Fancy text isn't a font. That's why your hashtags stop working.

Paste 𝐡𝐞𝐥𝐥𝐨 into a search box and search for hello. You get nothing.

That surprises people, and it shouldn't. The bold text you copy from a "fancy text generator" is not your text in a bold font. It's a different set of characters that happen to look like your letters. 𝐡 is U+1D421, MATHEMATICAL BOLD SMALL H. It is not an h, any more than the Cyrillic а is an a.

Everything that follows comes from that one fact, and most of it is worse than a failed search.

What these tools actually do

They don't apply a font. They swap each character for a lookalike, mostly out of Unicode's Mathematical Alphanumeric Symbols — a block designed so mathematicians could write a bold vector and a script Lagrangian in plain text.

Which means the available inventory is exactly what a mathematician needed: A–Z, a–z, 0–9, and nothing else.

There is no mathematical bold é. There is no script ł, no gothic ế, no double-struck ş. Nobody was writing equations in Vietnamese. So the moment your text steps outside twenty-six unaccented letters, the lookup table runs out, and the tool has to do something with the leftovers.

The first cost: your name

The accent is not decoration. In Portuguese, avô is your grandfather and avó is your grandmother, and the mark is the only thing between them. Strip it and you haven't made the text ugly — you've swapped someone's grandfather for their grandmother.

I tested eighteen of these tools with the same multi-language string, five styles each, between June and August 2026. Eighty-nine cells. None of them handles an accented letter correctly, and they fail in three distinct ways:

• Eleven leave the accent behind. é stays plain while the letters around it turn styled, so the word ends up a visual mix of two typefaces. The largest group.
• Five decompose it into a base letter plus a detached combining mark. This one may render perfectly on your screen and fall apart on someone else's — the mark's placement is now the renderer's problem, not the text's. It fails invisibly, which is worse than failing.
• Two strip or transliterate: one returns the letters with the diacritics simply removed; one converts Cyrillic to Latin before styling it, which is a different text, not a styled one.

Eleven plus five plus two is eighteen. There is no fourth behaviour, and no exception.

(How I know: sites driven by a scripted browser, every result stored with its evidence. Ten of the eighteen only entered the corpus when I searched in their languages rather than mine — which is itself the finding. A category that works in English is not a category that works. **The 22 sites that appear further down are a separate sample, about editing capability: I keep the two measurements apart on purpose.)

The second cost: nobody warns you about this one

Here is the part that has nothing to do with accents, and that I think matters more.

If 𝐡 is not an h, then #𝐝𝐞𝐬𝐢𝐠𝐧 is not #design. It's a hash sign followed by six mathematical symbols. Instagram will not file it under that hashtag. A styled @mention doesn't notify anyone. A styled URL isn't a URL: in a separate test, on WhatsApp Desktop for Windows, a styled link renders beautifully and simply isn't clickable — one destination, one measurement, not a survey.

Nothing tells you. The text looks right, it pastes fine, and it has quietly stopped doing its job. Your bio is still readable and no longer findable.

This is where the tooling shape matters, and it's the one measurement I'll state exactly as it was taken, because summarising it would overstate it:

No site among the 22 profiled offers multi-line editing with mixed styles on a single line.

I'm being pedantic on purpose — one has a contenteditable area, another a live preview, a third is built like a notes app, and "none of them is a real editor" would be my interpretation rather than my measurement. But none of them lets you style one word inside a paragraph. And if you can only transform a whole line at once, you cannot keep a hashtag plain while styling the words around it. The consequence isn't aesthetic. It's that the format forces you to choose between looking good and being found.

One of them handles it well, and should be said

dobrykalkulator.pl, a Polish generator, fails the same accent test as the rest. But it is the only one of the eighteen that tells you so before you paste:

"In Unicode mathematical alphabets, styled equivalents of Polish letters often do not exist — which is why the generator honestly marks them as partially supported." (my translation)

It labels every style as full or partial Polish support and lets you filter on it. That's a compatibility mode, built by someone who understood the constraint and decided the user deserved to know. I'm naming it because writing "the mechanism nobody explains" while one of them explains it on its own site would be the kind of omission that gets an article dismissed — deservedly.

If you're building one of these

1. Decide what you do with characters outside A–Z before you write the lookup table. It is the only interesting design decision in this category, and it is usually made by accident.
2. Assume your user's name has an accent in it, and test in their language rather than yours.
3. If you can't style something, label it. dobrykalkulator does; it costs nothing.
4. If you transform whole lines, you are silently breaking hashtags, mentions and links. Say so, or let people style one word at a time.
5. Don't sell "we never upload your text" as a feature. I checked: 22 out of 22 do the work locally. It's true of everyone, so it distinguishes no one — the honest version is "no ads, no tracking", which is rarer: 19 of the 22 carry advertising, 20 carry analytics.

Disclosure

I built one of these, so weigh what follows accordingly. It exists because of the study above rather than the other way round.

And it doesn't do the impossible either. By default it decomposes, like the five above: a styled 𝐞 plus a combining accent, which renders well on most screens and badly on some. Its Safe Mode does the opposite and leaves the é intact but plain, like the eleven. Judged by the test in this article, it wouldn't pass either. What it does differently is say so in the app, let you choose, and show you the result on your screen before you copy.

That's all I'll say about it here; the numbers stand on their own and I'll defend each of them in the comments.

Editing capability was profiled on 24 June 2026; accent testing ran from 24 June to 20 August 2026. Sites change. If one of the eighteen has fixed its accent handling since, tell me and I'll say so.

How this was written: with the help of AI (Claude). The survey itself was run with AI-driven browser automation under my direction, every result stored with its evidence, and each number here has been checked against that evidence.