Anyone looking for an agency, a tool or a service provider today increasingly no longer lands on a list of results, but on a ready-made answer. Assistant systems such as ChatGPT or Perplexity summarise, recommend and name specific providers in the process. This is exactly where a tactic currently doing the rounds comes in: the self-promotional listicle – a list article on your own website that lists the supposedly best providers in an industry and, of course, includes your own company.
An experiment suggests that a single such article can be enough for the systems to include your own offering in their recommendations. The effect therefore apparently sets in much faster than is known from classic search engine optimisation, where rankings are built up over months.
Why the tactic works at all
Language models do not answer questions from a static index, but draw on web content for current topics and condense it. Lists are ideally prepared for this process: clear structure, named candidates, short justifications. A text that already has the form of an answer can be adopted into an answer more easily than running text that contains the same information scattered throughout.
In addition, there is simply little structured material on many niche questions. Where only a handful of sources exist, a new, cleanly structured article carries more weight than in a fiercely contested topic area. That explains why the effect can occur so quickly in isolated cases – but says nothing about how stable it is.
Three reasons that speak against the shortcut
First, credibility. A ranking in which the publisher places itself in one of the top spots is not a comparison, but advertising in comparison form. Human readers see through this in seconds. Anyone perceived on first contact as someone who misappropriates a neutral format for self-promotion starts the sales conversation with a minus.
Second, the half-life. Operators of answer machines have a strong self-interest in their recommendations appearing reliable. Patterns that can be detected automatically – and a self-praise listicle is such a pattern – are, experience shows, devalued sooner or later. The history of search engine optimisation is a series of tactics that worked briefly and afterwards became a burden. Anyone producing dozens of these articles today will tomorrow be sitting on an archive they have to clear out again.
Third, the quality of contacts. Visibility is not an end in itself. A recommendation based on a self-written ranking brings enquiries from people whose expectations were shaped by a text that overstates your own performance. That creates conversations in which expectations have to be corrected first.
What works instead
The observation behind the experiment is still useful – you just have to draw the right lesson from it. It is not: write about yourself in list form. It is: answer machines prefer content that answers a question cleanly and verifiably. A few practical approaches can be derived from this:
- Questions instead of keywords. Formulate headings as the questions customers actually ask, and answer them in the first paragraph below – not only after an introduction.
- Publish what can be evidenced. Process models, decision criteria, project reports with specific framework conditions. Content that only you have because you did the work is harder to copy than a ranking.
- Keep the structure machine-readable. Clear heading hierarchy, short paragraphs, lists where they factually fit, clean structured data in the source code. This helps assistant systems and people alike.
- Conduct comparisons honestly. A comparison of several solutions can be very good content – if it discloses criteria, names disadvantages and also describes cases in which your own offering is not the best choice.
- Build up third-party mentions. How a brand appears in answers also depends on what others write about it – trade media, directories, customer references. These signals are slower, but considerably more robust.
Measuring what changes
Anyone preparing content for answer machines should observe the effect instead of trusting it. A simple approach is practical: define ten to twenty questions with which potential customers realistically enter a conversation with an assistant system. At regular intervals, check which providers and which sources are named, and log this. A look at the server logs is also worthwhile: accesses by the AI providers' crawlers show which pages are being read at all.
The expectation is important here. Answers from language models are not deterministic – the same question can deliver different results. Individual observations are therefore not suitable as proof; trends over several weeks rather more so.
Classification
For web and software projects, this shifts part of the requirements: content must not only look good, but be clearly structured, unambiguously assignable and delivered in a technically clean way – tasks that are decided in the content management system and in the information architecture, not subsequently in marketing.
Short-term tricks can be implemented in just a few hours, but they lose their effect as soon as the providers readjust.Anyone who instead invests in verifiable, well-structured specialist content gains visibility more slowly, but keeps it even when the rules of the answer machines change again.