How to make AI content that doesn't look AI-generated

Two years ago, AI-generated posts got attention because they were novel. That window has closed. Audiences have seen enough to recognise it in under a second — the particular sheen on the image, the particular rhythm of the caption — and recognition is the problem. The moment a post reads as machine-made and nobody-specific, it stops being a message from a business and becomes filler. People scroll past filler.

This is not an argument against using AI. It is an argument against shipping its first draft. The businesses getting real results are not the ones with better prompts; they are the ones who treat generation as one step in a process that also includes a locked brand look, real inputs from the business, and a human deciding what goes out.

Why unedited output underperforms

A general-purpose model produces the average of everything it has seen, which is exactly what you do not want. The average is the most generic possible version of your message, so it looks and sounds like every competitor who typed a similar prompt. Marketing works by being distinctive and being credible, and raw generation undermines both at once: not distinctive because it is averaged, not credible because it contains nothing that could only have come from you.

So the failure is not about quality in the abstract. A generated image can be technically impressive and still useless, because it does not look like your business. Recognisability is the asset, and averaged output has none.

The visual tells

People cannot always explain why an image looks generated, but they clock it anyway. These are the signals doing the work:

• The gloss. Over-rendered highlights, surfaces with an even plastic sheen, everything lit as though wet. Real photographs have dull patches, blown highlights and dust.
• Stock-photo composition. The perfectly centred subject, shallow depth of field on everything, the pristine empty workspace. It is the visual grammar of a stock library, and people have been trained to ignore it.
• Brand colours and type drifting between pieces. One post slightly warmer, the next in a different typeface weight, a third with its own accent colour. Individually invisible, obvious the moment they sit together in a profile grid.
• Mangled text and hands. Signage with almost-letters, a price that is not a real number, a logo that is nearly right, six fingers. Small, but it is the detail people screenshot and mock.
• Lighting that does not match across a set. Three images for one campaign, each lit from a different direction at a different time of day. A real shoot happens on one afternoon and looks like it.
• Subjects who look like nobody at your business — aspirational, thirty-two, wearing clothes nobody wears on that job. Local audiences notice fastest, because they know what the staff look like.

The written tells

Written output gives itself away differently, because reading is slow enough for the reader to notice the pattern.

• Uniform paragraph rhythm. Every paragraph the same length, every sentence the same length inside it. Human writing is lumpy — a long explanation, then four words.
• Hedge words everywhere. Things are often, typically, generally, may potentially help. Nothing is ever simply true, so nothing lands.
• Listicles with no specifics. Five tips that would fit any business in any industry, because none of them names a product, a price, a place or a situation.
• Claims with no numbers and no named examples. Improved results, significant growth, better outcomes. A reader who has been sold to before treats unsupported superlatives as noise.
• Nothing only this business would know. No mention of the job that went wrong last March, the question every customer asks about delivery, the material you stopped using and why. It is the biggest tell of all, and it is an absence rather than a mistake, which is why people miss it.

The last three are the same problem wearing different clothes: specificity. A model can imitate a tone of voice quite well. It cannot invent the details of your business, because it has never been there.

The fix is a system, not a better prompt

Every tell above is a production failure, not a technology failure. They are fixable with process, and the same process works whether you are making one post or thirty.

1. Lock the brand kit before you generate anything

Decide the palette, the typefaces and their weights, the logo placement rules and the tone of voice, then write them down and stop re-deciding them. Most inconsistency in AI content comes from the look being re-invented each time an asset is made. A locked kit turns hundreds of small aesthetic decisions into zero.

2. Approve one reference piece, then scale

Produce a single asset, get it right, get it signed off — and only then make the rest. That approved piece becomes the specification: later assets inherit its lighting, crop, colour treatment, type sizing and caption voice instead of starting from nothing. This is what converts a pile of individually acceptable images into a coherent set.

It is how we run production at Booltspace. A client approves one style sample per campaign, and the team then produces the whole monthly quota in that approved style — so the review happens once, where it changes everything downstream, rather than thirty times on finished work.

3. Feed it real inputs

Specificity is the one thing a general model cannot fake, so supply it. Photographs of your actual products under your actual lighting. Real questions customers asked this month, in their words. Photos from real jobs, including the unglamorous ones. Your real prices. The name of the town. The objection you hear on every call.

Content built on these inputs stops sounding generated for a simple reason: it contains information that only exists inside your business. Generic output is what you get when the prompt is all the model has.

4. Put a human in the curation seat

Generation is cheap, so generate freely and then be ruthless. Someone has to decide what ships: rejecting the image with the broken sign, rewriting the caption that hedges, killing the post that says nothing. The failure mode of AI content is not bad generation; it is publishing everything that came out.

5. Check the set, not just the asset

Before anything goes out, lay the month's assets side by side at phone size, in a grid. Drift that is invisible at full resolution is glaring in a grid, and the grid is how a profile gets judged.

Where AI genuinely wins

The answer to generic output is not to go back to posting twice a month, so it is worth being clear about the upside.

• Volume. A consistent posting rhythm is what most small businesses fail at, and cheap production makes it survivable.
• Iteration speed. Look at a version, say what is wrong, see the next one immediately. That loop used to take days.
• Cheap ad testing. Running many creative variations to find the one that works used to need a large budget. Now it does not.
• Repurposing. One good idea becomes a square post, a vertical video, an email and a blog section. The idea is the expensive part; the formats no longer are.

Where it should not be used at all

There is a line, and it is not a stylistic one. Do not use generated content to imply that a real person said or did something they did not. That covers synthetic voices or likenesses of named people, invented quotes from staff or customers, and fabricated endorsements.

It also covers two things businesses talk themselves into surprisingly often: invented customer testimonials, and before-and-after results that never happened. Both are deceptive, and consumer-protection regimes in the United States, the UK and the EU all treat fake reviews and unsubstantiated performance claims as enforceable matters. The trust cost is worse than the legal one — a single exposed fake testimonial invalidates every true thing on your website.

The same applies to synthetic imagery presented as documentary evidence. A generated illustration used as an illustration is fine; a generated image passed off as a photograph of a job you completed is not. In trades where the photograph is the proof, that is the fastest way to lose a client permanently.

What this means for search

There is a persistent worry that using AI will get a site penalised. That is not how Google describes its position: the guidance focuses on whether content is helpful, original and written for people, rather than on how it was produced. Automation is not itself the problem.

What is targeted is scaled content abuse — mass-producing unoriginal pages primarily to manipulate rankings rather than to help anyone. That distinction lines up with everything above. Content built on your real inputs, curated by someone who decided it was worth publishing, is original and people-first by construction. Content generated in bulk to fill a keyword list is exactly what the guidance is aimed at, and it was a bad idea anyway.

The short version

1. 1Lock the brand kit — palette, type, tone — before generating anything.
2. 2Approve one reference piece, then make everything else inherit its look.
3. 3Feed in real products, customer questions, job photos and prices.
4. 4Have a human decide what ships, and reject freely.
5. 5Review the whole set at phone size before anything goes out.
6. 6Use AI for volume, iteration, ad testing and repurposing.
7. 7Never fake a person, a testimonial, a result, or a photograph of work you did not do.