Here’s an admission that sounds wrong for someone in my job: the most useful production tool I own is not a video model. It’s a reading habit from an old corner of linguistics.
I’m the in-house AI creative director at a major FMCG company. I’ve spent four years shipping AI-made commercials past real brand teams, and in that time the models have solved almost everything demos care about — resolution, motion, light, skin. What they have not solved, and cannot solve on their own, is meaning. An AI ad can now look like anything. The open question is whether it says anything.
Semiotics — the study of signs and how audiences read them — is the discipline that deals with exactly this gap. Most people meet it as academic jargon. I use it as a checklist, every week, on deadline. This article is that checklist, with the theory kept to the minimum that earns its place.
What is semiotics, and why should an ad maker care?
Strip the vocabulary away and semiotics says one practical thing: an audience never watches your footage. It reads signs — compact units of meaning it decodes without noticing. A slim tea glass says home in one market and nothing in another. Golden hour says nostalgia. A hand opening a pack says this is for you.
The classic distinction worth keeping is denotation versus connotation. Denotation is what’s literally in the frame: a family, a table, breakfast. Connotation is what the frame means to a specific culture: comfort, aspiration, ritual, us. Two ads can share a denotation and mean opposite things.
Why does an AI ad maker specifically need this? Because generative models are trained on denotation at planetary scale and on your audience’s connotations not at all. A model knows what ten million breakfasts look like. It does not know what breakfast means on your street. That missing layer doesn’t ship in any model update, and it’s precisely the layer advertising is made of. Someone on the team has to supply it. In my pipeline, that supply job has a name and a process — and it’s the part of my work no tool has touched in four years.
Why do AI ads drift toward the generic?
In my article on brand consistency I described “regression to the generic”: a model’s constant statistical pull toward the average of everything it has seen. There I treated it as a visual problem — packs and colors drifting off-model. The semiotic version of the same force is worse, because it’s invisible in any single frame.
A generative model is an averaging machine, and meaning lives in deviation. Your brand’s signs — the odd pack shape, the specific gesture, the particular light your brand lives in — are meaningful because they differ from the category default. The model, left alone, sands every one of those deviations back toward the mean. The output is what I call beautiful nothing: footage with the grammar of advertising and none of the vocabulary of your brand. Cinematic, polished, and interchangeable with a competitor’s ad by swapping the pack.
Here’s the trap: beautiful nothing passes review. It looks expensive. Nobody can point at a defect. In my own selects, where I reject 60–70% of generated takes on brand work, the visually broken takes are the easy rejections. The dangerous ones are gorgeous and mean nothing — and someone always fights to keep them. A brief with no semiotic layer doesn’t fail loudly. It fails three months later, in tracking, when nobody remembers whose ad that was.
What does a sign audit look like in practice?
Before day one of my six-day production pipeline, I run what I call a sign audit: list every sign the brand actually owns, decide its lock level, and only then start generating. For a typical FMCG brand the meaningful list is short — 5 to 9 signs. If your list has thirty, you’ve listed styling, not signs.
Here’s the audit structure, with rough numbers from my own project files:
| Sign layer | Examples | Typical count per brand | Lock policy | Share of my sign-related rejections |
|---|---|---|---|---|
| Identity signs | pack shape, logo, brand color | 2–3 | absolute — never varies | ~55% |
| Ritual signs | how it’s opened, poured, shared | 1–2 | locked per campaign | ~20% |
| Cultural signs | objects/settings that say “us” to this market | 1–2 | locked per market | ~15% |
| World signs | light, pace, season the brand lives in | 1–2 | directional — may flex | ~10% |
| Styling | everything else: kitchens, wardrobes, weather | unlimited | free to vary | 0% — variation is the advantage |
Two things this table encodes. First, the bottom row is not a compromise — it’s the economics. Free variation in the styling layer is how AI production gives you 10 hook versions for testing at marginal cost. Guard the signs; free the world. Second, the last column is the honest part: when I trace my rejected takes to a reason, sign violations dominate, and identity signs dominate the sign violations. The model attacks your most distinctive assets first, because distinctiveness is exactly what it’s built to normalize.
How do I use semiotics inside an AI pipeline?
A concrete case from my FMCG work. We were making a home-moment scene for a beverage product aimed at our regional market. The model, across take after take, put a big Western ceramic mug in every hand. Denotatively fine — it’s a hot drink in a container. Semiotically, dead. In our market the sign that carries home, warmth, family is the slim tea glass. A viewer here doesn’t consciously register the mug; they just feel the scene is about someone else’s home. That feeling has no line item in any QC checklist — unless the checklist is semiotic.
Because “tea glass, never mug” was written into the sign audit before generation, the fix cost minutes: it was a rejection criterion, applied at selects like any other. Caught after the edit, it would have cost a day of regeneration. Caught by the audience, it would have cost what off-target ads always cost — politely ignored media spend.
That’s the whole method, honestly. The sign audit becomes part of the reference set. Selects check meaning alongside accuracy: right pack and right signs. And on the assembled film I do one dedicated pass asking a single question per sign: is it present, is it legible, is it ours? None of this adds a day — it rides inside the checks I already run. What it adds is a reason behind every rejection that “I don’t like it” can’t give you. When a stakeholder asks why I killed the prettiest take, “it violates ritual sign #2 from the audit we agreed on” ends the argument. Taste, documented, stops being taste and becomes direction.
Can AI actually generate meaning, or just style?
Style — yes, endlessly, better every quarter. Meaning — no, and I want to be precise about why, because this is where hype and dismissal both get it wrong.
Meaning isn’t a property of an image; it’s a transaction between an image and a culture. Models generate the carriers — frames, motion, light. Which signs those carriers activate in your market, this season, for your audience: that’s a reading, and reading requires a reader. The model has statistics where the reader has a childhood. This is also why “the models will get better” doesn’t close the gap. Better models produce better carriers. They still can’t know that the mug should be a tea glass, because that fact isn’t in the pixels — it’s in the people watching.
I’ve written elsewhere about mythology in advertising — the deeper layer where brands plug into stories a culture already tells itself. Semiotics is that same conviction at production scale: myths are built out of signs, one locked frame at a time. And it’s why I keep insisting, against the prompt-course industry, that the scarce skill in AI advertising is not prompting. Prompting is spelling. This is writing.
If you take one thing from this article: before your next AI campaign, spend one hour listing the 5–9 signs your brand actually owns and the meanings they carry for your specific audience. That hour will do more for the work than any tool subscription this year.
FAQ
Is “semiotics” just a fancy word for brand guidelines?
No — guidelines and semiotics answer different questions. Guidelines specify assets: this logo, this hex value, this typeface. A sign audit specifies meanings: which elements carry them, for which audience, and what must never vary. Most guidelines I’ve worked with cover identity signs well, ritual and cultural signs barely, and world signs not at all — which is exactly where AI production drifts first.
Do I need a semiotician on staff to do this?
No. You need one hour, the questions in this article, and someone who genuinely knows the audience — often a founder or a local marketer, not a consultant. The audit is a one-page document: 5–9 signs, their meanings, their lock levels. The discipline is applying it to every take, which is process, not credentials.
Does this apply to performance ads, or only brand films?
Performance work arguably needs it more. When you’re testing 10 hook versions, the styling layer varies wildly by design — and the sign layer is the only thing telling the audience all ten came from the same brand. Sign-locked variation is how you test aggressively without spending media budget teaching people someone else’s ad.
How does this change what I actually type into the model?
Less than you’d expect — the audit changes what you reject, not just what you prompt. Signs go into prompts and reference images, yes. But models ignore instructions often enough that the real control point is selects: every take checked against the sign list before it enters the edit. Prompting is a request; rejection is a decision.
I’m Mohammad Nasravi — in-house AI creative director at a major FMCG company. Everything here comes from campaigns that shipped past a real brand team, tea glasses included. If you want ads that mean something to your market — not beautiful nothing — see my services or get in touch. I reply within 24 hours.