# A machine reads literally

> Cookieless or cookie less? To a person the same word, to a machine two strings. Why you have to declare your own spelling variants to be found.

Picture someone who knows your surname in exactly one spelling. If the doorbell says it differently — hyphenated, without the umlaut, with one space too many — they say: never heard of them. Not out of malice. They simply know that one string of characters.

A machine reads exactly like that. It breaks your page into small units and compares strings — we explain it in [How an AI reads your website](https://btlabs.dev/en/posts/how-an-ai-reads-your-website).

## Two spellings, thirty positions apart.

One space in the query costs the same page thirty positions. For the search “cookieless” our glossary page appears 48 times in the results, at position 32. For “cookie less” — the same word, just with a space — 15 times, at position 63. Both figures come from our own Search Console, as of 6 September 2026.

Same page, same content, same meaning. To a person that is a typo. To a machine they are two different words.

## It happened to us.

Two of our articles wrote “cookielose Statistik”, the inflected German form. The term itself, though, was stored only in its base form: [Cookieless](https://btlabs.dev/en/glossary/cookieless). The machine saw two unconnected things — a text here, a term there.

The connection appeared only once we added the inflected forms. Immediately, and retroactively in every article that uses the word. We did not rewrite a single line. We simply said once how else that word is written.

## Your words have variants too.

Every business has terms that can be written more than one way:

- **The place.** South Tyrol, Südtirol, Alto Adige, Sudtirolo.
- **The service.** Refrigeration systems and cooling technology. Electrical installation and electrician.
- **Your own name.** With Ltd, without, with “&”, with “and”.
- **Abbreviations.** AI and artificial intelligence. NAP and name, address, phone.

A person makes that connection in their head without thinking. A machine does not. It needs the variants written down — by you.

## But not in the body text.

Scattering every spelling through the text is the obvious idea — and the wrong one. It makes the text unreadable and achieves the opposite: term stuffing is the only method studied that pushes visibility in generative answers **below** the baseline. Demonstrated in the [Princeton GEO study](https://arxiv.org/abs/2311.09735) (KDD 2024).

The variants belong where the term is explained once. From there they act on everything that points to it.

That is exactly what a maintained glossary is for. In btlabs Core every term carries its synonyms and inflected forms with it. Add a variant and it connects retroactively to every text it appears in — per language, because German and Italian have different variants.

## The one thing to take away.

A machine does not recognise meaning, it recognises characters. Tell it yourself how else your word is written. Otherwise chance decides whether it finds you.

---
Source: https://btlabs.dev/en/posts/a-machine-reads-literally
Last-Modified: 2026-09-12T07:00:01.149Z
Languages: [de](https://btlabs.dev/llms/de/posts/eine-maschine-liest-woertlich) · [it](https://btlabs.dev/llms/it/posts/una-macchina-legge-alla-lettera)
See also: [llms.txt](https://btlabs.dev/llms.txt) · [ai.txt (Policy)](https://btlabs.dev/ai.txt) · [identity.json](https://btlabs.dev/identity.json)