How long did it take to go from 38 to 93?

Eight months, though not evenly spread. The first large jump came in the first six weeks: access and self-description are one-off pieces of work, so the score rises quickly once they are done. The rest was won page by page, and since each page has to own its own question there is no shortcut. Running four languages at once was what lengthened the calendar.

Why is the score not 100?

A 100 means every check passes at its strictest threshold, which is not a meaningful target for every page. The ratio of question-shaped headings, for instance, is naturally lower on a case study than on a guide — that page's job is to describe work, not to answer a query. The score is a diagnostic showing which layer is thin, not a target.

← Work

GEO case

GEO build for a four-language software site

Services: GEO
Sector: Software and AI
Year: 2026

GEO readiness score

Baseline

93/ 100

Last measured

The last measurement was taken with Galata Media's public audit tool: because the tool reads only publicly available files, anyone running the same audit on the same domain gets the same result. The baseline was recorded as the engagement opened and measures the site as it stood that day — it cannot be re-run after the fact.

Last measured: September 15, 2026
Baseline: February 4, 2026

+55 points

Short answer

How does a software company get cited in generative search engines?

The GEO engagement Galata Media ran for a software company moved the site from a recorded baseline of 38 out of 100 to a measured 93. The gain came from rebuilding existing content so engines could read it, not from writing more: in all four languages, every page took ownership of one real search question, gained an answer that survives being lifted out of context, and bound every number to a verifiable source.

Key takeaways

A GEO readiness score recorded at 38 measured 93 after eight months — and the 55-point gain came from restructuring the content rather than rewriting it.
The 93 is measured from public files; a reader can run the same audit on the same domain and confirm it.
Writing all four languages by hand kept machine-translated, context-free sentences out of the answer passages.
Because access and self-description were finished before content, every page written afterwards was readable the moment it shipped.

Where it started

The site was complete for a human reader and close to invisible to an engine.

The baseline recorded in February 2026 was 38 out of 100. The loss was not a technical fault: the site was fast, it worked, and a visitor found what they came for. The problem is that an engine producing an answer does not read a page end to end. It splits the page into passages, scores each one against the query on its own, and carries only the strongest into its answer — and on this site no passage stood up alone.

No page owned a single question; all four languages repeated the same general description.
There was no answer sentence that still said what it was about once lifted out of context.
The site did not introduce itself: who it is, what it sells and where it is were not where an engine looks.
Not one number on the site was bound to a source — an engine skips the page it cannot confirm.

What was done

Every page took one question

The questions were chosen in the wording people actually type, verified rather than invented, and no two pages were given the same one. Two pages targeting a single question cancel each other out: rather than pick between them, an engine passes over both.

Every page gained an answer that stands alone

Answers were written so that, pulled off the page and pasted into a conversation, they still say what they are about. A sentence that leans on "we" or "this service" instead of naming its subject loses its referent the moment it is quoted.

The site introduced itself

Who the company is, what it does and where it sits were written where an engine looks — and written identically everywhere else on the site. A site that describes the same organisation two different ways stops being one organisation as far as an engine is concerned.

Every number was bound to a source

Claims that could not be verified were not softened but removed outright. An engine skips the page it cannot confirm rather than quoting around it — so an unsupported number costs the whole page it sits on.

Four languages were written separately

The locales were not left to translation. Translation carries a language-specific example across verbatim, which makes the sentence wrong for that language's reader — and a wrong sentence does not get quoted. Each locale was written with its own examples.

City pages were not multiplied from a template

Each city page was written one at a time, with the similarity between them measured and tracked. Repeating one text hundreds of times with the city name swapped makes the pages indistinguishable, and an engine then picks none of them.

What the engagement covered

Content architecture rebuilt across four languages (TR, EN, DE, AR)
One question owned per page, with a quotable answer passage
The layer through which the site introduces itself to engines
A source registry, and removal of claims that could not be verified
City pages written one at a time
Per-engine citation tracking and a monthly re-measurement cadence

Result

Now cited for

software company istanbul
ai integration services
custom software development

Engine

ChatGPT
Google AI Overviews
Perplexity
Claude

On GEO

Why a GEO score climbs in steps rather than all at once

Visibility in generative search is built in three layers, and the layers work in order. The first is access: whether an engine is permitted to read the page at all. The second is legibility: whether the page says on its own what it is, who published it and which question it answers. The third is evidence: whether every claim on it has something verifiable behind it. Investing in an upper layer while a lower one is missing is painting the far side of a closed door.

That is why the score climbs in steps rather than along a line. Once access and legibility are in place there is one large jump; after that, every page that ships uses the same foundation and moves the number a little at a time. The movement in the first month is usually larger than the next six months combined — which means the work started in the right order, not that it finished.

Why visibility on a four-language site is not divided by four

An engine answers in the language it was asked in, and looks for its sources mostly in that language too. So a four-language site enters four races rather than one: being cited in Turkish wins nothing on the same question asked in German. The layer through which the site introduces itself, by contrast, is shared across the languages and is built only once.

In practice that means the cost does not quadruple: there is a fixed base, and only the writing multiplies. With one condition — that the writing is not translated. Translation carries a language-specific example across verbatim, which makes the sentence wrong for that language's reader, and a wrong sentence does not get quoted. When all four locales are written with their own examples, a single base feeds four races at once.

When does an answer get quoted?

An engine does not read a page through and summarise it. It splits the page into passages, scores each against the query on its own, and carries only the strongest into its answer. The consequence is direct: what gets quoted is not the page but a single paragraph. If that paragraph does not make sense alone, the rest of the page never comes into play, however good it is.

A paragraph that makes sense alone has two properties. It names its subject outright — a sentence built on "we", "this service" or "as described above" loses what it is about the moment it is lifted out. And it finishes one question; a paragraph that turns to a second subject halfway through comes second on both.

Frequently asked

How long did it take to go from 38 to 93?: Eight months, though not evenly spread. The first large jump came in the first six weeks: access and self-description are one-off pieces of work, so the score rises quickly once they are done. The rest was won page by page, and since each page has to own its own question there is no shortcut. Running four languages at once was what lengthened the calendar.
Why is the score not 100?: A 100 means every check passes at its strictest threshold, which is not a meaningful target for every page. The ratio of question-shaped headings, for instance, is naturally lower on a case study than on a guide — that page's job is to describe work, not to answer a query. The score is a diagnostic showing which layer is thin, not a target.

Queries this page answers

geo case study
generative engine optimization example
ai search optimization

You next

Let us run the same measurement on your site.

Start by seeing your score. The audit takes about ten seconds, asks for no sign-up and reads only your site's public files — the same measurement this page opens on.

Measure your score

Free, no sign-up, ten seconds.