# What happens every month on a GEO programme: Berlitz UK

> A client example. After the audit and the first fixes, the work becomes a monthly routine: measure again, refresh, publish, mark up, link, work the sources, then report. What each job is for Berlitz UK, what we check before release, what goes in the report, and what counts as success. With three sets of figures: 12 buyer questions put to one answer engine (Berlitz named in 1), a check of seven public pages, and the berlitz.com llms.txt experiment where 6 crawler hits a week became 332 and no visibility number moved.

**Mark Bodnar**, Founder, CrawlerPal · November 18, 2026 · Field notes

Canonical: https://crawlerpal.com/blog/berlitz-uk-monthly-geo-work

---

After the audit and the first fixes, a GEO programme becomes a monthly routine. For Berlitz UK that routine has six jobs: measure again, refresh, publish, mark up, link, and work the sources. Then one report. Here is what each job is, what we check, and what counts as success.

Berlitz UK teaches languages in centres such as London and Manchester, online, and to companies. It is one of the four clients that have paid for this work, and its programme follows the same sequence as [Berlitz Canada](/blog/berlitz-canada-geo-programme-start-to-finish): baseline, keyword map, comparison content, source outreach, re-test. The programme's own baseline is not published yet. The figures on this page come from three checks made for it: 12 buyer questions put to one answer engine and a read of seven public pages, both on 17 November 2026, and an experiment on berlitz.com, the domain Berlitz UK shares with every other Berlitz market. Together they show the monthly work in practice.

## Why is there monthly work at all?

Because an answer is rebuilt every time someone asks. The pages it leans on change when a directory updates its list or a competitor publishes a comparison. And [roughly 86% of cited domains are unique to one engine](/blog/sources-ai-engines-trust), so four engines mean four moving sets of sources. A fix that held in March can be gone by June.

We have not found a reliable published figure for how fast cited sources turn over. That is one reason we measure it per client, on a fixed schedule, and do not assume.

## What do we do each month?

Six jobs. Each one leaves something you can open and check.

| Job | What it is | What you can check |
| --- | --- | --- |
| 1. Measure again | The same prompt set, on the same four engines, in clean sessions, three runs each | The updated table: named, cited, who was named instead |
| 2. Refresh | Pages an engine reads and passes over are rewritten to answer in the first lines. Dated pages get current figures | The list of pages changed |
| 3. Publish | The next comparison pages and buying guides from the plan | The new URLs |
| 4. Mark up | FAQ, author, review and organisation schema checked, corrected and deployed | The validation result |
| 5. Link | New pages tied to the course pages they support | The updated link map |
| 6. Work the sources | Outreach to the directories and comparison pages the engines cite | Listings won, pending and refused |

## What did a first pass on the answers show?

For this page we ran job 1 once, on one engine. On 17 November 2026 we put 12 questions a UK buyer might ask to Claude, with web search on and the location set to London, three for each of four moments. They included "best business English courses in London", "where can I learn German in Manchester" and "which provider should a UK company use for corporate language training for employees". None contains the brand name, because [a prompt with your name in it flatters the score](/blog/branded-vs-unbranded-ai-visibility).

| Type of question | Questions | Berlitz named | berlitz.com among the pages retrieved |
| --- | --- | --- | --- |
| Category discovery | 3 | 0 | 0 |
| Comparison | 3 | 0 | 0 |
| Recommendation | 3 | 1 | 0 |
| Problem | 3 | 0 | 0 |
| All questions | 12 | 1 | 0 |

Source: CrawlerPal first pass, Claude Opus 5.5 through the API with web search on, one run per question, 17 November 2026.

Berlitz was named once, first on the list, for corporate language training. It was not named in the other eleven answers, and no berlitz.com page was among the pages the engine retrieved for any of the twelve. The names that came up most were two cultural institutes, a tutoring marketplace and two London schools.

Read it with its limits. This is one engine and one run, and three of the twelve answers were given without a search. The audit proper samples four engines, three runs each. It is still enough to show where the monthly lists would start: the questions about schools and courses in London and Manchester.

## What did a check of the UK pages show?

On 17 November 2026 we read seven public pages on berlitz.com/en-gb the way a crawler does, from the page source and without running JavaScript: the home page, the school locations page, the London and Manchester school pages, group classes, online courses and business English.

| Check | Result on 7 pages |
| --- | --- |
| Main text readable without JavaScript | 7 of 7 (598 to 1,905 words per page) |
| AI crawlers allowed in robots.txt | Yes. One set of rules for every crawler, none blocked |
| Structured data in the page source | 1 of 7 (the London school page, FAQ markup) |
| Headings phrased as a question | 10 across the seven pages, 9 of them on the London page |
| llms.txt and /ai-instructions | Both exist at domain level |

Source: CrawlerPal check of seven public pages on berlitz.com/en-gb, page source only, 17 November 2026. Markup added later by JavaScript would not show in this check.

Access is fine. What differs from page to page is how they are built. The London school page has questions as headings and FAQ markup, which is the form an answer engine can reuse. The Manchester page has neither. Lines like that go straight onto the refresh and mark-up lists for the month.

## What do we check before anything goes live?

Every change passes a release check by someone who did not make it. The page has to render for a crawler, the markup has to validate, and the page has to answer the question it was written for.

The same separation applies to results. The person who gives the verdict on the month is not the person who did the work.

## What do we track?

The main number is the share of unbranded prompts where the brand is named, per engine and per group of questions. Next to it we log whether one of your pages was cited, who was named instead, and which sources the answer used.

From your side we read Search Console for the pages we wrote, analytics for visits from the answer engines, and your forms or CRM for enquiries. [The full list of seven numbers is in the Berlitz Canada example](/blog/berlitz-canada-geo-programme-start-to-finish). If Magic Pulpo, our parent studio, also runs your search programme, the same report carries the search figures, so you read one document.

## What is in the monthly report?

-   The prompt table, against the baseline and against last month.
-   What shipped, with the URLs.
-   What did not land, and what we think the reason is.
-   The source list: listed, pending, refused.
-   Search figures for the pages we wrote.
-   Next month's list, in order.

Then a call to read it together. If you already use a visibility tool, the figures go there. You also keep the prompt set, the engines and the dates, so anyone on your team can repeat the test.

## What does a result that did not land look like?

In September 2026 an experiment ran on berlitz.com. On 10 September the llms.txt file was added to robots.txt, to see whether pointing crawlers at it would change anything. The crawlers came. Nothing else moved.

| Measure | Week before (3 to 9 Sep) | Week after (10 to 16 Sep) |
| --- | --- | --- |
| Crawler hits on /llms.txt | 6 | 332 |
| Indexing signal | 6 | 187 |
| Training signal | 6 | 318 |
| Citations, as the tool counts them | 0 | 0 |

Source: Agent Analytics report for www.berlitz.com, 13 August to 16 September 2026.

Over the full five weeks, /llms.txt had 338 bot visits and 0 citations as the tool counts them. [An ordinary page of facts](/blog/how-to-write-an-ai-instructions-page) on the same site, /ai-instructions, had 215 bot visits and 73. The crawlers of Google, Meta, Perplexity and Copilot never requested llms.txt, although nothing blocked them. Berlitz's visibility score in its category stayed flat through the period, between 0.173 and 0.184, ranked seventh or eighth.

So the verdict in the report was plain: crawl volume rose 55-fold and no result changed. Even the 73 needs care. [The tool's citation count comes from crawler logs](/blog/ai-visibility-dashboard-counting-wrong-thing). It records a bot fetching the page, which is different from an answer in which a buyer saw Berlitz named. This is why the monthly test stays the prompt set, and why [we do not count an llms.txt file as a fix](/blog/llms-txt-almost-nobody-reads-it).

## How do we measure success?

By the same test every month. The same prompts, in the same conditions, name the brand more often than at baseline in the groups we worked on. Being cited without being named is a lead to follow: the engine read your page and recommended someone else.

We do not promise a number or a date. If the prompts we worked on have not moved after 90 days, we re-test at no extra cost, tell you which fixes landed, and rework the ones that did not. The monthly work is optional and can be stopped at the end of any quarter. The pages, the markup and the prompt set stay yours.

**Three client examples.** This page covers the monthly work. [The Berlitz Canada example](/blog/berlitz-canada-geo-programme-start-to-finish) covers the whole sequence with published figures, and [the International Academy of Football example](/blog/international-academy-of-football-ai-visibility-audit) goes through the audit step by step.

## What this means for you

If you are comparing monthly proposals, ask each agency two questions. What is measured again every month? And what are we allowed to run ourselves? A fee that buys articles and a traffic chart cannot tell you whether an answer engine names you.

To go through the monthly cycle for your market, [book a consultation](https://crawlerpal.com/#book).

## Frequently asked questions

### Is the monthly work only for clients who had a Sprint first?

Yes. The monthly cycle continues an Implementation Sprint, because measuring again needs a baseline to compare with and a plan to work from. The Sprint is a fixed one-time fee per market. The monthly work is optional, and you can stop it at the end of any quarter.

### Can we see the raw answers?

Yes. You get the prompt set, the engines, the dates and the logged answers. Run the same prompts in a clean session and you will see what we saw, within the normal variation of systems that do not answer the same way twice.

### Why three runs per prompt?

Because an engine does not give the same answer twice. One run can name a brand that the next two leave out. Three runs on each engine dampen that variance enough to compare one month with the next, which a single run cannot do.

### What happens when a month shows a drop?

The report says so. We look at which sources changed in the answers that dropped and whether a competitor published something new, and the fix goes on next month's list. If the drop is on prompts we worked on, we explain it and rework the pages.

---

CrawlerPal · https://crawlerpal.com
