What makes a ‘good’ corporate presentation?

Over the past 18 months we've worked with hundreds of companies and generated thousands of presentations, not to collect human-written slides but to understand what it takes for a model to produce slides that look like a given company's. What surprised us in the process was how specific to a company, and to a team inside it, the definition of a good-looking presentation turned out to be. This benchmark comes out of that work.

Building a presentation is a multi-stage process: storylining, content creation and analysis, then visualisation. Doppel Benchmark covers the last stage, for two reasons. One model is unlikely to be best at all three, so the stages are worth measuring separately. And visualisation is what users see first: a deck with the right argument in the wrong visual language gets rejected before the argument is read.

Any company with a defined identity on its slides, whether enterprise or mid-market, wants generated output to look like theirs. That is hard for two reasons. A company's slide identity is far richer than its brand guideline and its master layout. And the parts that aren't in the guideline differ across teams, topics, clients, and managers, and are difficult for anyone to state out loud.

Almost every one of these companies has a design team that sets the initial track, usually through a slide master and a set of published rules. Very few of those guardrails survive contact with actual daily work. What emerges instead are conventions attached to a situation: how this team builds a QBR, how financial reporting differs from strategy work, what this manager expects, what this client has come to expect. These conventions often diverge from what the design team specified. That isn't a failure of the design team, but rather that when a team is building an actual artifact, they diverge from a rigid design system; sometimes for strategic reasons, sometimes out of preference, sometimes for reasons no one remembers.

Capturing that requires two axes. One is the brand guideline - typically published, stable, and company-wide. The other is situational preference, which is emergent, local, and mostly unwritten. We call the combination a company's visual intelligence system. A generator that reads only the first axis produces slides that are on-brand, rigid, and still not what the team would have made.

In practice it looks as follows. The brand guideline tells you what a chart looks like: colours, fonts, axis treatment. Then a manager tells you that for revenue by channel, they want the three charts stacked one above the other. On a different topic, revenue against customer churn, the same manager wants them side by side. Neither preference is written down anywhere, both are firm, and a generator that only reads the guideline gets both wrong.

Doppel evaluates whether models can capture a company's visual intelligence system from its own material and hold to it when generating something new.

Current state of slide benchmarking - what's missing

Several benchmarks for AI slidemaking have been published in the past year. They measure real things, but none answers the question an enterprise actually asks, which is whether a generated deck looks like the company's own. Three reasons that question needs a different kind of benchmark.

Past the basics of design, visual quality can only be judged against a reference. Some properties are objectively measurable: alignment, overlap, elements running off the slide, contrast, whitespace balance. But measuring them and judging them are different steps, and the thresholds are company-specific. Text over an image is deliberate in one house style and a defect in another. Density that reads as cramped for a consumer brand is standard in financial reporting. The universal layer is also the one frontier models have largely cleared, so scoring decks in absolute terms, or by broad category such as strategy or finance, has little separating power left. Categories capture density and slide type, but not identity: two teams inside the same firm will disagree on what a finished slide looks like. This sets a requirement on the source material too. If the reference presentations carry little visual identity, or all carry the same one, a model that reproduces a company's conventions and one that ignores them score about the same, because there was little to miss.

Visual conventions have to be extracted, not prompted. Frameworks that ask users to state their guidelines assume people can articulate their own visual identity. In practice they can't, or won't spend the time, and what does get written down is generic enough to be true of any deck. The conventions live in past assets, and most were never written down anywhere, so extraction is part of the benchmark rather than a precondition for it.

A single blended score is no longer actionable. Storylining, content, and visualisation are increasingly handled by different models at different stages of a pipeline, and specialisation is only possible if the evaluation resolves to the parameter level. A team needs to know which visual conventions the output deviated on, not one number averaging layout with writing.

Doppel closes those three gaps: it scores a generated presentation against a specific company's reference material, extracts the conventions rather than asking for them, and reports where the deviation sits.

Overview

To compare the performance of different models on a single data point, we follow these steps:

1. Start with a real corporate presentation - i.e. a presentation created by a human expert in a well-known company.
2. Extract the company's visual intelligence system from this presentation
3. Give to each model the reference presentation as a PDF, and a prompt (related to the reference presentation) to generate a new PPTX presentation
4. Extract the visual intelligence system from the generated presentation
5. Compare the visual intelligence system of the generated presentation to the visual intelligence system of the reference presentation
6. Score the generated presentation based on the comparison

For each model, we compute a score for every reference-presentation-prompt-generated-presentation tuple using the aggregation method described below, then take an equally weighted average across those tuple-level scores. This gives us the model's final percentage score.

The pilot results reported here are based on three reference-presentation-prompt tuples, with three independent generations per model for each tuple. We plan to expand the evaluation to more reference-presentation-prompt tuples from the dataset, so these preliminary scores are likely to change, and the model rankings may change.

Focus on visual design

Defining what goes into making a 'good' presentation is fairly nebulous. We attempt to define it more clearly by splitting it into a three-stage process:

• Storylining: The presentation's overarching narrative: how ideas are selected, structured, sequenced, and connected to build a coherent argument for its intended audience and objective.
• Content and Analysis: The quality of the information and reasoning within that narrative, including its factual accuracy, relevance, analytical rigour, supporting evidence, insights, and recommendations.
• Visual design: The presentation's diagrams, frameworks, composition, consistency, reuse patterns, and brand guidelines, including its layouts, typography, colour, imagery, charts, and other visual conventions.

For this benchmark, we restrict our focus to the Visual design stage, as well as some selected writing conventions.

Reference dataset

To run this benchmark, we've built Doppel Benchmark as a set of 25 corporate PowerPoint presentations from real enterprise and mid-market companies, with the identifiable information anonymized, validated by business leaders and management consultants. The median presentation length is approximately 40 slides. Only three of these 25 presentations are included in the pilot results reported here. We are currently working on expanding this dataset to 1,000 presentations with:

• Additional real corporate presentations - i.e. presentations created by business leaders and presented to a professional audience - from enterprise and mid-market companies
• Presentations generated specifically for the benchmark by human experts - typically ex enterprise business leaders, ex bankers, or ex management consultants.

Defining a presentation's visual intelligence system

A naive way to evaluate how well a produced presentation replicates a source presentation would be to run it through a VLM and ask it to provide a score. However, this approach would be:

• Slow
• Costly
• Highly variable
• Difficult to reproduce
• Difficult to break down and understand the different components of the score

We therefore define a visual intelligence system for presentations. This addresses the issues above and, most importantly, allows us to be more deterministic in our checks and to isolate the criteria where a model is measurably deviating from the company's branding and writing guidelines under our rubric.

A presentation's visual intelligence system is formed of a set of individual criteria across key categories. We do not show the full list of categories here, but they include:

• Structure
• Spacing
• Positioning
• Colour
• Typography

The full list of criteria is kept closed-source for the time being. Some examples of these criteria:

• The hex color of the slide title in content slides (Typography)
• The EMU or cm spacing between the slide content and the left edge of the slide in header slides (Spacing)
• The vertical alignment of section headers in agenda slides (Positioning)

The visual intelligence system schema represents criteria using five broad representation families:

• Point values (e.g. a hex colour)
• Range values (e.g. the range and median of acceptable pt font sizes of a paragraph header)
• Bucketed values (e.g. very low, low, medium, high, very high density of content on a slide)
• Conditional values (e.g. if a bar chart is present on a slide, it must have a title above it)
• Free-form values (e.g. the writing style of the slide title)

[Note: these are high-level authoring families rather than scoring rules. During comparison, they are mapped to more specific forms such as exact values, ranges, floors and ceilings, categorical values, colours, presence checks, bands, mixtures, count ranges, conditional rules, and free-form prose.]

Defining these criteria and their representation families allowed us to build a structured visual intelligence system schema, which can then be used to more systematically compare the visual intelligence systems of two presentations.

Extracting the visual intelligence system

As described above, Doppel's methodology depends on extracting the visual intelligence system from a reference presentation. To do this, we have built an extractor workflow. While some parts of it are deterministic (done with scripts), other parts are done with an LLM, and the whole workflow is orchestrated with an agent (which we henceforth refer to as the "extractor agent").

To ensure the extractor agent is able to extract the visual intelligence system correctly, it was crucial that we ground it in a set of worked examples. We did this as follows:

1. Populating a visual intelligence system schema for 5 reference presentations manually, by human experts with at least 10 years of experience in corporate slidemaking.
2. Building an extractor skill from these 5 manually authored reference profiles. The skill combines written instructions, visual verification procedures, and Python measurement and validation scripts.
3. Running the skill on the same 5 presentations, comparing its outputs with the manually authored profiles, and iterating until it could reproduce them reliably.
4. Applying the skill to 10 additional presentations and manually reviewing the resulting profiles. We supplied identified errors and omissions to a coding agent, which proposed changes to the skill; we then manually reviewed those changes before incorporating them and rerunning the extractor.

Target presentation generation

For each reference presentation, we assign a human-provided prompt based on a real-world scenario that asks a model to generate a brief presentation - typically 5-10 slides - around that scenario. We generate that target presentation by:

• Using the web interface, on a premium subscription account, of the model we are evaluating (e.g. claude.ai, chatgpt.com, etc.)
  • Note this means we are evaluating the frontier lab's product rather than the model itself, which is closer to the user experience. For simplicity, we are referring to "the product set with a given model and reasoning effort level" as "the model".
• Selecting the model and reasoning effort level, as allowed by the model's product interface
• Starting a fresh 0-context conversation
• Providing the prompt, and the PDF of the reference presentation attached as reference material
  • We ask the model to generate a presentation that is as close as possible to the reference presentation in terms of visual design and style
  • We ask the model to generate a downloadable PPTX file of the presentation
  • We retain the first output
• If PowerPoint reports that the generated file needs repair, we allow its standard one-click Repair operation before evaluation. When that operation opens the presentation successfully without further intervention or visible content loss, we record the repair but do not penalise the model. Repairs that require manual reconstruction, fail to produce an openable presentation, or visibly alter or remove content are not treated as trivial.

Hybrid evaluation pipeline

Doppel Benchmark does not rely on a single LLM looking at two presentations and returning a holistic similarity score. The evaluator is deliberately hybrid, with different parts of the visual intelligence system handled by the method best suited to them:

• Deterministic measurement. Scripts analyse a PDF rendering of each presentation and measure properties such as layout geometry, fixed-element positioning, typography, spacing and density. Measurements are retained as evidence alongside the inferred convention rather than being reduced immediately to a score.
• Structured agent interpretation. A model classifies slide layout templates, assigns semantic roles to elements, identifies components, and authors the parts of the profile that cannot be recovered reliably from geometry alone, including writing style, palette structure, iconography, and chart conventions.
• Visual and structural verification. The pipeline renders labelled overlays of the extracted elements and components for visual inspection, checks that the measurements correspond to what is visible on each slide, validates the profile against the schema, and verifies that quoted writing-style exemplars appear in the source presentation. Errors found at this stage are corrected before scoring.
• Parameter-level comparison. Once the reference and generated presentations have structured visual intelligence system profiles, numerical and categorical criteria are compared using fixed rules wherever possible. Criteria requiring semantic judgement, such as similarity of title-writing style, are evaluated by an LLM judge rather than forced into an unreliable numerical proxy.

The resulting benchmark is therefore not fully deterministic. Most measurement and comparison steps are reproducible, but variation can enter through layout classification, semantic labelling, the authored parts of the visual intelligence system profile and the LLM-judged criteria. The judge is blinded to the identity of the product that generated each target presentation. We have measured how much that variation is worth in points, and report it under Evaluator reliability below.

We have found best results using Anthropic's Opus 5 model with high reasoning effort, run via Claude Code CLI, for both running the extractor agent and the LLM judge.

Scoring - comparing the visual intelligence systems of the generated and reference presentations

Some of the criteria are deterministic - e.g. "the value and standard deviation of the X position of the title for content slides". For these, we can deterministically compare the value in the visual intelligence system of the generated presentation to the value in the visual intelligence system of the reference presentation. We have tried to keep as many of the criteria deterministic as possible. However, some criteria need to be scored with an LLM judge - e.g. writing style of the slide titles.

Each criterion is scored 0, 1, or 2, where 0 is a failure, 1 is a partial success, and 2 is a full success. Criteria that do not apply - a layout template the generated presentation never uses, an element the reference presentation itself does not have - are excluded rather than scored 0, and we report how many were excluded alongside the score.

Criteria are then rolled up in two stages.

1. Within each category, we average the applicable criterion scores. A criterion associated with a slide layout template is weighted by that template's share of slides in the reference presentation, while a presentation-wide criterion receives a weight of 1.
   • For example, in a 40-slide reference presentation with 30 content slides and 4 header slides, a criterion applying to content slides receives a weight of 30/40, or 0.75; one applying to header slides receives a weight of 4/40, or 0.10; and a presentation-wide criterion receives a weight of 1 because it governs the whole presentation. Criteria that do not apply are excluded.
2. The resulting category scores are averaged equally to produce a final 0-100% score. Categories with no applicable scored criteria are excluded. Equal weighting at this stage is deliberate: it prevents categories that naturally contain more criteria from dominating the result. For example, typography contributes dozens of criteria while narrative structure contributes only one, so averaging all criteria flat would make this largely a typography benchmark.

We report the per-category scores alongside the headline number because they are what a team can act on. For each model, we then average the final scores across all reference-presentation-prompt-generated-presentation tuples, weighted equally. This gives us the model's final percentage score.

Evaluator reliability

Because the evaluator is hybrid rather than fully deterministic, we measured how much of a score it contributes before publishing the scores themselves. Holding 24 generated presentations fixed, we replicated both evaluator stages: judging three times over, and extraction three times with a fresh judging on each.

Against the evaluator, the rankings are more robust than the scores. 69 of the 84 model pairs among the eight products in this study separate on evaluator noise alone, and averaging five replicates reproduces the ordering on each reference presentation to within one adjacent swap - in every case between two models the evaluator cannot separate anyway. Individual scores moved by up to 6 points; the ordering did not.

That claim is bounded, and the next section is where it stops. It was measured by re-evaluating a fixed set of generated presentations, so it answers would this presentation score the same if we measured it again. It does not answer would this model score the same if we asked it again, which is the question a reader of a leaderboard is usually asking. The second question is measured below, and the answer there is weaker.

Every replicate scored the same set of generated presentations, so these figures isolate the evaluator.

Generation variance

We then measured how much a score moves when a model simply generates the presentation again: three independent generations of every reference-and-model combination, each extracted and judged once. Generation variance is what remains once the evaluator's contribution is subtracted.

Regenerating a presentation moves its score about 1.8 times as far as our entire evaluator does, so roughly 77% of the spread on a single result is the model and 23% is us. Five cells carry a PowerPoint repair unevenly across generations; we pre-registered both figures, and excluding them raises the term to 4.52 rather than lowering it.

Repeatability differs between models. Within-cell standard deviation runs from 0.17 to 8.83 points: one model lands within a third of a point of itself three times running, another moves 17.5 points across three attempts. We publish it beside each score once it has been measured. Two models separate on this corpus only at a gap of 4.52 points, which 14 of the 28 pairs among the eight products in this study clear.

Operational observations

PowerPoint repair. 90% of the PPTX files generated by the OpenAI 5-series models triggered PowerPoint's repair flow when opened. Because PowerPoint repaired these files after the user clicked "Repair", we treated the issue as trivial file repair under our protocol and did not penalise it, but it is still an issue worth noting.

Generation time. We estimate that the Gemini PPTX outputs were produced approximately 5-10x faster than the other evaluated models. The Gemini outputs also received the lowest scores in this pilot. The timing estimate is descriptive only and does not establish a speed-quality trade-off; we did not instrument or formally measure generation time.

Future improvements

Some points in this methodology can be considerably improved in the future. We plan to prioritize them based on researcher and user feedback.

Expanding and refining the visual intelligence system schema

A company's or team's visual intelligence system is rich and complex. As such, there's still a lot in it that we don't measure well, or at all. We're constantly improving the schema and making sure we're capturing it better from companies' reference presentations. As a natural consequence of our definition of the visual intelligence system improving, this benchmark will become more difficult to pass, and the scores will likely trend downwards.

Expanding the evaluated set of reference-presentation-prompt tuples

The pilot results reported here are based on three reference-presentation-prompt tuples. We plan to expand the evaluation to more tuples from the dataset, so the preliminary scores are likely to change, and the model rankings in the leaderboard may change.

Expanding the reference dataset

We are currently working on expanding this dataset to 1,000 presentations with:

• Additional real corporate presentations - i.e. presentations created by business leaders and presented to a professional audience - from enterprise and mid-market companies;
• Presentations generated specifically for the benchmark by human experts - typically ex enterprise business leaders, ex bankers, or ex management consultants.

Expanding to the other pillars of slidemaking culture

The current methodology focuses on Visual Design. We plan to expand it to cover the other two pillars - Storylining and Content & Analysis - so the benchmark can evaluate a broader range of the conventions that shape how a company creates presentations.

Expanding the list of evaluated models

We are currently evaluating a handful of frontier models, but there are many more providers/models to benchmark.

Adding token cost and generation time to the benchmark

We are currently not tracking token/dollar cost and generation time for the evaluated models. These factors are important in real world professional document generation workflows, and we are considering building instrumentation into the benchmark to accurately track them.

Multiple prompts per reference presentation

With each reference getting a unique prompt, a low score could reflect a harder content task rather than weaker style imitation. We've made best efforts to balance prompt difficulty across reference presentations by keeping the prompts as similar as possible (and only varying as per the reference company's context), but this is not perfect. Multiple prompts per reference would make comparisons much stronger.

Improving the speed of the extractor and the LLM judge

The extractor and the LLM judge are currently key bottlenecks in the evaluation pipeline, both in terms of speed and reproducibility.

More generations per reference-presentation-and-model combination

We now run three independent generations per reference-presentation-and-model combination and report the mean, the median and the within-cell standard deviation. Three is enough to show that generation variance dominates evaluator variance; it is not enough to characterise the shape of the distribution, and a standard deviation from three points is itself an unstable statistic. We plan to raise N and to report best-of-N separately as a measure of peak performance under a fixed generation budget, using the same N and selection method for every model.