# How a GEO programme runs, from audit to monthly report: Berlitz Canada

> A client example. What we audited for Berlitz Canada, what the audit turned into, what ships, what happens every month, what we track and report, and how we decide whether it worked. With the published baseline: named in 5 of 26 ChatGPT answers, cited in 4.

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

Canonical: https://crawlerpal.com/blog/berlitz-canada-geo-programme-start-to-finish

---

Berlitz Canada asked what most brands ask us first: what do you do, in what order, and how will we know it worked? The answer has four parts. An audit, a ranked plan, the fixes, and a monthly cycle that measures again. Here is each part, with the Canada figures where they are published.

## How does the work start?

It starts with an audit, and the audit starts with questions. We write down what a buyer asks before choosing, without the brand name in the question. For Berlitz Canada that meant 26 questions: the best French school in a given city, preparation for the TEF and TEFAQ exams, language training for a team, French for newcomers.

In May 2026 we ran each one in ChatGPT in a clean session and logged four things per answer: whether Berlitz was named, whether a Berlitz page sat in the sources, who was recommended instead, and which pages the answer leaned on. Berlitz was named in 5 of 26 answers and cited in 4. [The full baseline is in our first field notes](/blog/berlitz-canada-chatgpt-baseline-field-notes).

That first pass used one engine and one run per question. The standard audit is wider: about 20 prompts, four engines (ChatGPT, Perplexity, Claude and Google AI Overviews), three runs each. That is around 240 answers, and it takes five working days per market.

## What does the audit check besides the answers?

Five things. Each one explains a different kind of absence.

| What we check | How | What it tells us |
| --- | --- | --- |
| The answers | The prompt set in clean sessions, named and cited logged in separate columns | Which groups of questions leave the brand out |
| The competitors | Who is recommended instead, and how often | Who holds the answer today |
| The sources | The pages each answer leans on | Which directories, comparisons and threads decide the answer |
| Your own pages | Structured data, FAQ blocks, answer-first structure, whether the page renders without JavaScript, whether AI crawlers are let in | What an engine can read and reuse |
| Your presence elsewhere | Directories, review sites, communities, and your links set against competitors' links | Where you are missing from the pages engines cite |

For Berlitz Canada the answers pointed at two groups. Six questions were about Montreal or Quebec, and Berlitz was named in none of them. On the TEF and TEFAQ questions, the answers went to providers with a programme built for that exam.

## What comes out of the audit?

A ranked plan made of three working lists. You approve it before anything ships.

1.  **A keyword map.** Each query is tied to one target URL and one action. For Berlitz Canada: 49 bottom-of-funnel queries, 26 new articles to write and 23 existing pages to rework.
2.  **A source list.** The directories and comparison pages the engine leaned on, with a note on whether you are listed. Of the first four directories we logged, Berlitz Canada was on one.
3.  **A conversation list.** The forum threads that showed up in the sources, where a useful answer from the brand belongs. For Berlitz Canada, 15 Reddit threads.

Every line is ranked by effort and impact, and an analyst reads the whole plan before it reaches you.

## What do we ship?

Five kinds of work, in the order the plan sets.

1.  **Answer pages.** Comparison articles, buying guides and FAQ blocks with question-and-answer markup, written for the questions that returned nothing.
2.  **Trust markup.** Author, review and organisation schema, so an engine can tell who is speaking and why to believe them.
3.  **Internal links.** A map that ties each new article to the course or product page it supports.
4.  **Source work.** Getting the brand listed, or corrected, on the directories and comparison pages the engines cite.
5.  **Site structure.** One round of input on architecture, where the way the site is built keeps pages out of reach.

Each release is checked by someone who did not build it. For Berlitz Canada the writing runs inside the search programme of Magic Pulpo, our parent studio: two to three articles a week, each in English and French. The September 2026 monthly report, read from Search Console, counts 115 Google clicks in three months to the comparison articles written for AI answers.

## What happens every month?

The month has a fixed shape, because a result you measure once tells you little.

-   **Measure again.** The same prompts, on the same engines, in the same conditions, set against the baseline and against last month.
-   **Refresh.** Pages the engine reads and passes over get rewritten to answer first. Dated pages get current figures.
-   **Publish.** The next answer pages on the keyword map.
-   **Mark up and link.** Schema checked and deployed, new pages linked to the pages they support.
-   **Work the sources.** Outreach to the pages on the source list.
-   **Verify and report.** A release check, a verdict on the results from someone outside the delivery work, then a call to read the report with you.

[The Berlitz UK example goes through the month job by job](/blog/berlitz-uk-monthly-geo-work).

## What do we track and report?

Seven numbers, and we say where each one comes from.

| Number | Source | What it is for |
| --- | --- | --- |
| Named rate on unbranded prompts | Our re-run of the prompt set | The main result |
| Cited rate | The same run, sources logged | Finds pages that are read and passed over |
| Who is named instead | The same run | Shows which competitor holds the answer |
| Listings on cited sources | The source list | Progress away from your own site |
| Clicks and impressions on the pages we wrote | Search Console | The search effect of the same pages |
| Visits from chatgpt.com, perplexity.ai, claude.ai and copilot.microsoft.com | Your analytics | A floor, since few answers are clicked |
| Enquiries from those pages | Your forms or CRM | The link to revenue |

Read the visits row with care: [about one AI answer in a hundred is clicked](/blog/citation-vs-mention-measuring-geo), so referral traffic understates what the answers are doing.

You get one report a month and a call to read it: the prompt table against baseline, what shipped, what did not land, and what comes next. If you already pay for a visibility tool, the figures are reported there. You also get the prompt set, the engines and the dates, so you can run the test yourself.

## How do we decide whether it worked?

By one test. The same unbranded prompts, run again in the same conditions, name the brand more often than at baseline in the groups we worked on. Every other number supports that one or explains it.

For Berlitz Canada the baseline is 5 of 26 named and 4 of 26 cited. The search side is public: 4,140 organic clicks in September 2026, up 60% on September 2025, in [Magic Pulpo's case study](https://magicpulpo.com/works/berlitz-ca-seo-blog). Search clicks are a good sign. They are not the test. The re-test of the 26 prompts is, and we will publish it once Berlitz has signed off the figures.

The rule cuts both ways. In September 2026, adding llms.txt to robots.txt on berlitz.com took crawler hits on that file from 6 in a week to 332, and no visibility number moved. The report said so. [The figures are in the Berlitz UK example](/blog/berlitz-uk-monthly-geo-work).

We do not promise a number, because the engines are third-party systems that change without notice. 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.

**Three client examples.** This page covers the whole sequence. [The Berlitz UK example](/blog/berlitz-uk-monthly-geo-work) goes through the monthly work job by job, 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

Before you sign with anyone for this kind of work, ask for four things in writing: the prompt list, the conditions of the test, the baseline, and the date of the re-test. If one is missing, you cannot check the result.

To see what ours would look like for your market, [book a consultation](https://crawlerpal.com/#book). Prices are in [the pricing section](https://crawlerpal.com/#pricing).

## Frequently asked questions

### What do you need from us to start?

Read access to Search Console and analytics, one person who can approve the prompt list and the pages, and your proof points: pass rates, client numbers, testimonials you are allowed to use. We also ask early whether your brand may appear in comparison articles next to competitors, because those are the pages engines recommend from.

### Is the audit a separate purchase?

No. It is the first phase of the Implementation Sprint, which is a fixed one-time fee per market. The free AI Visibility Score is something else: an estimate from five answers in a chat, useful as a first look. The audit is the measured version, with the prompts, the engines and the dates written down.

### Do you report in a dashboard?

No. We do not sell a dashboard. If you already use a visibility tool such as Profound or Peec, the before and after figures are reported there. If you do not, you get the monthly report and the sheet behind it, with every prompt, engine and date, so the numbers can be checked without us.

### Why was the Berlitz Canada baseline run on ChatGPT only?

It was a first measurement, taken in May 2026 to decide where to work. One run on one engine was enough to show which groups of questions returned nothing. It is too thin to call a market share. The standard audit samples four engines, three runs each.

---

CrawlerPal · https://crawlerpal.com
