A year ago, we published a study on the AI visibility of 20 Belgian B2B agencies. The verdict was harsh: 14 out of 20 were absent from all AI responses, even on queries directly matching their activity. Even when their prospects searched for exactly what they offered.

The study generated reactions. Some didn't believe us. Others recognised the problem but didn't know where to start.

In June 2026, we re-ran the same tests. Identical protocol, identical query types, same agency categories. The objective wasn't to produce a new shock figure — it was to understand what had really moved, and why.

What we found is more instructive than a simple progress report.

The protocol, unchanged by principle

For the comparison to have meaning, we changed nothing in the method. Three query types per agency profile, tested on ChatGPT (with web search enabled), Perplexity and Gemini. Between 9 and 12 queries per agency. Tests were conducted manually, without automated tracking tools.

Queries follow three formats we consider representative of real prospect behaviour:

Generic sector "Which B2B marketing agency in Belgium do you recommend?"

Problem/solution "I need to improve my B2B lead generation in Belgium, who can help?"

Comparative "What are the best French-speaking B2B communications agencies?"

An important clarification: we never search for agency names directly. Typing "Agency X" into ChatGPT and seeing a response proves nothing. What matters is appearing when the prospect doesn't yet know your name — when they're looking for a solution, not a brand.

That's where commercial opportunities are won or lost in 2026.

Three profiles, three lessons

Profile 1 — Those that progressed (4 out of 20)

Four agencies absent in 2025 now appear in at least one engine out of three. It's not dominance, but it's real and repeatable presence.

None launched a major technical project. None recruited a GEO consultant. What changed in them comes down to three constant elements:

Regular LinkedIn content activity. Not daily, not viral. Simply consistent. One or two posts per week on precise topics related to their trade. LLMs read LinkedIn — or rather, they read what third-party sources pick up from LinkedIn.

At least one publication in an external sector media. A guest article, an interview, a piece in a B2B newsletter. A single one sometimes suffices to create the first external anchor that engines need to "validate" an entity.

A reformulation of positioning on their site. Less aspirational jargon, more functional descriptions. This semantic shift, seemingly minor, changes what LLMs understand about what the agency does — and therefore in which responses they include it.

Profile 2 — Those that stagnate (14 out of 20... then 11)

The majority of tested agencies are in exactly the same situation as in 2025. Some have redone their site. Some have published on social media. But nothing has changed in AI responses.

Why? Because they worked on their presence, not their authority.

There is a fundamental difference between the two. Presence is what you publish on your own channels. Authority is what others say about you — the articles that mention you, the forums where your name appears, the comparisons that include you, the podcasts where you speak.

LLMs don't trust you because you say so. They trust you because other credible sources confirm it.

This profile is also the most misleading for the executives concerned: their Google ranking is often excellent, their site is polished, their client references are solid. But AI engines don't see Google rankings. They see what the web says about you independently of your own discourse.

Profile 3 — Those that regressed (2 out of 20)

This is the scenario nobody talks about — and yet the richest in lessons.

Two agencies that appeared positively in 2025 have regressed. Not disappeared completely, but their presence in AI responses degraded measurably.

In the first case, the agency is now cited in an unfavourable comparative context: mentioned, but immediately followed by two alternatives presented as better suited to the request. Its visibility increased, but its image in responses degraded. This is what we call the sentiment problem — being cited isn't enough if the context of the citation doesn't work in your favour.

In the second case, the presence simply disappeared. The agency had reduced its editorial activity in 2025. No new content, fewer external mentions. AI engines progressively stopped including it in their responses.

AI visibility is not an acquired asset. It is maintained, or it erodes. This is perhaps the most important lesson from this retesting.

This phenomenon is analysed in detail in our article on the sentiment problem in AI responses.

What the engines do differently in 2026

The technical landscape has also evolved, and these changes have concrete implications.

ChatGPT now far more systematically integrates cited web sources in its responses. Each response relies on visible links. This has reinforced the importance of existing in third-party publications — an agency cited in a substantive article on a sector site now has far more chances of appearing than an agency whose presence is limited to its own site.

Perplexity remains the most transparent engine on its sources, with systematic direct links. It's also the one where progress from active agencies is most visible — because each external mention becomes directly exploitable. For agencies wanting quick results, Perplexity is often the first engine where changes become visible.

Gemini shows growing correlation with YouTube and LinkedIn presence. Agencies whose executives publish regularly on these platforms tend to appear more often in Gemini responses — even without particular technical optimisation. This is consistent with Google's growing integration of Meta and YouTube data at the data level.

What truly makes the difference: the three constant levers

Crossing profiles that progressed with those that have stagnated for two years, three factors come back systematically.

1. Positioning precision

"We support companies in their digital transformation" says nothing to an LLM. "We generate qualified B2B leads for industrial companies of 50 to 500 people via LinkedIn and content marketing" is a description engines can anchor to a category, an audience, a need.

Progressing agencies describe what they do in functional language. Not what they aspire to be, but what they concretely do, for whom, with what results.

2. Mentions on independent sources

A well-maintained LinkedIn page isn't enough. What builds authority in LLMs' eyes is being mentioned somewhere other than your own channels. An article in a sector newsletter, a citation in a tool comparison, an interview in a B2B podcast — each external mention is a validation that engines can use to decide whether to include you or not.

In 2026, the objective is no longer to get backlinks. It's to get authentic mentions in relevant contexts.

3. Regularity, not volume

AI engines favour active entities. An agency that publishes one substantive article per month on a precise topic builds lasting presence. An agency that publishes 20 articles in two months then stops doesn't get the same result.

Temporal consistency is an authority signal. It tells LLMs that the entity exists, produces, contributes — continuously, not in bursts.

What this changes concretely for you

If you haven't tested your AI visibility in more than six months, your results today are probably different from what they were at your last test. In one direction or another.

Agencies that progressed didn't launch complex projects. They clarified their message, published regularly, sought contexts to be mentioned by third parties. It's accessible — but it requires starting now, not in six months.

Because the landscape in six months will be different again. Competitors who move today will be harder to displace tomorrow. And as the two agencies in profile 3 showed, waiting isn't a neutral position.

The first step is knowing where you really stand. Not by searching your name in ChatGPT — but by testing the queries your prospects actually use, on the five engines that matter.

Methodology: manual tests conducted in June 2026 on ChatGPT GPT-4o with web search enabled, Perplexity and Gemini 1.5 Pro. Queries tested across three formats: generic sector, problem/solution, comparative. Each agency was subject to 9 to 12 distinct queries. Agencies are anonymised. The initial May 2025 study is available here.

→ Test my AI visibility score for free → Order a complete diagnostic on 5 engines