Typical AI visibility mistakes I have run into

There is a company with serious manufacturing capacity that spends several hundred thousand forints a month on Google Ads. Fine, that is not a crime in itself.

The problem is that its entire digital presence is, to a machine, essentially unreadable. The company profile states two contradictory capacity figures on the same page. One is two and a half times the other. Take your pick.

The "fresh" company deck turns out to be a renamed copy of material from years earlier, references are nowhere to be found, and all of it exists in Hungarian only — at a company that works for export.

In this situation the AI does not necessarily start wondering "hm, something is off here". It simply says nothing.

A language model is not going to argue with you about which capacity figure is the true one. It will quote someone else instead, someone with nothing to untangle.

Meanwhile the advertising keeps taking the money. Because advertising happens to be the channel where you can also pay for not having to be particularly comprehensible.

This is one of my personal favourites.

A healthcare provider spends a hundred thousand forints a month on Google ads while its robots.txt file — because of a misconfigured Cloudflare setting — blocks Google's bots.

Not one of them. Not a few.

Every crawler. Everything.

"But the advertising works!"

It does. And here comes the technical punchline that turns the whole thing from funny to expensive: Google's advertising bot, AdsBot, bypasses the general robots.txt block.

So precisely the part of the system that takes money out is working. The part that could bring visitors in for free is, in practice, dead.

Organic results: zero.

AI mentions: zero.

Invoice: precise, monthly, dependable.

It is roughly like paying a doorman to invite the guests in while the sign on the door reads: NO ENTRY.

The doorman is enthusiastic. The door is locked.

A marketing provider nicely rewrote the human-facing part of its website. New messaging, new positioning, new copy.

Congratulations, it genuinely got better.

Except that the machine layer — the metadata, the structured data and everything else a language model works from — went untouched.

So the human already sees the new company. The machine still reads the old one.

The result is a kind of digital split personality: ask the AI about the business and it describes a company that, by its own account, no longer exists.

A bit like sitting through a video call in a shirt with pyjama bottoms on. Except in this case the machine happens to see under the desk too.

Yes, I brought an example from my own trade as well. The cobbler's shoes are alive and well around here too.

A company working on AI visibility cobbled together its own structured data, its JSON-LD, from elsewhere. That alone would merely be embarrassing.

The trouble is that values survived the copy according to which the company provides hotel services, and its products include the preparations of a foreign cosmetics firm.

A human visitor notices none of this.

The machine, however, reads exactly that, since the JSON-LD is made for it. This is, in effect, the company's official statement about itself to machines.

And that statement says they rent out rooms and sell face cream.

The AI does not necessarily open an investigation. It processes what it was given.

Anyone who would build someone else's AI visibility had better tidy up their own first.

A classic WordPress story.

What was needed was a language switcher between the Hungarian and the English version. What was born instead was a separate domain, carrying /hu and /en slugs, with no meaningful connection between them.

No hreflang. No canonical relationship. Nothing to tell the machine that these are in fact two language versions of the same company.

In the machine's eyes there is therefore no bilingual business. There are two separate, individually half-finished entities that also duplicate each other's content.

The company manufactured itself a digital twin, and now the two share the visibility that would have been thin for one.

Brand equity, sadly, does not multiply by cell division. It halves.

A service company spent serious money on Facebook campaigns for years.

Followers, likes, reach. The social report is beautiful.

Then someone asked ChatGPT who it would recommend in that category, and the company was nowhere.

Total bewilderment:

"But we are everywhere!"

No.

You are in one place. And that place is a walled garden.

AI models do not see Facebook's content, do not index it, do not learn from it. Every forint you pushed in there stayed inside the fence.

It is like building on leased land for ten years and then being surprised that the land registry has nothing in your name.

The AI is a student of the open web. What happened in the garden stays in the garden.

A beautiful site.

Animations, elements floating in on scroll, spectacular transitions, portfolio-grade design.

Then you look at what an AI crawler gets out of it — and the vast majority of them still do not run JavaScript today — and the answer is:

an empty div.

No text. No structure. Nothing.

The entire company is a white page.

This is digital haute couture: breathtaking on the runway, and simply non-existent to a machine.

The AI does not sit through the demo. It reads the source. And if there is nothing there, there is not much about the company for it to read out either.

This one is a genre of its own.

Every piece of substantive information — prices, services, technical data — lives in a downloadable PDF. Ideally in a scanned or photographed version.

Meanwhile the website says this much:

"Details in our downloadable material!"

To a machine, a scanned PDF is a heap of pixels. The information is technically there, only locked into a form that answer engines do not, in practice, unpack.

As if you kept the price tags in the safe instead of the shop window.

Secure, no question.

From the customer as well.

The website footer carries one company name. The Google Business Profile a different spelling. Facebook the trade name. And the invoice the full legal name, which nobody outside the company really uses.

A human steps over this without noticing.

The machine, however, thinks in entities, and out of four loosely similar names it does not necessarily assemble one strong company. It assembles four weak ones.

So none of them reaches the level where the AI would recommend it with any confidence.

The first rule of machine-facing brand building is disappointingly boring:

call yourself the same thing everywhere.

If you cannot hold that, the rest hardly matters.

Finally, the absolute baseline.

A hospitality venue's opening hours and menu exist exclusively in a photo uploaded to a Facebook post.

When a tourist asks the AI where to eat nearby, the model can more easily recommend the places whose data also exists in readable form.

Opening hours written onto a photo are not data to a machine.

They are decoration.

At first glance these are ten different mistakes.

In reality they are ten variants of the same problem: the vast majority of businesses build for the human eye while knowing almost nothing about the machine layer.

They spend on advertising, on design, on social media. On spectacular, comfortably invoiceable things.

Meanwhile the layer AI models assemble their answers from is empty, contradictory, or simply closed.

And here comes the unpleasant part.

The AI does not resent this. It does not take offence. It does not punish.

It simply recommends someone else.

The competitor whose data is readable, whose name is the same everywhere, and who — according to their own structured data — does not happen to run a hotel.

Invisibility is therefore not a neutral state. Invisibility means that someone else's name goes into the answer to the question — several billion questions a day.

The good news is that nearly every point on this list is measurable and fixable.

There is no magic in it and no "AI wizardry". It is engineering work: we look at what the machine sees, then we put in order what it sees.

If you are curious which point your company appears at in this post — from experience, at least one of them, probably — an AI visibility audit shows exactly that.

In the meantime, some free homework: ask ChatGPT, Claude or Perplexity about your own category.

If you are not in the answer, you now roughly know which chapter to look for the reason around.