A new study by Adobe published today exposes a measurable and widening gap between how confident marketers feel about AI prompting and how well they actually perform when tested on it - with consequences that are already slowing down campaign production.

Adobe today published research examining how marketers write prompts for generative AI tools, drawing on a survey of more than 1,000 AI users in the United States. The study, conducted through Adobe Acrobat and Firefly, assessed both self-reported confidence and actual prompt quality using a structured scoring methodology. The findings are blunt: the average prompt scored 57 out of 100, equivalent to a C grade, and not a single respondent achieved an A.

The announcement date is May 21, 2026.

Abandonment rates and the cost of poor prompting

The most striking headline figure is the abandonment rate. According to Adobe, 91 percent of AI users surveyed have abandoned a generative AI task and reverted to manual work because the output was not usable. That figure applies specifically to situations where the respondent did not have the time or skill to craft a prompt that matched their requirements. The practical consequence is straightforward: teams investing in AI tools are still losing hours to rework, revision cycles, and manual fallback processes.

The study identifies a patience threshold that differs by task type. For image generation tasks, respondents expected a usable result after roughly four attempts. By the seventh attempt, most had given up. For text-based tasks - such as drafting emails or social media posts - the window was even narrower: respondents hoped for a solid result after two prompts and walked away around the fourth try. This short tolerance is compounded by the economics of AI tools. When a user burns through a free trial refining output, according to Adobe, they often conclude the tool is broken rather than their instructions.

Scoring methodology and what a C grade actually means

To move beyond self-reporting, Adobe asked respondents to study a target image and write a text-to-image prompt to recreate it. The original prompt used to generate that image in Adobe Firefly's AI image generator was: "You are a commercial artist. Please generate a charming, stylized cartoon bumblebee with large, expressive eyes and a friendly smile as it zips through a whimsical field with oversized, dew-kissed blades of grass under a crescent moon and twinkling stars, all rendered in the vibrant, neon-lit, somewhat psychedelic color."

Each submitted prompt was then scored against ten criteria, each worth ten points, for a maximum total of 100. The ten categories were: whether the prompt explicitly mentioned a bumblebee; whether it included movement words such as "flies," "zips," or "zooms"; whether it referenced celestial elements like a moon, stars, or night setting; whether it included natural elements such as grass, field, or flowers; whether it incorporated emotional adjectives like "happy," "cute," or "charming"; whether it contained a persona assignment (such as "you are a concept artist"); whether the word count fell within plus or minus five of the original's 49 words; whether extraneous detail was avoided; whether it included at least two style-specific terms such as "cartoon" or "photorealistic"; and whether it specified a color.

Across the full sample of 1,000 prompts, the average score was 57 out of 100. An A grade required a score between 81 and 100. Nobody reached it. The most common outcome was a C.

According to Adobe, the B-level prompt captured much of the image's tone and detail but missed two key elements - a persona assignment and a word count close to the original. Those omissions were enough to hold it below the A threshold. The C-level prompt identified the main subject but omitted movement, emotion, color, and a defined role for the AI. The result, according to the study, lost the neon lighting and expressive stylized look of the original "because those details never made it into the text prompt."

Generational confidence versus measured performance

The generational findings complicate a widely held assumption. Gen Z respondents reported the highest confidence in their AI prompting skills, with 54 percent self-rating their ability as four or five on a five-point scale. They were also 26 percent more likely than Millennials to claim high proficiency in text-to-image prompt writing for tools such as Adobe Firefly. But when prompts were actually scored, Gen Z averaged 56 out of 100. Millennials averaged 58.

The gap is modest in absolute terms - two points on a hundred-point scale. Its significance lies in direction: the generation with the highest stated confidence underperformed the generation that expressed more restraint. According to Adobe, this suggests Millennials are underselling their skills while Gen Z respondents are overestimating how good their prompts actually are.

This finding has practical implications for how companies allocate AI training resources. If Gen Z employees are assumed to be AI-native and therefore require less instruction, that assumption is not supported by the scored output in this study.

Hierarchy flattened by prompt quality

One of the more counterintuitive results involves seniority. Entry-level employees and directors scored identically in Adobe's assessment - both groups averaged 55 percent. Remote employees performed seven percent better than on-site workers. Men reported being 15 percent more confident in their prompting abilities than women, but their actual scored prompts were only five percent better on average.

These patterns suggest that prompt engineering skill - at least at its current stage of industry development - does not track neatly with years of experience, job title, or demographic profile. The skill is sufficiently new that most people are roughly at the same starting point regardless of seniority. That has implications for hiring, team structure, and how organizations think about distributing AI-related responsibilities.

PPC Land has previously documented how the AI skills gap is showing up across the broader advertising industry, including findings that internal expertise and training represent the primary barrier to broader AI adoption, cited by 45 percent of respondents in IAB Europe research.

Where prompts go wrong: omissions, not inclusions

Adobe's analysis of the 1,000 submissions found that the most common prompting failures were not about what people included but what they left out. Omitting an important detail causes the AI to fill in the gap itself, producing output that does not match the user's intent. The three most frequently cited mistakes were: failing to define the soft elements of the output, such as tone or personality; neglecting to provide a clear example of the desired output style; and omitting a specific role or persona for the AI to assume.

The persona assignment category is particularly notable. Writing "you are a commercial artist" before specifying the task is a structural technique that tells the model how to interpret and weight the rest of the instructions. According to Adobe, this type of role-priming shapes the output's approach, vocabulary, and stylistic defaults. It is one of the clearest technical distinctions between how most users prompt and how prompt engineers approach the same task.

The ALL CAPS finding and prompting myths

The study surfaced at least one widely held belief about AI prompting that does not hold up. About one in seven respondents believed that writing in ALL CAPS leads to better output. According to Adobe, men were 80 percent more likely than women to agree that "shouting" at AI produces better results. In practice, capitalizing words does not influence a language model the way it might affect a human reader. Adobe notes that capitalizing certain constraints - such as "DO NOT include personal information" - can serve as an emphasis technique by helping a model prioritize specific instructions, but this is a narrow functional application, not a general-purpose improvement strategy.

Politeness also varies by sector. According to the study, respondents in finance and banking said "please" 43 percent of the time. Those in education, transportation, and logistics said please 42 percent of the time, followed by creative arts at 38 percent and healthcare at 36 percent. Adobe's position is that politeness rarely translates to better output and that focusing on tone, style, and persona assignment is a more effective use of prompt length.

Reuse, templates, and the manager gap

Not all patterns in the data reflect individual skill. Some reflect organizational habits. According to Adobe, more than half of AI users surveyed - 55 percent - have adopted a workflow in which they ask the AI tool itself to review and improve their prompts before proceeding. This is described in the study as an advanced workflow that functions as an efficiency loop: generate a draft response, ask the AI to critique it, then revise the prompt before generating a final output.

There is a notable gap by seniority in how prompts are saved and reused. Manager-level employees were 39 percent more likely than entry-level respondents to save successful prompts as templates for future use. This templating behavior - treating effective prompts as reusable assets rather than one-off instructions - is one of the clearer markers of mature AI integration in a marketing workflow. The study frames it as a form of institutional knowledge management that compounds over time.

Research published earlier in 2026 on PPC Land found that when AI generates vast volumes of content at near-zero marginal cost, competitive advantage shifts toward harder-to-automate skills including creative judgment and strategic framing. Prompt quality sits at the intersection of those two forces: it is both a technical skill and a form of creative direction.

Fact-checking and industry-level gaps

The study also examined post-generation behavior. According to Adobe, one in seven respondents does not fact-check their AI outputs at all. Fact-checking rates varied considerably by industry. Business professionals were most likely to skip the step, with 24 percent reporting they do not verify AI-generated content. Healthcare professionals came in at 17 percent, and retail and e-commerce respondents at 16 percent.

Adobe's framing here is that AI prompting skill is not determined by job title or industry - it is determined by the habits people bring to the tools they use. The fact-checking gap is treated as a workflow problem rather than a literacy problem. Engineers, according to the study, treat fact-checking as a standard part of the process whether they are reviewing text or scanning AI-generated imagery for errors such as distorted body parts, incorrect reflections, or unreadable text.

What distinguishes a prompt engineer from a typical user

The study draws a clear line between how ordinary AI users approach prompting and how engineers structure the same tasks. Most respondents relied on single, catch-all prompts. The highest performers, according to Adobe, took a different approach: they broke prompts into smaller, sequential steps, giving the AI focused guidance at each stage rather than asking it to handle everything at once.

The study outlines eight practices associated with higher-quality outputs. These include setting voice and tone explicitly, assigning the AI a role at the start of the prompt, providing an example of the desired output structure, incorporating fact-checking as a step rather than an afterthought, reviewing images carefully for common generation errors, saving strong prompts for reuse, using an efficiency loop in which the AI critiques an initial draft, and breaking complex tasks into a sequence of smaller prompts.

None of these practices require specialist technical knowledge. They are workflow habits - structured approaches to instruction-writing that produce more consistent results. The study's implicit argument is that the gap between a 57 out of 100 and a score in the 80s is primarily a matter of learned behavior, not innate ability.

Context for the advertising and marketing industry

The findings land at a moment when AI tooling is deeply embedded in marketing campaign production. Adobe's own Firefly platform has expanded considerably over the past year, adding video generation through a partnership with Runway announced in December 2025 and a strategic infrastructure deal with NVIDIA announced in March 2026 that targets next-generation Firefly foundational models. Adobe also launched an enterprise tool for AI visibility optimization in October 2025, which generated a fivefold increase in Firefly citations within one week of deployment on Adobe's own properties.

Separately, data published on PPC Land in January 2026 showed that AI usage among marketers concentrates heavily in tasks requiring high creativity and cognitive complexity - precisely the category where prompt quality has the most direct impact on output. If the tools are being used most intensively for the tasks where prompting is hardest, the 57 out of 100 average represents a real drag on the value that AI investment is generating.

The study was conducted by Adobe Acrobat and Firefly and surveyed 1,008 AI users across different skill levels. The data carries a 95 percent confidence level with a three percent margin of error. Because the research relied on self-reported data for some components, Adobe notes that respondents may have biases and that discrepancies may exist between their answers and actual experiences.

Timeline

Summary

Who: Adobe, through its Acrobat and Firefly products, surveyed 1,008 AI users across different skill levels and experience types in the United States.

What: The study scored 1,000 text-to-image prompts against a ten-category rubric, found an average score of 57 out of 100 (a C grade), recorded a 91 percent task abandonment rate due to unusable AI output, and identified a systematic gap between generational confidence and actual measured prompt quality - with Gen Z scoring lower than Millennials despite reporting higher confidence.

When: The research was published on May 21, 2026. The survey data carries a 95 percent confidence level with a three percent margin of error.

Where: The study focused on American AI users working across industries including finance, healthcare, retail, education, transportation, and the creative arts.

Why: As AI tools become embedded in standard marketing and content production workflows, the quality of the instructions given to those tools - the prompts - determines the quality of the output. The study makes the case that prompting is a learnable workflow skill, that most current users are performing well below their potential, and that the gap between a C-grade prompt and a B-grade prompt often comes down to a small number of structural choices that can be practiced and systematized.

Share this article
The link has been copied!