AI in digital marketing
In marketing, AI produces the first version and not the last; it is good at drafting and at finding patterns, and bad at telling what is true. The most valuable thing you can do is build the difference between those two into your own workflow.
- Lesson 8 of 8
- Intermediate
- Free, no signup
The six jobs AI does around a marketer
Each cell is a job a model genuinely does; the failure mode of each is in the body of the lesson.
None of these six cells includes making the decision. The decision stays in the core, and this figure deliberately leaves it there.
Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.
Where AI works in marketing, and where it fails
Six jobs, and each has one clear failure mode that stops surprising you once you know it.
Drafting copy. Genuinely fast, and it removes the blank page. Failure mode: the text comes out averagely good and distinctly flavourless, and every sentence that needs a fact about your business is either left empty or invented. Audience research. Excellent for sorting hundreds of comments and messages and finding the questions that repeat. Failure mode: give it no data and the model builds a picture of the average global customer, who has nothing to do with yours.
Creative variants. Ten versions of a headline or ten crops of a banner, in minutes. Failure mode: choosing is still your job, and the model does not know which of them is wrong for your brand. Analysis. Finding the unusual rows in a table, which the analytics lesson carries a recipe for. Failure mode: let it draw conclusions and it fabricates reasoning rather than numbers, and fabricated reasoning is harder to spot.
Chat handling. Answering the repeated questions, around the clock. Failure mode: the moment a question leaves the circle of what it was told, the model invents a price, a condition or a delivery time, and the customer treats it as your promise. Translation and localisation. Carrying a message into another market without starting over. Failure mode: tone does not translate, and trade terms come out wrong.
The pattern repeating across all six: wherever the work is shaping and sorting, the model helps; wherever the work is knowing, the model is a hazard. The speed you gain belongs to the first kind, and if you spend it on the second, you spend the same time again later cleaning up.
Why every competitor's feed looks the same now
Because everyone's input has become the same. Three competitors in one trade call the same model with a similar request, and naturally get three texts sitting in the middle of one distribution. The model has done nothing wrong; writing the most probable next sentence is exactly its job, and the most probable sentence is the one everybody writes.
For the reader the result is a vague feeling rather than a diagnosis: nobody says this text was written by AI, they simply stop remembering any of these posts. And that loss of effect is the real cost, far larger than any penalty you might be imagining.
The way out is not a matter of taste either; it is a matter of input. What a model cannot produce is the things only you have: a number you measured in your own work, a mistake you made and learned from, a sentence a customer said last week, a boundary where you are willing to say your work does not reach. Every sentence of that kind is one your competitor cannot obtain with any prompt. If you do not fill the machine draft with these, there is nothing left to be different about.
What the reader sees, and what sits underneath
A flawless, forgettable post
nobody says a machine wrote it; they simply do not remember it
-
One model for every competitor
three competitors, one model, three texts from the middle of one distribution
-
A near-identical request
the same few sentences everybody puts in the prompt
-
No raw material from your own work
no number you measured, no mistake you made
-
No position a competitor would avoid
text that upsets nobody convinces nobody either
This figure does not claim readers can identify machine text. Its claim is that losing effect is itself the cost, even when nobody notices anything.
Three things not to do with AI
These are not matters of taste; all three are written into somebody's policy text and the source can be read.
Buying engagement. Bots for likes, comments and followers are cheap and available now, and Meta's spam policy prohibits it in as many words: "attempting to or successfully selling, buying, or exchanging for engagement, such as likes, shares, views, follows, clicks, use of specific hashtags, etc.", and likewise selling or buying "platform assets, such as accounts, groups, pages, etc." Beyond the rule, the practical reason is that the only number genuinely yours is your comparison against your own previous month, and buying engagement destroys exactly that number.
Manufacturing reviews and customer experience. This is where the temptation is greatest and so is the risk. In 2024 the US Federal Trade Commission finalised a rule prohibiting the sale and purchase of fake reviews and testimonials, and its announcement names "AI-generated fake reviews" explicitly; it also prohibits paying or incentivising someone on the condition that their review express a particular sentiment. That rule is American law and we make no claim about its legal counterpart in Iran, nor did we find a source to cite; but if you sell to customers outside Iran, it applies to you directly.
Speaking as a person without saying it is a machine. An answering bot in the inbox is fine; an answering bot that presents itself as human and then invents a price both costs you trust and creates a legal problem. In this building we decided that an automatic decision never earns the "reviewed" badge, and the reason is in the first-hand section of this page.
What the speed is worth having, and what it is not
Worth it
- the first draft, when you rewrite it yourself
- summarising customer comments and messages
- producing several versions of a headline to choose from
- a bot that says it is a bot and answers only what it was told
Not worth it
- buying likes, follows and views; Meta's policy prohibits it
- invented reviews and customer experiences, even "as a sample"
- a bot presenting itself as a human
- publishing a price or specification the model wrote and nobody checked
The Federal Trade Commission rule is American law. We make no claim about an equivalent legal position in Iran, because we found no source to cite.
What Google and the platforms do about AI content
Google's position is clearer than the market repeats. On its own page it says "appropriate use of AI or automation is not against our guidelines", meaning it is not used to generate content primarily to manipulate search rankings, which is against its spam policies. And on the same page: "Using AI does not give content any special gains. It is just content." So there is neither a penalty for using it nor a reward; the usual criteria stay where they were. The lesson on AI content and SEO opens this up in more detail.
With images the story is different, and this is where the surprise usually happens. Google has a tool called SynthID which, in its own words, "embeds digital watermarks directly into AI-generated images, audio, text or video"; those watermarks "are embedded across Google's generative AI consumer products" and "are imperceptible to humans but can be detected by SynthID's technology". Google also says it is designed to survive modifications such as cropping, filters, frame-rate changes and lossy compression. So an image coming out of these tools carries a label you cannot see and cannot easily remove. The image watermark lesson in this section covers it.
As for detecting text: AI text detectors err in both directions and we trust none of them for a decision. A positive result is not proof and a negative result is not innocence. If you read somewhere that a given tool detects with a stated accuracy, you heard that number from the vendor of that tool.
The fast path, with AI
If you only do one thing from this lesson, do this: sit down once and build a short file recording your tone, your permitted claims and your boundaries, and from then on put that file first every time you write, and your request second. It is the only move that directly stops you sounding like your competitor, because it separates your input from theirs. Building it takes half an hour and it works for months.
- Write five sentences that are allowed in your copy and five that are forbidden. The forbidden ones matter more: no guarantee of results, for instance, and no number that is not already on your site.
- Add three to five pieces of real raw material: a number you measured yourself, a mistake and its lesson, a real sentence from a customer, and one thing you do not do.
- Paste in two examples of copy you wrote yourself and were happy with. A model learns more from an example than from a description.
- Put the file first in every conversation. For mechanical work such as producing short variants, a cheap Flash-class model is enough; for text that will actually be published, use the stronger one. Today's pick is in the AI section of this site.
- Refresh the file every three months and put new raw material in. A file that never changes eventually becomes a repetitive template of its own.
Copy-ready recipe
### Brand voice file for {business name}
Tone: {three adjectives, for example plain, direct, no slogans}
Audience: {exactly who}
Permitted claims (only these may be written):
- {claim one}
- {claim two}
Forbidden claims:
- any guarantee of a result or a ranking
- any number or percentage not in the list below
- {a forbidden claim specific to your work}
Real, permitted numbers:
- {number one, and where it came from}
- {number two, and where it came from}
Raw material (use these, not generalities):
- a mistake we made: {what happened and what we learned}
- a real customer sentence: {quote}
- something we do not do: {the boundary}
An acceptable sample:
{a short piece of copy you wrote yourself}
---
Using this file, write {what you want}.
Rule: wherever a number or a fact is needed that is not in the file above, write {unknown} instead of writing it, and list at the end what I have to fill in.Before you trust the output: This file fixes tone, not honesty. Three things remain your job. First, wherever the model left {unknown}, either fill it with a fact yourself or delete the whole sentence; leaving it is the worst option. Second, the raw material you put in the file has to have actually happened, otherwise you have merely made one lie repeatable. And third, text produced with this file is still a draft; publishing it unread is exactly what this lesson says not to do.
AI in this kind of work
This lesson is itself about AI, so this block has one job only: to say which tool suits which part of marketing work, and which of them an Iranian can actually reach. The second answer is the same and unwelcome for all three tools below, and we write it as it is; we neither work around it nor teach anyone how to.
Tools that actually help
- Claude For text that will be published, because it holds to constraints and leaves a blank empty when you say write nothing outside the file. Iran is on neither of Anthropic's two supported-country lists, and we read that on Anthropic's own page.
- Gemini For bulk mechanical work: ten versions of a headline, sorting hundreds of comments, summarising a table. Persian is on the supported-language list for the web app but Iran is not on the supported-country list; both from Google's own help pages. An image out of Google's generative tools also carries the SynthID watermark.
- ChatGPT The most widely used tool in Iran, and it works for drafting and ideas. We make no claim about official access: OpenAI's supported-countries page returns 403 to our server, and we do not write a claim with nothing behind it.
Where it backfires
Three risks, and all three have sources. First images: Google's generative tools embed the SynthID watermark in their output, "imperceptible to humans but detectable by SynthID's technology", and built to survive cropping, filters and compression (the DeepMind page). So a brand filling its feed with such images is choosing to make that fact detectable, even when it does not know it.
Second, bought engagement. Meta's spam policy prohibits selling, buying or exchanging engagement and platform assets, and also states that content designed to deceive or mislead users in order to artificially increase viewership is not allowed (the policy text). Third, fake reviews and testimonials: the US Federal Trade Commission's 2024 final rule prohibits their sale and purchase and names "AI-generated fake reviews" in its own announcement (the FTC announcement). That last one is American law; if you have customers outside Iran it applies to you directly, and if you do not, the ethical reason stands on its own.
And one risk with no external source, because it comes from inside the work: the more of the customer conversation you hand to a machine, the less you understand what the customer wants. Those strange questions that arrive in the inbox once a week are where the next product comes from; if the bot answers them and you never read them, you have traded speed for the only free feedback you had.
Sources: Google DeepMind: SynthID Meta: Community Standards, spam policy FTC: final rule banning fake reviews and testimonials Anthropic: supported countries and regions Google: where Gemini Apps are available
Where this advice stops
Four things are not here, and it is worth knowing why. First, any number about how much Persian content is produced with AI today; no such measurement was available to us, and guessing is precisely the behaviour this lesson criticises. Second, we recommend no detection tool and quote no accuracy figure; these tools err in both directions and their accuracy numbers are usually published by their own vendors. Third, we have no claim about the legal position on AI disclosure in Iran and found no source to cite; what we quoted is an American rule and we said so where we quoted it. And fourth, the tools and the policy pages are all changing, and what is quoted here was read on the verification date of this page; read those pages again before deciding anything.
From our own work
We run an AI classifier across three of our own marketplaces: freelance, the digital-product market and the advertising market. One shared file at wp-content/mu-plugins/rgb-modai.php takes a verdict and a confidence number for each new listing and then publishes it, rejects it, or holds it for a human. Two decisions we made there are useful for this lesson, because both run against what you would first assume.
First, the threshold to reject is higher than the threshold to publish: in the default settings min_publish is 60 and min_reject is 78. We reject deliberately more reluctantly than we publish, because a wrong rejection costs a real seller while a wrong publication can be taken down afterwards. Symmetry looks like the default option in a system like this, and it is wrong.
Second, a listing published automatically never receives the "reviewed" badge. That badge comes from a separate meta field that only the approve button in the moderation queue writes, and the comment above that line in includes/ads/ai-moderation.php states the reason: the badge means a human looked. This is exactly what the section on the three forbidden things said: print a trust mark only when somebody stands behind it. We enforced it on our own tool, and the same rule holds for reviews and customer testimonials.
Real follow-up questions
If I write content with AI, will Google penalise me?
Google states that appropriate use of AI is not against its guidelines, and that using it gives content no special gains either. What is against its policies is generating content primarily to manipulate rankings, with or without a machine. The criterion remains the usual one of quality and usefulness.
Should I put an answering bot in my inbox or not?
If the volume of repeated questions is high, yes, on two conditions: it says it is a bot, and it answers only from what you gave it rather than from the model's general knowledge. And read the messages it could not answer yourself once a week; those are the most valuable part of your inbox.
My competitor posts several times a day with AI; should I do the same?
Do the same and you become one more text from the middle of the same distribution, and nobody sees a difference. What works is writing less but writing what they cannot write: a number you measured yourself or a mistake you made yourself. Volume is something a competitor can double tomorrow; first-hand experience is not.