From zero to 63% visibility in AI in five weeks. A case study of a lime manufacturer

A family-run lime producer from the Łódź Voivodeship increased its visibility in AI models from 31% to 63% between 24 June and 27 July 2026. The number of queries in which the brand is mentioned by name rose from zero to seventeen out of eighteen. The citation index rose from 6% to 50%. Despite this, the brand recognition test in both scans showed the same result: three out of four models recognise the brand, but none recommend it.

We describe this case in its entirety, including the parts that didn’t go to plan, because it is precisely those parts that are the most instructive.

Starting point: a measurement that never took place

The first document was drawn up on 31 May 2026. It was not yet a survey, but rather a preparation for one: the set of questions was still being drawn up, so the indicators stood at zero, as there was nothing to count.

This is the situation most companies start out in. It is not known whether the brand’s models are aware of this, as no one has ever checked. We explain how to do this yourself in the article how to find out what AI says about a company.

We finalised the set with eighteen search queries, covering four areas: building lime complying with the PN-EN 459-1 standard, fertiliser lime and soil deacidification, local search terms including a town and region, and the product’s specific name. The same set appeared in both subsequent scans without any changes, as otherwise the results could not be compared. We explain why this is a necessary condition in the entry frozen query set.

First reading: 31% visibility and zero mentions

The scan from 24 June produced a result which, at first glance, appears contradictory.

Thirty-one per cent visibility with zero mentions means this: the models were happy to answer questions about building and fertiliser lime, providing detailed answers on the subject, but did not link them to that specific brand.

The divergence between the models was striking. Gemini and Mistral addressed the topic in 89%, Perplexity in 72%, whilst Claude, ChatGPT and Grok did not address it at all. The average came to 31%, which in itself doesn’t tell us much. This is a good example of why breakdown by model is more important than a single figure.

Instead of brands, the models listed the major players: the country’s two largest manufacturers, each with over twenty search terms and the top ranking for each. In the fertiliser segment, there were four medium-sized brands.

There was one detail that proved to be crucial. The product’s brand name was recognised by the models in eighteen phrases, but as the category ‘lime’ rather than as a specific company’s product. The brand had a recognisable product, but was unable to associate it with its own name. This phenomenon is addressed in the entry brand aliases.

What has been done over the past five weeks

The report identified four barriers and the same number of actions. All of these relate to the customer, as Aelo does not publish anything on their behalf.

Signals from the standard on its own website. The standards page was given a clearly highlighted statement of compliance with PN-EN 459-1 and a link to the declaration. The phrases referring to the standard were the strongest in the first scan, so emphasising them was the most cost-effective move.

Product price and specifications. A table of specifications for the flagship product and a section on pricing have been added to the building lime webpage. The search term combining the product name, price and manufacturer had the lowest result in the entire set at that time, at twenty per cent.

Structural data. Include ‘Product’ and ‘Organisation’ on key subpages so that the models do not have to make assumptions based on the text alone. See the entry for more details structural data.

Improving the weakest aspects of local communities. The subpages for agricultural lime in Poznań and Warka were labelled „manufacturer” and given an indicative price. Not a single new subpage was created: the brand already had several dozen of them.

This is the conclusion that crops up most frequently in our scans. The problem is rarely the amount of content. Far more often, it is whether the existing content is suitable for quotation, that is, the area described as AEO.

Second survey: 63% and seventeen mentions

A screenshot from 27 July, the same set of eighteen queries, the same models.

The breakdown by model varied even more significantly than the average. Claude went from zero to one hundred per cent, with an average ranking of 2.3. ChatGPT went from zero to one hundred per cent, but with an average ranking of tenth – meaning the brand only appears at the very end of a long list. Gemini 94%. Perplexity 67%, and it was this one that proved to be the weakest link, precisely where the lack of external citations carries the most weight.

The strongest results were seen for sales-related and local search terms: bulk agricultural lime with delivery throughout Poland 72.5%, orders in the city where the company is based 71.67%, certified manufacturer from the Łódź region 71.25%. Product and price-related phrases remained the weakest: product name with price 45%.

What we were unable to change

The brand recognition test for both scans yielded the same result: three out of four models recognise the brand, but none of them would recommend it. Visibility doubled, mentions increased eightfold, but this one aspect remained unchanged.

The reason is stated explicitly in the report: a lack of independent, public editorial content. There are no rankings, comparisons, tests, press articles or external reviews of this particular brand.

This is the limit to which working on one’s own domain leads, and the point at which it stops. We describe this phenomenon separately in the article knows it, but doesn’t recommend it, as it applies to most companies after their first few months of operation.

The domains cited by the models in this category belonged to competitors and to industry websites specialising in lime and aggregates. The brand did not appear on any of them. Until this changes, the recommendation will not appear, regardless of how well the brand’s own website is described. We explain why the models work this way in the article Why does AI recommend the competition?.

Three lessons that can be applied to other companies

Average visibility can hide anything. In June, 31% meant, in practice, that three models did not recognise the brand at all, whilst three recognised the subject without the brand. Without a breakdown by model, this figure is misleading.

A mention and a quotation are two different things, and they grow at different rates. Here, the number of mentions has gone from zero to seventeen, and the number of citations from 6% to 50%, so both metrics have increased, but from completely different starting points. We explain the difference in the entries reference i citation.

Recognition is earned through content; recommendations are earned through the words of others. Five weeks’ work on our own website resulted in an increase in all metrics except one – the most important one. The next stage of this implementation involves activities exclusively outside our own website: expert content for an industry-specific website, applications for certification schemes, and customer reviews on platforms that influencers read.

Methodological caveats

Two measurements are not enough to distinguish the effect of the implementation from the natural variability of the models. The models themselves also changed between June and July, and some of the increase may be due to this, rather than the work carried out on the website. We are therefore not presenting this as a guarantee of results, but merely as a description of a single case.

The set of questions was identical in both scans, which were carried out on the same day for each model, with no brand names mentioned in the questions. A full description of the method can be found in methodology, and an explanation of why we take measurements once a month can be found in the article How often should visibility be measured?.

We publish this data with the client’s consent, without disclosing the company’s name.

Questions