
| `contain` | Preserves aspect ratio and the whole image; may produce letterboxing or pillarboxing. [6] | Less of the slot is occupied by footage | Default for respondent evidence |
| `cover` | Preserves aspect ratio but clips mismatched content. [6] | Faces, hands, screens or context can be excluded | Only an explicitly reviewed derivative; retain original view |
| `fill` | Can stretch content when ratios differ. [6] | Distorts the source | Never use for respondent footage |
| Neutral letterbox/pillarbox | Fills unused slot space without altering footage | Visible bars, but minimal visual competition | Default background: a quiet, high-contrast neutral |
| Blurred duplicate background | Adds a decorative full-slot layer behind contained footage | Extra visual movement and processing; can confuse frame boundaries | Optional static treatment, not default |
| Colour-sampled background | Uses source-derived colour behind the full frame | Variable contrast and possible distraction | Optional, with a fixed contrast-safe overlay |
| Native-ratio container | Lets the frame dictate slot dimensions | Strong source fidelity, inconsistent card heights | Expanded viewer; selectively in editorial Hero panels |

**Geometry, not a usability result:** a contained 9:16 image in a 16:9 slot occupies approximately **31.6%** of the slot area. Conversely, filling that slot with `cover` retains approximately 31.6% of the portrait image area. This is why merely replacing cropping with containment inside the same wide thumbnail may still leave the respondent too small.

### Smart cropping is feasible, but not a trust substitute

Cloudinary provides face-aware and automatically moving crop regions. Its documented `face` mode targets the largest detected face, while `faces` targets all detected faces; analysis may take seconds or minutes and has additional transformation-count conditions. Google's AutoFlip also demonstrates saliency-based reframing, but its documented legacy solution is no longer supported. [12][10]

For research evidence, a technically successful face crop can still remove the object a respondent is discussing. No respondent-specific accuracy or cost-per-clip benchmark is established by these documents. **Recommendation:** ship containment first. Offer reviewed cropping later, with an original-frame toggle, crop provenance and manual override; never use face detection to infer respondent identity or relevance.

**Poster specification:** select a clear frame within the quoted source span, preferably near its beginning; avoid a blink, transition or misleading expression. Offer researcher selection and apply the same privacy treatment as the video. Explicit-offset extraction is documented by Cloudinary, whereas its automatic thumbnail mode selects using colour distribution, not quote relevance. Do not assume "automatic thumbnail" means "representative evidence". [34]

## 2. Use Four Evidence Treatments With Deterministic Selection Rules

### Dovetail separates evidence objects from their presentation

Dovetail turns highlighted transcript passages into individual searchable clips, allows highlights to be grouped under tags and embedded in documents, and supports links that preview highlighted content in other tools. This establishes a useful architectural distinction: the evidence object can remain stable while its presentation changes. [26]

Its AI-highlight workflow also includes **approve or reject** actions rather than making every suggested highlight final. Voxpopme should similarly separate selection, approval and rendering: a researcher-approved source span might appear as a Hero in an executive readout and as a simple citation in a detailed narrative. [26]

### Proposed launch taxonomy

| Variant | Selection rule | Default specification | Important exclusion |
|---|---|---|---|
| Hero evidence panel | A section-defining finding where the delivery or visible behaviour adds material meaning | One clip; full quote up to 60 words; proposed clip target 10-45 seconds; desktop quote:video split 40:60 | Not awarded automatically to the longest or most emotional clip |
| Compact evidence card | A discrete supporting finding in a list, grid or end-of-paragraph block | Quote preview up to 30 words; proposed clip target 5-60 seconds; video:quote approximately 40:60, capped at 240px media width on wide containers | Not injected into a sentence |
| Attributed quote plus evidence link | Running prose, a text-led summary, insufficient media quality, or evidence that needs no visible performance | Quote, safe descriptor and "Watch clip" or "View source" | Not an unattributed quotation or an unlinked AI paraphrase |
| Evidence group | Multiple sources support, qualify or challenge one finding | One representative visible item; "View all N examples"; expandable list with respondent and source labels | No autoplay carousel; no implication that N clips means N people |

The numerical thresholds above are **editorial defaults**. A short but decisive behavioural clip can justify promotion; a long recording should usually yield a focused, reviewed excerpt rather than a giant card.

**Width rule:** use approximately **40% video / 60% quote** for compact cards in medium containers. Once the container reaches 720px, cap the media column at **240px** and give the remaining width to text. This avoids a small supporting quote acquiring a disproportionately large video panel. For a Hero, reverse emphasis to **60% video / 40% quote**, subject to responsive stacking.

**Long quotes:** keep the full verbatim source span in the evidence record. Display a 30-word preview with a visible "..." and an **Expand full quote** button; never use a nested scrolling text box. If the omitted portion changes meaning, do not truncate: use a text-led block or a better source span. Expansion must not change the underlying quotation or clip binding.

**Multiple respondents:** show variety deliberately, including a materially contradictory example where relevant. Group by finding first, then segment or stance. A manually operated carousel may be an optional presentation mode, but a labelled list is the default because it exposes membership and supports comparison.

**Recommendation:** launch three visual states - Hero, compact and quote-only - plus a grouping wrapper. Do not build separate evidence models for each appearance.

## 3. Link Evidence To Prose Instead Of Interrupting It

### Citation interfaces preserve a readable primary argument

Google's documented citation pattern lets readers inspect quoted text on hover and select a citation to reach the quote in context. Tufte CSS offers another useful pattern: sidenotes sit beside text on sufficiently large screens and become user-toggled content on smaller screens. Neither requires replacing part of a sentence with a media player. [43]

The principle is **progressive disclosure**, not hiding evidence. A stakeholder should understand the report without playing every clip, while being able to verify a contested claim immediately. Hover can be a desktop enhancement, but tap and keyboard activation must offer the same evidence; no essential information should exist only on hover.

### What the reading research supports - and does not

The **spatial contiguity principle** predicts benefits when related words and pictures are placed near one another. A meta-analysis reports benefits for integrated designs, but examines multimedia learning rather than enterprise research reports. Separately, the seductive-details literature finds that interesting, irrelevant additions can hinder learning, with important moderating conditions. These findings support relevance and proximity, not a blanket ban on media or proof that side-by-side cards are optimal. [48][9]

This is an important counterpoint to "video always breaks reading flow": **relevant, well-signalled media can help**. The problem in the supplied brief is the uncontrolled placement and ambiguous relationship, not video as a medium. Treat the expected benefit of an evidence drawer or end-of-paragraph block as a testable design hypothesis.

### Proposed placement policy

| Reading context | Evidence attachment | Behaviour |
|---|---|---|
| Claim within a paragraph | Explicit source marker immediately after the supported clause | Opens a labelled evidence viewer without replacing paragraph text |
| Quote within prose | Attribution followed by "Watch clip, 24 seconds", using actual duration | Keeps the quotation readable; playback is optional |
| Paragraph with several supported claims | Separate claim markers; optional evidence block after the paragraph | Never assign one clip ambiguously to the whole paragraph |
| Short list of findings | One compact card directly under each applicable finding | Relationship is visible through shared heading and spacing |
| Wide, text-led report | Optional margin evidence aligned with its claim | On narrow containers, convert to a trigger or paragraph-end block |
| Executive narrative | At most a deliberately selected Hero at a finding boundary | Do not split a sentence or surround prose with floating players |

Use ordinary footnote-like reference conventions where they fit, but make the action understandable: "Evidence for checkout confusion" is better than an unexplained play icon. Academic citation-style linking supplies the model of claim-to-source navigation; video adds a time-based source span.

**Recommendation:** quote plus attribution is the default inside prose. Promote it to a card only when seeing or hearing that respondent adds meaning worth interrupting the reading sequence.

## 4. Make Claim, Quotation And Source Span Independently Verifiable

### Compass's public promise should become a visible interaction

Voxpopme says Compass validates every quote against original video timings, surfaces sample sizes and identifies evidence that contradicts assumptions. The new component should make that promise inspectable, rather than expressing it through a generic "AI-powered" badge. [13]

NIST identifies a specific failure mode: generated outputs can contain confident errors and fabricated citations that appear to justify an answer. Source availability therefore does not establish that a source supports the generated claim. A valid link and a valid claim-source relationship are separate checks. [8]

### Proposed evidence anatomy

Each evidence unit should show, in order: **supported finding label; verbatim quotation; safe respondent descriptor; play action and actual duration; study and fieldwork context; source/context action**. Keep the finding visually distinct from the respondent's words. Do not put an AI interpretation inside quotation marks.

The source record should retain participant ID, study ID, question ID, transcript version, source language, word- or segment-level timing, clip boundaries, consent scope and approval status. The stakeholder surface should show only the subset permitted for that audience. Recommended attribution is a pseudonymous label plus a relevant broad segment, study name and fieldwork month, where approved; precise job titles, locations or dates can be unnecessarily identifying.

**Playback alignment:** start at the reviewed beginning of the quoted passage, not the beginning of the full response. Boundaries should respect complete words and meaningful qualifications. If the quote depends on the interviewer's question or an earlier "but", enlarge the evidence span rather than silently removing that context.

Add **View surrounding context**, initially offering 10 seconds before and after where available, followed by a permission-checked full-source action. Show the selected span on a transcript timeline. Do not concatenate non-contiguous passages as one continuous quotation; present separate excerpts with visible boundaries.

**AI signposting:** use "AI-selected evidence" and a distinct "Researcher checked" state only when that review occurred. Include "Why this evidence?", "Open source" and "Report mismatch". A badge is provenance, not a confidence score or a guarantee of truth.

### Privacy must survive the whole sharing chain

MRS guidance limits recordings to the research purpose. Condens demonstrates that practical protection may involve face or whole-video anonymisation, voice distortion and selective text, audio or video redaction. Blurring a face alone is therefore not a complete sharing policy. [2][29]

Apply permissions to the quote, poster, captions, audio, source link and export, not just the player. Where a descriptive segment could identify someone, show a broader label or quote-only evidence. Revoking access must remove cached derivatives and future exports where operationally possible, with clear limits for already downloaded files.

**Recommendation:** make claim support, transcript fidelity, privacy approval and source availability separate statuses. None should be inferred from successful video playback.

## 5. Give Long Reports Rhythm Without Turning Every Finding Into A Film

### UserTesting shows why report structure and evidence are separate concerns

UserTesting Reports can include clips and highlights alongside headings, tables, dividers, checklists and a table of contents. They can also be presented directly or exported as PDF. Evidence therefore sits within a larger information architecture rather than determining every page's structure. [22]

Dovetail's reusable highlights suggest the complementary approach: keep a complete evidence collection available without rendering every item in the main narrative. The report can communicate a hierarchy while the repository preserves the underlying material. [26]

There is **no established optimum for Hero counts or evidence-per-finding density in the sources reviewed**. Do not import a social-feed density target into an enterprise report. The following is an implementable starting policy, not a market benchmark.

| Report situation | Proposed Hero budget | Main-body evidence policy | Supporting material |
|---|---|---|---|
| One-page executive summary | Maximum 1 | Text-first findings; 1 primary evidence action per major finding | Grouped examples behind links |
| Short findings readout | Maximum 1 per section, 2 across the report | Normally 1 visible compact card per finding | Additional examples collapsed |
| 30-page report | Maximum 1 per section, 3 across the report | 1 visible compact card per finding; a second only for a meaningful contrast or qualification | Searchable or grouped evidence appendix |
| Source-heavy appendix | No automatic Hero promotion | Group by finding, then respondent/segment | Full source navigation subject to permissions |

These are **caps, not quotas**. A report can legitimately contain no Hero panels. An illustrative 30-page report with ten principal findings might contain three Heroes and ten visible compact evidence cards, with the rest accessible through evidence groups; this is a layout scenario, not a measured norm.

Use a stable reading column with a proposed **960px report maximum** and **68ch narrative measure**. Allow section openers and Heroes to use the wider grid; retain the narrower measure for paragraphs. Place a Hero after the finding has been stated, so the reader knows why it matters before deciding to watch.

Keep section summaries predominantly textual, with source actions rather than repeated players. Do not repeat the same full card in an executive summary, detailed finding and conclusion: use a compact cross-reference to the same evidence ID. In paginated exports, keep the quote, attribution and poster together rather than allowing a page break between them.

**Trade-off:** less visible media can reduce incidental plays, but can also improve report scanability. **Recommendation:** spend visual prominence on changes in meaning, not on achieving a fixed video density.

## 6. Make Components Respond To Their Container, Not Just The Screen

### YouTube adaptation and CSS containment solve different parts of the problem

YouTube's adaptation concerns the video and viewer's device. CSS container queries address a different issue: a component can style its descendants according to the size of its containing element rather than the viewport. This matters when the same evidence card appears in a full-width report, a narrow column or an evidence drawer. [3][38]

MDN documents declaring a query container with `container-type: inline-size`, and notes that queries use a containing ancestor. Voxpopme should therefore place containment on an outer wrapper and query the card inside it. A wide desktop viewport must not force a two-column card into a narrow report sidebar. [38][41]

### Proposed responsive specification

All values below are **CSS pixels and launch defaults**, measured against the card container's inline size.

| Container width | Compact treatment | Hero treatment | Media policy |
|---|---|---|---|
| Below 480px | Stack quote, attribution, then media | Stack in the same order | Portrait/square slot 1:1; landscape 16:9; maximum media height `min(320px, 60svh)` |
| 480-719px | Quote:video 60:40 with 16px gap | Remain stacked | Compact media maximum height 240px; footage contained |
| 720px and above | Remaining text width plus 240px media column; 24px gap | Quote:video 40:60; 24px gap | Hero landscape slot 16:9; portrait slot 4:3; maximum height 480px |
| Expanded viewer | Text/transcript and player may use separate regions | Same shared viewer | Original frame contained; media maximum height 75svh |

The consistent **quote-first DOM order** preserves the text-led reading sequence on mobile. On desktop, put the quote on the left and media on the right rather than visually reversing the DOM. This is a recommendation, not a claim that all users prefer video below text.

A portrait recording should not automatically produce a full-height 9:16 card on mobile. A square preview slot with the entire portrait image contained limits page domination; the viewer provides a larger uncropped view when requested. This deliberately trades immediate image size for reading continuity.

Treat small compact media as a **poster-plus-play trigger**, opening the shared viewer, rather than squeezing a complete control bar into a thumbnail. With the proposed 480px threshold, 16px padding and gap, the medium compact media column is approximately 173px wide at its smallest. That is a preview size, not a proven minimum for recognising expressions.

Use proposed **44px minimum standalone play and expand targets**. Do not claim that a particular pixel width guarantees face legibility: lighting, framing and source resolution vary. Include actual low-quality recordings in testing, and keep an obvious enlarge action.

**Recommendation:** use container queries for component layout, viewport/small-viewport units for player height limits, and a stacked fallback when container queries are unavailable.

## 7. Meet Accessibility Requirements Without Overstating The Law

### The BBC player pattern favours native semantics and progressive enhancement

BBC GEL recommends preserving native video controls until custom controls are available, grouping the player and controls semantically, using native buttons, and making timeline interaction accessible to keyboard and screen-reader users. Its guidance also calls for clear focus and caption support. [56]

The implication is practical: the new visual shell should not become a reason to replace a functioning accessible player with custom icons. Prefer an existing audited player; if custom controls are necessary, retain a usable fallback and test the complete experience, not merely the card's static appearance.

### Prerecorded-media requirements

For ordinary prerecorded synchronised video, **WCAG 1.2.2, Level A, requires captions**, subject to its clearly labelled media-alternative-for-text exception. A selected quote is not automatically a complete caption or transcript alternative. Level A 1.2.3 requires an alternative for time-based media or audio description; **Level AA 1.2.5 requires audio description for relevant prerecorded visual content**. A transcript alone does not generally replace that AA requirement. [14]

There is an important qualification for this product: W3C technique G203 explains that additional audio description is unnecessary for a single speaker against an unchanging background when there is no important time-based visual information. Associated static text can convey speaker identification and environmental context. If a respondent points to a product, demonstrates a task, or relies on visible text not conveyed in audio, reassess the description requirement. [36]

### Component acceptance rules

- **Keyboard:** every action works without a pointer; no keyboard trap; logical focus order; visible focus that is not entirely obscured. Opening the viewer moves focus appropriately; closing returns focus to its trigger. [14]
- **Targets and contrast:** WCAG 2.2 AA 2.5.8 sets a **24 x 24px** pointer-target requirement with spacing, inline, equivalent-control, user-agent and essential exceptions. Use 44px standalone targets as the product's more generous default. Text contrast is normally **4.5:1**, or **3:1** for large text, with the standard's exceptions. [37][14]
- **Screen readers:** label the evidence group with a short finding heading; expose the quote once as text and attribution separately. Give the play control an action label with respondent descriptor and duration. Use empty alt text for a decorative poster when the adjacent control already identifies it. Do not use the entire quote as the player's accessible name.
- **Captions and transcript:** provide reviewed captions, speaker identification and relevant non-speech information; offer the full excerpt transcript separately. Keep translated captions and original-language material distinguishable.
- **Autoplay and motion:** proposed policy is no autoplay, no animated poster and no auto-advancing carousel. Respect `prefers-reduced-motion`. WCAG's automatic-audio and moving-content provisions remain relevant, but this policy intentionally avoids creating the problem. [14]
- **Cognitive load:** one active player, consistent controls, no essential hover-only content, and no requirement to watch every clip to understand the finding.

### Jurisdictional scope for a global enterprise product

| Jurisdiction | Relevant obligation | Implication for Voxpopme |
|---|---|---|
| EU consumer services | The EAA covers specified services provided to consumers after 28 June 2025; its consumer definition excludes activity within a person's trade or profession. [40] | Enterprise-only SaaS is not automatically covered simply because it is online. Assess the actual service, audience and national implementation. |
| EU public sector | The Web Accessibility Directive concerns public-sector websites/apps; its legal requirements are underpinned by EN 301 549. [55] | Procurement and customer deployment can create requirements beyond a general consumer-service analysis. |
| UK | Public-sector guidance specifies WCAG 2.2 AA and an accessibility statement; service providers also have reasonable-adjustment duties under the Equality Act 2010 or Northern Ireland's DDA. [57] | Distinguish public-sector regulations from wider service-provider duties; do not market a blanket legal certification. |
| US federal | Section 508 applies when federal agencies develop, procure, maintain or use ICT. [35] | Support federal procurement assessments and contractual accessibility requirements. |
| US ADA | This guidance addresses website accessibility for state and local governments and businesses open to the public under the ADA. [47] [47] [47] It states that covered entities have flexibility in how they comply, and DOJ does not set detailed technical standards; WCAG is helpful guidance, not a blanket WCAG 2.2 mandate for all private B2B software. [47] [47] | Assess customer and service scope rather than claiming that every private B2B product has the same statutory WCAG version. |

**Recommendation:** make WCAG 2.2 AA the engineering target, commission a scoped accessibility assessment and maintain procurement-ready conformance documentation. Legal applicability and technical conformance are related, but not interchangeable.

## 8. Implement One Evidence Model With Lightweight Renderers

### Reserve layout before downloading media

MDN documents that `videoWidth` can be zero before metadata is available; dimensions are known once `loadedmetadata` fires. Waiting for that browser event to discover orientation would therefore make layout dependent on a media request. Store display dimensions during ingestion and send them with the report instead. The storage recommendation is an engineering inference from the documented browser lifecycle. [54][53]

Use **display width/height after orientation normalisation**, not merely raw encoded dimensions. Store rotation, pixel/display aspect information where needed, duration, source language, poster reference, source span and available caption tracks. Derive orientation from normalised display dimensions; reserve a deterministic fallback slot when metadata is unavailable.

### Proposed data and component contract

| Contract area | Required fields or behaviour |
|---|---|
| Evidence identity | `evidenceId`, `studyId`, `sourceId`, `participantId`, `questionId`, immutable source/transcript versions |
| Media | `displayWidth`, `displayHeight`, `rotation`, `durationMs`, `posterId`, rendition references |
| Quotation | `quoteText`, `quoteStartMs`, `quoteEndMs`, transcript span IDs, original language, optional translated text and review state |
| Governance | `consentScope`, audience permissions, redaction version, retention/revocation state, human approval |
| Component props | `evidenceId`, `claimId`, `variant`, `presentationContext`, `expanded`, `showAttribution`, `onOpenSource`, `onReportMismatch` |
| Variants | `hero`, `compact`, `quote`; separate `EvidenceGroup` wrapper |
| States | Poster, loading, ready, playing, paused, ended, failed, restricted, withdrawn, captions unavailable |

Pass IDs and permitted presentation data to the component. Do not let client-supplied URLs or generated HTML determine source access.

### CSS starting point

This implements the proposed quote-first layout. `--slot` describes the **presentation slot**, not the source ratio; the source remains contained. The wrapper is the query container. `aspect-ratio` needs at least one automatic dimension to influence box sizing. [38][33]

```css
.evidence-wrap { container: evidence / inline-size; }
.evidence {
  display: grid;
  grid-template-columns: minmax(0, 1fr);
  gap: 16px;
  padding: 16px;
}
.evidence__copy { min-inline-size: 0; overflow-wrap: anywhere; }
.evidence__media {
  position: relative;
  inline-size: 100%;
  aspect-ratio: var(--slot, 16 / 9);
  max-block-size: min(320px, 60svh);
  background: #17191c;
  overflow: hidden;
}
.evidence[data-orientation="portrait"] .evidence__media,
.evidence[data-orientation="square"] .evidence__media {
  --slot: 1 / 1;
}
.evidence__media > video,
.evidence__media > img {
  position: absolute;
  inset: 0;
  inline-size: 100%;
  block-size: 100%;
  object-fit: contain;
  object-position: center;
}
@container evidence (min-width: 480px) {
  .evidence--compact {
    grid-template-columns: minmax(0, 3fr) minmax(0, 2fr);
  }
  .evidence--compact .evidence__media { max-block-size: 240px; }
}
@container evidence (min-width: 720px) {
  .evidence { gap: 24px; }
  .evidence--compact {
    grid-template-columns: minmax(0, 1fr) 240px;
  }
  .evidence--hero {
    grid-template-columns: minmax(0, 2fr) minmax(0, 3fr);
    padding: 24px;
  }
  .evidence--hero .evidence__media { max-block-size: 480px; }
  .evidence--hero[data-orientation="portrait"] .evidence__media {
    --slot: 4 / 3;
  }
}
```

Use a real button over the compact poster to open a shared accessible player. Hero playback can remain inline. Provide ordinary video controls, `playsinline`, caption tracks and a transcript action; do not suppress controls until a replacement has successfully initialised.

### Performance and export behaviour

Web.dev recommends `preload="none"` for non-autoplay video and explains that posters occupy the video space while loading. It also cautions against lazy-loading an LCP video; an important above-the-fold poster may need priority loading. Native video lazy-loading is support-dependent, so it should not be the sole loading strategy. [11]

**Proposed implementation:** render posters and text first; attach media sources or initialise the streaming player on intent. Allow **one active player** across the report, pausing the previous one when another starts. Use the existing adaptive-streaming pipeline where appropriate; for short clips, compare a small MP4 rendition against streaming overhead rather than mandating a new player stack. Limit poster requests to near-viewport content, except the principal above-the-fold image.

For PDF, print and non-playing slides, render **poster, quote, attribution, duration, study and a stable permission-checked source link**. Never export a broken play button as the only representation. Condens documents downloadable clips and reels for embedding in presentations; this is a separate delivery mode from a static report. [42]

**Recommendation:** maintain one canonical evidence record and separate web, print and slide renderers. Export permissions must govern both downloadable footage and textual derivatives.

## 9. Give Compass An Evidence Schema, Not A Quote-Detection Rule

### A citation can be valid while its interpretation is wrong

Discuss says each takeaway links to its underlying quote and clip, while its interface can acknowledge thin evidence. Dovetail likewise links generated insights to source material and leaves validation with the team. These are useful patterns, but source-linked AI still needs a check that the linked passage actually supports the statement. [19][23]

NIST recommends reviewing and verifying sources and citations during pre-deployment measurement and ongoing monitoring. For Compass, that means testing claim support, not merely checking that a clip ID resolves or that quoted words occur somewhere in a transcript. [8]

### Proposed generation pipeline

1. **Retrieve authorised evidence:** exclude withdrawn, inaccessible or unapproved derivatives before generation.
2. **Identify canonical source spans:** preserve question, respondent, language and transcript version.
3. **Generate findings with explicit evidence relationships:** label each reference `supports`, `qualifies`, `challenges` or `context`.
4. **Validate literal quotation and timing:** reject altered wording presented as verbatim; preserve negation and qualifications.
5. **Assess claim support:** a matching transcript string is necessary for a quote, but insufficient for the generated interpretation.
6. **Apply report policy:** the renderer, not a free-form model response, chooses placement, responsive behaviour and Hero budgets.
7. **Review and publish:** require human approval for headline evidence and changed translations; preserve an audit trail.

### Structured output example

The identifiers and text below are **schematic placeholders**, not findings from a real study. Compass references approved evidence; it does not invent video URLs, timestamps or attribution.

```json
{
  "schemaVersion": "1.0",
  "claims": [
    {
      "id": "claim_01",
      "text": "Researcher-approved finding text",
      "kind": "interpretation",
      "evidence": [
        {"evidenceId": "ev_01", "relationship": "supports"},
        {"evidenceId": "ev_02", "relationship": "qualifies"}
      ]
    }
  ],
  "blocks": [
    {
      "type": "paragraph",
      "segments": [
        {"type": "claimRef", "claimId": "claim_01"},
        {"type": "evidenceTrigger", "evidenceId": "ev_01"}
      ]
    },
    {
      "type": "evidenceGroup",
      "claimId": "claim_01",
      "evidenceIds": ["ev_01", "ev_02"],
      "placement": "afterParagraph",
      "preferredVariant": "compact"
    }
  ]
}
```

The evidence service resolves each ID to the canonical quotation, source span, metadata and audience-safe rendition. `preferredVariant` is advisory: the policy engine can downgrade a Hero when the budget is exhausted, the clip is unsuitable or the report context is narrative.

**Guardrails:** reject unresolved IDs, source-span mismatches, overlapping duplicates presented as independent respondents, and unsupported segment/date claims. Do not equate "three supporting clips" with "three respondents" or a population percentage. Deduplicate by participant, study and source. Surface sampling limitations and disagreement rather than filling an empty evidence slot with a vaguely related recording.

For translated quotes, retain the original, label the translation, record its review state and keep timing anchored to the original-language source. If evidence is insufficient, render a qualified finding or no evidence card; never manufacture a more persuasive quotation.

**Recommendation:** remove regex-style quote-to-video injection. Use typed document blocks, canonical evidence references and policy-controlled renderers across reports, Share Pages and exports.

## 10. The Global Landscape Is Converging On Traceable Evidence

### Dovetail and Discuss raise the baseline beyond highlight reels

Dovetail describes answers containing a participant's video, exact quote and source session, with access controls that can separate polished findings from raw material. Discuss similarly says takeaways link to quotes and clips and can signal when evidence is thin. This makes inspectability a competitive expectation, not a novel feature on its own. [23][19]

UserTesting exposes a particularly important trade-off: its "Anyone with a link" report setting makes the report and embedded evidence public to the web, whereas Private and Workspace settings constrain access. Frictionless sharing and respondent protection can pull in opposite directions; the new Share Page needs an explicit audience model, not simply a convenient URL. [22]

### Documented competitor workflows

| Player | Video and quotation workflow | Reports, sharing or reels | Provenance and transferable lesson |
|---|---|---|---|
| **Dovetail** | Transcript highlights create searchable clips; source answers can include exact quotes and participant video. [26][23] | Highlights embed in documents and can be shared by link or grouped under tags. [26] | AI-highlight suggestions can be approved or rejected. Separate reusable evidence from layout. [26] |
| **Marvin** | Documents video clips and highlight reels; dedicated quote-card geometry is not specified in the cited material. [25] | Dynamic reports and cross-data answers. [25] | Original-source citations and automatic PII blur are marketed capabilities. Provenance and privacy are already part of the category. [25] |
| **Condens** | Transcript/note highlights become evidence; timestamped transcription, captions and reels support media analysis. [32][29] | Reports, findings and other Artifacts; stakeholder-facing Insights Magazine. [32] | Participant-name visibility controls and permission-respecting published-research access illustrate audience-specific delivery. [32][29] |
| **UserTesting** | Selecting transcript text creates clips; transcript text jumps to corresponding video points. [28] | Workspace clips/highlights can be added to Reports, which support presentation and PDF export. [22] | Explicit visibility settings are critical because public links also expose embedded evidence. [22] |
| **Medallia LivingLens / Medallia Video** | Medallia Video (formerly LivingLens) is a multimedia intelligence platform that provides tools to help companies capture multimedia, search for and navigate to specific moments within media uploads, and create and edit showreels. [20] | Moment navigation and edited showreels are documented; a specific side-by-side quote-card treatment is not established here. | Transfer searchable source moments and curated reels, not an assumed report geometry. |
| **Discuss.io** | Takeaways link to the quote and clip behind them. [19] | Transcript-text selection can generate shareable video quotes; the current platform presents research as a queryable library. [19][31] | "When the evidence is thin, it says so" is a useful model for Compass. [19] |
| **Qualtrics** | Video responses are transcribed and analysed; video editing creates clips. [1] | Clips can be compiled into highlight reels within dashboards. [1] | Transfer the connection between quantitative context and qualitative evidence; a specific card layout is not documented by this source. |
| **Remesh** | Timestamp-based clips, theme clips and manual clips share a Clip Library. [27] | Individual/combined downloads, cited PPTX output and password-protected read-only analysis sharing. [27] | Do not treat Remesh as text-only: its documented video workflow now includes clips and readout assets. [27] |
| **Grain** | Transcript text or an AI note can create a shareable clip. [30] | Playlists, stories and cross-meeting reports with citations. [30] | Useful model for a low-friction transition from summary to original conversation. |
| **Gong** | With Gong recording your calls, you can review conversations and understand what was discussed, what stood out, and what happens next. [51] AI summaries include a recap, key points, and next steps. [51] Click timestamps to go to that part of the call, or click labels to view related snippets. [51] These are call-review patterns, not documented research-report cards. [51] | The cited material concerns call review, not a demonstrated qualitative findings-report card. | Transfer timestamp navigation and summary-to-source inspection, not assumptions about respondent consent. |

These are overlapping competitive arenas: research repositories, experience-management platforms, testing tools and meeting-intelligence products. Their published capabilities show convergence around searchable sources, clips, generated synthesis and reusable outputs; they do not establish relative market shares or audited superiority.

### What is transferable - and what is not

From social/video platforms, borrow adaptive framing, clear play affordances and optional enlargement, not feed-style autoplay. From news, borrow intentional media boundaries and robust player controls: BBC guidance preserves native functionality until enhancement succeeds. From documentation and citation tools, borrow short claim-linked triggers and context inspection. [3][56][43]

The defensible opportunity is a **coherent evidence contract across surfaces**: identical claim/source bindings, visible context, uncropped playback, safe attribution, accessible controls and useful static exports. Competitors' public descriptions do not establish how consistently they satisfy that entire contract; that is a differentiation hypothesis, not proof they lack it.

**Recommendation:** position the redesign around "evidence that remains faithful and verifiable wherever it is shared", rather than "video plus AI" alone.

## 11. Design Failure States As Evidence States, Not Broken Players

### Condens demonstrates that privacy changes the asset, not just the label

Condens documents whole-video or face anonymisation, voice distortion and selective redaction across text, audio and video. This illustrates why an anonymised respondent can require different poster, transcript and playback derivatives, rather than a pseudonym pasted onto the original recording. [29]

Likewise, Cloudinary's automatic crop processing can remain pending while analysis completes. A report should not disappear or become misleading because a derived media treatment is not ready. The quotation and source status need independent rendering paths. [12]

### Proposed failure-mode matrix

| Situation | Required behaviour | Avoid |
|---|---|---|
| Clip shorter than 5 seconds | Default to text plus an optional source action; consider a reviewed contextual excerpt | Looping to make the clip feel longer |
| Clip longer than 60 seconds | Select a focused approved excerpt or open a longer source in the viewer | Making the full response the default evidence card |
| Very long quote | Explicit expansion or a text-led block; preserve qualifications | Silent truncation, tiny type or nested scrolling |
| No visible face / screen recording | Preserve the whole frame; label the source meaningfully | A face-shaped placeholder suggesting missing identity |
| Poor lighting / low resolution | Show an honest poster and enlargement; permit quote-only presentation | Unreviewed enhancement that changes apparent expression |
| Captions missing or inaccurate | Flag the publishing issue; require correction or a genuinely suitable alternative representation | Presenting a transcript excerpt as if it captions the whole clip |
| Non-English source | Label language and translation; expose the original and reviewed translation | Making translated words look like verbatim original speech |
| Anonymisation required | Use approved poster, video, audio and text derivatives | Blurring only the video while captions or audio identify the person |
| Slow connection | Render quote and attribution first; play on request; offer transcript | Empty media space blocking the finding |
| Playback failure | Keep the quote; show Retry, transcript and source availability status | Endless spinner or automatic substitution with another clip |
| Permission denied / withdrawn | Replace the asset with a clear restricted/withdrawn state; re-evaluate claim support | Leaking the poster, cached quote or source details |
| Email preview | Use a safe static summary and permission-checked destination | Assuming embedded video will play in every client |
| Many clips | Use grouped disclosure and a shared player | Multiple simultaneous audio streams or autoplay previews |

The short/long thresholds above are **editorial defaults**, not evidence-validity boundaries. A three-second clip can still be compelling evidence; its meaning, not its duration, determines whether it is publishable.

For Share Pages, distinguish **not yet loaded**, **failed**, **restricted** and **withdrawn**. These states have different implications for reader trust. A failed stream does not invalidate an approved transcript; a withdrawn source may invalidate publication of both.

**Recommendation:** guarantee a useful text-first state, but never let fallback content bypass the same privacy and evidentiary checks applied to video.

## 12. Validate Comprehension And Calibrated Trust Before Optimising Engagement

### HEART makes enterprise task success more important than activity

The original **HEART framework** names Happiness, Engagement, Adoption, Retention and Task Success, and pairs measurement with a **Goals-Signals-Metrics** process. It explicitly notes that engagement may be less meaningful where enterprise users are expected to use a product for work. [58]

For this redesign, a stakeholder who reaches a correct decision quickly and watches one relevant clip may be better served than someone who spends longer scrolling and plays five loosely connected videos. This is the central measurement hypothesis: increase correct verification and appropriately calibrated confidence, not video consumption for its own sake.

### Proposed measurement plan

| Metric | Operational definition | Interpretation |
|---|---|---|
| Claim-to-source accuracy | Correctly matched claim/source tasks divided by attempted tasks | Primary alignment outcome |
| Comprehension | Correct answers to finding, qualification and disagreement questions | Detects persuasion without understanding |
| Verification time | Time from a request to verify a claim to identifying the correct source span | Lower is better only when accuracy is preserved |
| Evidence play rate | Unique evidence plays divided by eligible card exposures, defined as at least 50% visible for 1 second | Diagnostic; not a north-star metric |
| Clip completion | Plays reaching 90% of the excerpt divided by started plays | Segment by clip length, device and variant |
| Source inspection | Context/transcript/source opens divided by evidence interactions | Indicates inspectability, but may also reveal confusion |
| Trust calibration | Confidence ratings compared with correctness and ability to identify weak evidence | Prefer justified confidence over uniformly higher trust |
| Report behaviour | Active reading time, scroll depth and return visits | Interpret jointly with task success |
| Sharing | Share actions and authorised recipient opens | Separate distribution from useful, safe consumption |
| Technical quality | Playback failures, time to first frame, layout shift and caption availability | Segment by network, language, orientation and surface |

All exposure and completion thresholds are **proposed instrumentation definitions**. Log pseudonymous evidence and claim IDs, not respondent quotations or unnecessary personal data.

### Test the real trade-offs

Run an initial **18-participant formative study**, proposed as six insight/research practitioners, six product or design stakeholders and six senior decision-makers. This is a coverage plan, not a statistically powered estimate. Include mobile Share Page use, keyboard-only operation, screen-reader use and poor-network playback; expand recruitment where accessibility or language coverage remains thin.

Compare: current inline-injected evidence; proposed paragraph citation plus viewer; compact side-by-side cards; and a text-led report with selectively placed Heroes. Hold findings and source material constant. Use matched tasks and counterbalance presentation order; NN/g explains that within-subject designs can improve comparability but create learning effects, while between-subject designs reduce carryover at the cost of more participants. [44]

Ask participants to: identify which claim a clip supports; find the exact quoted words; discover a qualification; compare two respondents; explain whether examples establish prevalence; verify a translation; recover from a failed load; and share the report without exposing restricted evidence.

Capture "I cannot tell why this is here", "I need the question they answered", and "I trust the quote but not that conclusion" as distinct failure signals. Ask enterprise clients about review responsibility, evidence reuse, approval bottlenecks, consent boundaries and whether the report changes a real decision.

**Proposed acceptance gates:** at least 90% correct claim-source matches in the evaluated task set; no critical privacy or keyboard blockers; no loss of comprehension against text-led presentation. Then run a powered comparative study or controlled rollout, with sample size based on observed baseline and the minimum worthwhile effect.

**Recommendation:** only optimise play rate after alignment, accessibility, privacy and comprehension pass their gates.

## Synthesis

The redesign is not primarily a choice between two card shapes. It is a choice between **evidence as decorative media** and **evidence as an inspectable source relationship**. The card should express that relationship consistently across reading, verification, sharing and export.

| Competitive approach | Mechanism and scope | Main trade-off | Strategic implication for Voxpopme |
|---|---|---|---|
| Dovetail, Marvin and Condens | Reusable research knowledge, source-linked insights and stakeholder-facing outputs. [23][25][32] | Broad reuse requires careful separation of raw and approved material | Preserve canonical evidence IDs and audience-specific views |
| UserTesting | Observed experience flows into report evidence and presentation. [28][22] | Convenient public sharing can expand exposure of embedded evidence | Make Share Page audience and export scope explicit |
| Qualtrics and Medallia Video | Response analysis, searchable media and curated highlight/reel delivery. [1] Medallia Video (formerly LivingLens) is a multimedia intelligence platform that provides tools to help companies capture multimedia, search for and navigate to specific moments within media uploads, and create and edit showreels. [20] | A compelling montage is a different artefact from verification of one claim | Keep reels separate from claim-level evidence cards |
| Discuss and Remesh | Generated research analysis with source references and shareable outputs. [19][27] | Fast synthesis still needs support checks and acknowledgement of thin evidence | Bind generated claims to reviewed spans and visible limitations |
| Grain and Gong | Conversation summaries lead back to original moments. [30] With Gong recording your calls, you can review conversations and understand what was discussed, what stood out, and what happens next. [51] AI summaries include a recap, key points, and next steps. [51] Click timestamps to go to that part of the call, or click labels to view related snippets. [51] These are call-review patterns, not documented research-report cards. [51] | Operational call review is not the same consent or reporting context as respondent research | Borrow source navigation, not the entire sharing model |

Across these approaches, **scope**, **verification mechanism** and **sharing risk** differ more than the mere presence of video. For Voxpopme, the near-term advantage is not another montage format; it is a dependable path from finding to exact words, surrounding context and permitted source.

Three tensions should shape implementation. First, containment preserves all evidence but can make portrait footage small; orientation-aware slots and enlargement resolve that better than unreviewed cropping. Second, proximity helps association but forced inline playback harms narrative control; explicit claim-linked disclosure reconciles the two. Third, stronger emotional impact is not stronger evidentiary support; visible disagreement, context and sample information are essential counterweights.

**Delivery sequence:** first fix containment, metadata, quote-only references and accessible shared playback. Next introduce compact/Hero policy and typed Compass blocks. Then add context inspection, review feedback and static exports. Defer smart cropping until evidence shows a material benefit after review and operating costs are included.

**Final recommendation:** approve the side-by-side direction, but do not approve a two-card-only system. The complete pattern is **Hero for emphasis, compact for discrete support, attributed text for prose, and grouped disclosure for scale**, all backed by the same governed evidence model.

## References

- **Voxpopme:** Compass product documentation and its stated timing/sample-size validation. [13]
- **Video framing and posters:** YouTube aspect-ratio guidance; MDN `object-fit`; Cloudinary video gravity and poster documentation; MediaPipe AutoFlip documentation. [3][6][12][34][10]
- **Reading and multimedia learning:** Schroeder and Cenkci, spatial contiguity/split-attention meta-analysis; Sundararajan and Adesope, seductive-details meta-analysis. [48][9]
- **Accessibility:** W3C WCAG 2.2; W3C G203; target-size guidance; BBC GEL video controls. [14][36][37][56]
- **Legal scope:** Directive (EU) 2019/882; European Commission Web Accessibility Directive guidance; GOV.UK accessibility guidance; Section508.gov; US Department of Justice web guidance. [40][55][57][35] This guidance addresses website accessibility for state and local governments and businesses open to the public under the ADA. [47] [47] [47] It states that covered entities have flexibility in how they comply, and DOJ does not set detailed technical standards; WCAG is helpful guidance, not a blanket WCAG 2.2 mandate for all private B2B software. [47] [47]
- **Research ethics:** MRS observer guidance and permitted recording use. [2]
- **Browser implementation:** MDN container queries, `aspect-ratio`, `videoWidth` and `loadedmetadata`; web.dev video-loading guidance. [38][33][54][53][11]
- **AI risk and citation inspection:** NIST AI 600-1 and Google source-citation help. [8][43]
- **Research repositories:** Dovetail highlights/repository documentation; Marvin product documentation; Condens getting-started, AI and export documentation. [26][23][25][32][29][42]
- **Research and experience platforms:** UserTesting transcripts/Reports; Discuss product documentation; Qualtrics qualitative/quantitative research documentation; Remesh product update. [28][22][19][1][27]
- **Conversation intelligence:** Grain product documentation; Gong call-review documentation. [30] With Gong recording your calls, you can review conversations and understand what was discussed, what stood out, and what happens next. [51] AI summaries include a recap, key points, and next steps. [51] Click timestamps to go to that part of the call, or click labels to view related snippets. [51] These are call-review patterns, not documented research-report cards. [51]
- **Measurement and testing:** Rodden, Hutchinson and Fu, HEART/Goals-Signals-Metrics; NN/g, between- versus within-subjects study design. [58][44]

## Final Decision Table

| Content situation | Recommended card type | Reason |
|---|---|---|
| Section-defining evidence where voice or visible behaviour matters | Hero, within the section/report budget | Creates meaningful emphasis without making every finding equally prominent |
| One discrete supporting clip beneath a finding | Compact side-by-side card | Keeps quote and video visibly associated at manageable density |
| A quotation inside running prose | Attributed quote plus "Watch clip" evidence link | Preserves reading flow and explicit claim support |
| Several sources supporting one finding | Evidence group with one representative item and expandable list | Scales without a wall of players |
| Respondents disagree or qualify a finding | Labelled comparative evidence group | Makes the tension visible rather than hiding it in a reel |
| Portrait clip in a landscape report | Orientation-aware slot with contained video | Preserves face and surrounding context without a full-height card |
| Very long quotation or long source response | Text-led block plus expandable source viewer | Avoids excessive height and nested scrolling |
| Short clip whose meaning depends on context | Attributed text plus a reviewed contextual excerpt | Protects meaning rather than optimising duration |
| Poor-quality or non-visual footage | Quote-only evidence with optional source playback | Keeps usable evidence without pretending the picture adds value |
| Anonymisation required | Approved redacted card, or safe quote-only alternative | Applies privacy protection to poster, audio, captions and text together |
| AI-generated finding | Typed claim reference plus canonical evidence block/trigger | Prevents arbitrary quote-detection injection and misleading pairings |
| Mobile executive Share Page | Quote-first compact/stacked treatment; at most a deliberate Hero | Prioritises comprehension, optional playback and controlled page height |
| PDF, print or non-playing slides | Static evidence card with poster, quote and governed source link | Remains useful when playback is unavailable |
| Missing permission, withdrawn source or unsupported claim | Restricted/withdrawn state, qualified text, or no card | Never substitutes persuasive-looking but invalid evidence |

## References

1. *Qualitative + Quantitative Research & Analytics Software*. https://www.qualtrics.com/en-au/strategy/research/qualitative-quantitative
2. *MRS Guide to Observers' Legal & Ethical Responsibilities (For those viewing or listening to research) & Viewing Facilities - Commitments*. https://www.mrs.org.uk/pdf/Guide%20to%20Observers%20Legal%20%20Ethical%20Responsibilities%20%28with%20Committments%20to....pdf
3. *Video resolution & aspect ratios - Computer - YouTube Help*. https://support.google.com/youtube/answer/6375112?co=GENIE.Platform%3DDesktop&hl=en
4. *Video resolution and aspect ratios - Computer - YouTube Help*. https://support.google.com/youtube/answer/6375112?co=GENIE.Platform%3DDesktop&hl=en-GB
5. [
        A meta-analysis of Richard Mayer's multimedia learning research: Searching for boundary conditions of design principles across multiple media types
      -  Illinois Experts](https://experts.illinois.edu/en/publications/a-meta-analysis-of-richard-mayers-multimedia-learning-research-se/)
6. *object-fit CSS property - CSS | MDN*. https://developer.mozilla.org/en-US/docs/Web/CSS/Reference/Properties/object-fit
7. *European Accessibility Act (EAA) - European Commission*. https://commission.europa.eu/strategy-and-policy/policies/justice-and-fundamental-rights/disability/european-accessibility-act-eaa_en?prefLang=cs
8. *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile*. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
9. *ERIC - EJ1263249 - Keep It Coherent: A Meta-Analysis of the Seductive Details Effect, Educational Psychology Review, 2020-Sep*. https://eric.ed.gov/?id=EJ1263249
10. *mediapipe/docs/solutions/autoflip.md at master · google-ai-edge/mediapipe · GitHub*. https://github.com/google/mediapipe/blob/master/docs/solutions/autoflip.md
11. *Lazy loading video  |  Articles  |  web.dev*. https://web.dev/articles/lazy-loading-video
12. *Video gravity for crops*. https://cloudinary.com/documentation/video_gravity
13. *AI Research Agent | Voxpopme Compass*. https://www.voxpopme.com/research-agent
14. *Web Content Accessibility Guidelines (WCAG) 2.2*. https://www.w3.org/TR/WCAG22/
15. *Content*. https://help.gong.io/docs/content
16. *AI Briefs use stats | Community*. https://visioneers.gong.io/data-in-gong-71/ai-briefs-use-stats-1527
17. *Who or what is Marvin? | Marvin's Help Center*. https://help.heymarvin.com/en/articles/10130484-who-or-what-is-marvin
18. *Medallia LivingLens April 2022 (22LCR3) | Medallia Documentation*. https://docs.medallia.com/en/medallia-video/video-release-notes/medallia-livinglens-april-2022-22lcr3
19. *Qualitative Research Data Analysis Software for Market Insights*. https://www.discuss.io/
20. *Medallia Video | Medallia Documentation*. https://docs.medallia.com/en/medallia-video/medallia-video
21. *Grain | The AI Notetaker Built for Growing Teams*. https://grain.com/
22. *Insights Reports – Knowledge Base Home*. https://help.usertesting.com/hc/en-us/articles/23113393692061-Insights-Reports
23. *Research Repository*. http://dovetail.com/solutions/research-repository
24. *Remy Live: Interactive Analysis Demo*. https://help.remesh.ai/e3t/Ctc/S+50013/cw1Jk04/Jks2-6qcW69sMD-6lZ3pyW7Ypb046lWbKLW40JtD-113_F3Mp4bR7xG3wwW4Wp39h8yh22DW1_J7f05f1ZS2W71Mz--8nDVNbW5hhlW_2KrLwGW58F9zL2-kW3bW2yXk_47xQRn7W6b0p3N381P4SW8dkCs32NQg_ZW5dYjyD7zGMMcW2ctMPS2Qyr5vW7z3fsg3bzYzrN5hvF_fvJWqJVppHMb4WtVYjN6Zk96Lmp0b1VPj35n7sLddXW2zZlD63VYsPRW4DSj065yy0kbf10kwv004
25. *Marvin | The #1 AI-native customer insights platform*. https://heymarvin.com/
26. *Highlights*. https://docs.dovetail.com/help/projects/highlights
27. *Remesh 2026: Remy, AI Research & New Platform Features*. https://www.remesh.ai/resources/from-question-to-defensible-decision-2026-at-remesh-so-far
28. *Transcripts (UserTesting) – Knowledge Base Home*. https://help.usertesting.com/hc/en-us/articles/11880413169821-Transcripts-UserTesting
29. *How Condens AI Supports Every Step of Your Research Process*. https://condens.io/product-updates/condens-ai-supports-research-process/
30. *Grain | The AI Notetaker Built for Growing Teams*. https://www.grain.com/product
31. *Feature Series: How To Analyze Focus Group Data*. https://www.discuss.io/blog/feature-series-finding-sharing-quotes/
32. *Getting started*. https://condens.io/help/-/getting-started
33. *aspect-ratio CSS property - CSS | MDN*. https://developer.mozilla.org/en-US/docs/Web/CSS/Reference/Properties/aspect-ratio
34. *Video previews and posters*. https://cloudinary.com/documentation/video_previews_and_posters
35. *IT Accessibility Laws and Policies | Section508.gov*. https://www.section508.gov/manage/laws-and-policies/
36. *G203: Using a static text alternative to describe a talking head video | WAI | W3C*. https://www.w3.org/WAI/WCAG21/Techniques/general/G203
37. *Understanding Success Criterion 2.5.8: Target Size (Minimum) | WAI | W3C*. https://www.w3.org/WAI/WCAG22/Understanding/target-size-minimum.html
38. *Using container size and style queries - CSS | MDN*. https://developer.mozilla.org/en-US/docs/Web/CSS/Guides/Containment/Container_size_and_style_queries
39. *Anonymising data and protecting participants - UK Data Service*. https://ukdataservice.ac.uk/learning-hub/data-producer-support/preparing-data-for-sharing-and-reuse/anonymising-data-and-protecting-participants
40. *Directive - 2019/882 - EN - EUR-Lex*. https://eur-lex.europa.eu/eli/dir/2019/882/oj/eng
41. *CSS container queries - CSS | MDN*. https://developer.mozilla.org/en-US/docs/Web/CSS/CSS_Container_Queries
42. *Download files and highlight reels*. https://condens.io/help/import-and-export/export/download-files-and-highlight-reels
43. *Use chat in Gemini Notebook - Gemini Notebook Help*. https://support.google.com/notebooklm/answer/16179559
44. *Between-Subjects vs. Within-Subjects Study Design - NN/G*. https://www.nngroup.com/articles/between-within-subjects/
45. *ERIC - EJ747512 - Integrating Information: A Meta-Analysis of the Spatial Contiguity and Temporal Contiguity Effects, Learning and Instruction, 2006-Dec*. https://eric.ed.gov/?id=EJ747512
46. *AutoFlip: An Open Source Framework for Intelligent Video Reframing*. https://research.google/blog/autoflip-an-open-source-framework-for-intelligent-video-reframing/
47. *Guidance on Web Accessibility and the ADA | ADA.gov*. https://www.ada.gov/resources/web-guidance/
48. *ERIC - EJ1186641 - Spatial Contiguity and Spatial Split-Attention Effects in Multimedia Learning Environments: A Meta-Analysis, Educational Psychology Review, 2018-Sep*. https://eric.ed.gov/?id=EJ1186641
49. *View a call transcript*. https://help.gong.io/docs/view-a-call-transcript
50. *Intro to the call page*. https://help.gong.io/docs/intro-to-the-call-page
51. *Review what happened in a call*. https://help.gong.io/docs/review-what-happened-in-a-call
52. *Tufte CSS*. https://edwardtufte.github.io/tufte-css
53. *HTMLMediaElement: loadedmetadata event - Web APIs | MDN*. https://developer.mozilla.org/en-US/docs/Web/API/HTMLMediaElement/loadedmetadata_event
54. *HTMLVideoElement: videoWidth property - Web APIs | MDN*. https://developer.mozilla.org/en-US/docs/Web/API/HTMLVideoElement/videoWidth
55. *Web accessibility | Shaping Europe’s digital future*. https://digital-strategy.ec.europa.eu/en/policies/web-accessibility
56. *GEL | Video controls*. https://bbc.github.io/gel/components/video-controls/
57. *Understanding accessibility requirements for public sector bodies - GOV.UK*. https://www.gov.uk/guidance/accessibility-requirements-for-public-sector-websites-and-apps
58. *heart-metrics-author-version*. https://research.google.com/pubs/archive/36299.pdf