Independent AI Research, Intelligence & TestingInstruction following
NOT TESTED- Exact slide-count control
- Required headings and order
- Required content placement
- Formatting constraints
- Conflicting instructions
AI PRESENTATIONS · OCTOBER 2026 · REPORT ISL-20261006-001
HANDS-ON PILOT · COMPLETED
SlidesGPT
SlidesGPT
Presentation generation test
Our first AI presentation generator test: a six-slide business brief, supplied financial facts, and ten checks fixed before generation.
SINGLE-RUN TASK ADHERENCE50 / 1005 of 10 criteria fully met
First pilot, not a certification. This score measures adherence to one fixed presentation task. It does not measure overall product quality, consistency, security, or standing against competitors.
EXTENSIVE TEST PROGRAM · DRAFT
60 checks across 12 testing areas
Our broader evaluation will test the product with controlled tasks, repeated runs and inspectable output evidence. The ten-check pilot above is a separate first exercise; it does not complete this benchmark.
60 planned checks0 benchmark checks completedOverall product rating: not issued
Factual accuracy
NOT TESTED- Source-fact preservation
- Unsupported claims
- Missing-information handling
- Citation and link validity
- PDF extraction accuracy
Calculations and data integrity
NOT TESTED- Growth calculations
- Totals and percentages
- Money and unit consistency
- Chart-to-table agreement
- Targets versus actual results
Writing and communication
NOT TESTED- Audience adaptation
- Narrative flow
- Conciseness and readability
- Technical explanation
- Multilingual fidelity
Visual design and layout
NOT TESTED- Overflow and clipping
- Typography and hierarchy
- Theme consistency
- Brand adherence
- Image relevance
Charts, tables and diagrams
NOT TESTED- Table completeness
- Chart suitability
- Axes and labels
- Dense data readability
- Diagram relationship accuracy
Editing and user control
NOT TESTED- Outline editing
- Slide-level revision
- Regeneration stability
- Undo and recovery
- Save and reopen
Exports and compatibility
NOT TESTED- PowerPoint fidelity
- PDF fidelity
- Google Slides fidelity
- Font and media portability
- Notes and link preservation
Speed and repeatability
NOT TESTED- Generation time
- Fresh-run consistency
- Failure and retry behavior
- Long-brief handling
- Version-aware retesting
Usability and workflow
NOT TESTED- First-run completion
- Settings discoverability
- Mobile workflow
- Plan-limit clarity
- Support response
Accessibility
NOT TESTED- Keyboard navigation
- Screen-reader labels
- Text contrast
- Reading order and alternatives
- Zoom and small-screen legibility
Privacy controls and documentation
NOT TESTED- Disclosure consistency
- Sharing-default visibility
- Deletion-control behavior
- Retention and training options
- Sensitive-data guidance
First hands-on pilot · 10 checks
Structure & instructions
0 / 30Extra title and agenda slides shifted the requested order and table position.
Facts & arithmetic
50 / 50Supplied price, customers, MRR, growth and March target were preserved.
Presentation caveats
0 / 20The remainder was in speaker notes; required limitations were on a later slide.
What happened
The unedited outline matched all six requested headings. The rendered deck added a cover and an agenda, producing eight rendered slides: a cover, an agenda, and six numbered content slides. Our strict count includes all rendered slides. The comparison table and financial figures were retained, but strict placement requirements were missed.
Practical takeaway: A useful draft for this fictional brief, with an instruction-checking and editing step needed before delivery.
Read the full testing review →Detailed test results
First output · October 6, 2026 · 10 points per fully met criterion. Competitor and industry averages are not available for this pilot.
Structure & Instructions
Slide count, sequence and placement against the original task.
Method & more information →| Test case & observed evidence | Maximum | SlidesGPT |
|---|---|---|
| C01 · CONTROLLED TASKExactly six slides Eight rendered slides: cover, agenda and six numbered content slides. | 10points | 0NOT MET |
| C02 · CONTROLLED TASKRequired titles in exact order Rendered order begins Cedar Analytics, Agenda, Cedar Analytics; exact six-title sequence not met. | 10points | 0NOT MET |
| C07 · CONTROLLED TASKSlide 3 contains requested comparison table Requested table is on rendered slide 5, not rendered slide 3. | 10points | 0NOT MET |
Structure & Instructions score0 / 30
Facts & Arithmetic
Preservation of the supplied dataset and calculated values.
Method & more information →| Test case & observed evidence | Maximum | SlidesGPT |
|---|---|---|
| C03 · CONTROLLED TASK$40 monthly price preserved $40 per month appears on rendered slide 4. | 10points | 10MET |
| C04 · CONTROLLED TASKJanuary 120 and February 150 customers preserved 120 January customers and 150 February customers appear on rendered slide 5. | 10points | 10MET |
| C05 · CONTROLLED TASKJanuary $4,800 and February $6,000 MRR preserved $4,800 January MRR and $6,000 February MRR appear on rendered slide 5. | 10points | 10MET |
| C06 · CONTROLLED TASKCustomer and MRR growth both 25% Customer and MRR rows each show 25% growth on rendered slide 5. | 10points | 10MET |
| C09 · CONTROLLED TASK180 March customers explicitly a target 180 customers explicitly labeled target, not achieved, on rendered slide 7. | 10points | 10MET |
Facts & Arithmetic score50 / 50
Presentation Caveats
Visibility and placement of financial labels and limitations.
Method & more information →| Test case & observed evidence | Maximum | SlidesGPT |
|---|---|---|
| C08 · CONTROLLED TASK$3,200 costs and $2,800 remainder correctly labeled; not net profit $3,200 costs appear in the chart on rendered slide 6; $2,800 and its label appear only in speaker notes. | 10points | 0NOT MET |
| C10 · CONTROLLED TASKRequired limitations on slide 6; no invented facts or sources Required limitations appear on rendered slide 8, not rendered slide 6; no invented sources observed. | 10points | 0NOT MET |
Presentation Caveats score0 / 20
Conditions and evidence
Browser workflow: prompt → unedited outline → Show Slides. No regenerated output, manual text edits, model selection, subscription purchase or export test. Latency was not instrumented, so no speed score is assigned. No matched competitor run exists; an industry average is unavailable.