Part seven of the series AI Content: 10 Strategies That Work. A new part every Friday.
Before we get to practice, a detour that belongs in a series about AI content more than anywhere else. It is a story about what a marketing myth looks like from the inside.
The case of the sixty thousand
Almost every article about visual marketing carries the claim that the brain processes visual information 60,000 times faster than text. It is nonsense, and its history is worth retelling.
The earliest traceable appearance is an advertising section in Business Week from 1982, where it was stated without any reference by Philip Cooper, president of Computer Pictures Corporation. It entered wide circulation through a 3M corporate paper from 1997, where the figure sits in the second sentence, again without a citation.
In 2012 the educator Alan Levine tried to track down the original study. He went through libraries, called 3M, contacted Cooper himself and finally offered a 60-dollar reward to anyone who could find the primary source. The reward went unclaimed. The cognitive scientist Clark Quinn and the data journalist Jonathan Schwabish also searched without success.
The real numbers from research on stimulus processing look different. Detecting an animal in a photograph takes under 150 ms, recognising a face 130 to 160 ms, reading a word 100 to 200 ms and reading a sentence 300 to 400 ms. Visual and verbal processing take comparable amounts of time, so a 60,000-fold difference is nonsense on principle alone.
Why this story belongs in a series about AI
A language model will write that figure for you, because it appears online a thousand times over. It will write it confidently, in a correct-looking sentence, and quite possibly with a reference to 3M, because it really is there.
And you publish it, which makes it a thousand and one. That is exactly the mechanism by which errors stay alive, except that it used to take years and now it takes an afternoon.
Forty-four years in circulation is also a good calibration for how long your own mistake will last. An unpublished text is corrected for free. A published text you correct on your own site, but not on the sites that picked it up in the meantime.
What is actually documented about images
The peer-reviewed study by Li and Xie published in the Journal of Marketing Research in 2020 analysed three large datasets from Twitter and Instagram. The results are interesting precisely because of how conditional they are.
For airlines, an image raised retweets by 119 % and likes by 87 %. For SUVs, by 213 % and 151 %. But a tweet with a so-called linked image performed worse than a plain text tweet, so even „add an image“ does not hold unconditionally.
Human faces raised engagement on Twitter by 38 to 291 %, and had no effect at all on Instagram, where selfies are the norm. Colourfulness helped for air travel and hurt for SUVs. The only thing consistent across everything: high-quality, professionally shot images perform best and screenshots perform worst.
So do not use any single-number claim along the lines of „images raise engagement by X per cent“. The effect ranges from negative to several-fold depending on platform, sector and type of image, and anyone offering you one number either did not read the study or is hoping you did not.
How to brief images
A brief that works has three layers: what is in the image, in what style, and what feeling it should produce. An overly detailed description paradoxically narrows the result, an overly vague one leaves it to chance.
Instead of „create an image of investing“, try „minimalist illustration, a middle-aged man standing in front of a signpost with three arrows, flat colours, muted palette of blue and ochre, calm focused atmosphere, no text“. The difference is not the length, it is that the second brief decides for the model in the places where it would otherwise reach for the most common solution.
The instruction „no text“ pays off almost always, unless you specifically need text in the image. Models have improved considerably at generating text inside images, but with diacritics it is still a lottery and the accent above a letter goes missing in half the attempts.
If you do need text in the image, it usually works out cheaper to generate a clean visual and set the text into it in a graphics editor. It sounds old-fashioned, but it saves repeated generation.
Tools as they stand in August 2026
The situation changes month to month, so treat this as orientation and check prices with the provider. The cheapest bulk generation and editing of existing images today comes from Google via the Gemini API, where the price per image runs to single-digit cents depending on resolution.
For stylised aesthetics and building a visual identity, Midjourney remains strong, currently in version 8.2 from July 2026. For faithfully following a complicated brief and for text inside the image, OpenAI’s image model is a good fit. Adobe Firefly has the advantage of being trained on licensed data, which simplifies the question of legal cleanliness in commercial use.
For video, prices per second sit around ten cents for 720p with both Sora and Google’s models, and the credit systems of Runway and Kling start at roughly ten dollars a month. With video it goes double that the price of the first usable output is a multiple of the price of one generation, because it tends to be the third or fourth attempt that is usable.
Consistency matters more than any single image
Ten excellent but stylistically unrelated images look worse than ten average ones that look like they belong together. A brand is built by repetition, not by peaks.
Define three or four parameters you will hold across everything: the palette, the level of stylisation, a compositional rule and the type of lighting. Then put those into every brief as a fixed block you do not change.
It pays to keep them written down outside your head, ideally as a ready-made piece of text you copy into every prompt. As soon as it relies on memory it starts drifting within a month, and six months later you have a site that looks like a folder of downloaded images.
Previous part: Where AI ends and a human begins. Next Friday: Measurement that tells you something.
The whole series: AI Content
1. Why most advice about AI content no longer holds · 2. A publishing plan that assumes volume does not work · 3. The prompts that make a difference, and the fact-checking · 4. Personalisation that is more than a different salutation · 5. Visibility in the year people stopped clicking · 6. Where AI ends and a human begins · 7. Images, and how to spot a marketing myth · 8. Measurement that tells you something · 9. One piece of content, several channels, without copy-paste · 10. Law and ethics, specifically, as they stand in August 2026 · 11. Build a system, not a pile of articles · 12. What to take from the whole series