More and more people are no longer putting their research questions to a list of search engine results, but to an AI assistant.Instead of ten blue links, they get back a finished paragraph: which software is suitable for which purpose, which provider is worth recommending, how to assess a political position.Two studies currently being reported on suggest that these answers are considerably less stable than their authoritative tone would have you believe.
Software recommendations: the source decides, not the model
A study examined which sources the answer service Perplexity draws on when it recommends software.Perplexity is one of the so-called answer engines – systems that research live on the web, summarise the pages they find and output the result with source references.According to the study report, the analysis paints a picture of targeted manipulation: the pages drawn on did not come about by chance, but to a considerable extent are designed to appear in precisely such answers.
This is not a flaw in the model, but a consequence of its architecture.A language model that ties its statements to current web content is only as good as the selection of pages it reads at that moment.And the market for software comparisons has for years been one of the most commercially dense areas on the web: comparison portals, affiliate rankings, listicles with the same structure every time, guest articles with purchased placement.Anyone who serves this genre well gets cited – regardless of whether the assessment behind it is robust.
For users, this is hard to recognise.The AI answer smooths over differences in source quality.A paid ranking and a carefully researched specialist article appear in the result text in the same sober wording, and experience shows that the appended source links are rarely opened.The classic search results list had an underestimated advantage here: it left the assessment of the source to the human.
Political questions: the same mechanism, higher stakes
The second study comes from AlgorithmWatch and evaluated around 4,500 search queries in the context of upcoming state elections – specifically the AI overviews that Google now displays above the regular search results.The organisation's conclusion: anyone wanting to rely on these summaries for their voting decision should be cautious.
The second case is thematically far removed from the first, but technically close.In both cases, a question is translated into a handful of web sources, from which a model shapes a single, seemingly authoritative answer.And in both cases, the quality of the result depends on which content was well findable, well structured and appropriately worded on the topic at the time of the query.Wherever there is something to be gained – market share or votes – an incentive arises to put precisely such content into circulation.
What this means for your own visibility
The obvious reaction of many companies is to now do exactly what the study criticises: to produce content aimed primarily at being cited by answer systems.The term Generative Engine Optimization has become established for this – optimisation for AI answers instead of search results lists.The short-term effect can certainly occur.The medium-term price is foreseeable: search operators have in the past responded to every wave of industrially generated optimisation content with countermeasures, and content without intrinsic value rarely survives such adjustments.
A sober stocktake in two directions makes more sense.
Outward, with a view to your own findability:
- Check what AI assistants currently output about your own company, your own products and your own market – and which sources they cite for it.
- Clarify whether these sources even come from your own environment or whether third parties are shaping the narrative.
- Prepare your own content so that it can be captured cleanly by machines: clear structure, unambiguous terms, substantiated statements, a well-maintained technical foundation.That helps even when the systems change again.
Inward, with a view to your own procurement:
- Treat AI answers on software and service provider selection as a starting point, not as a basis for decisions.
- When selecting tools, continue to rely on traceable criteria: references, technical review, test installation, conversations with users.
- Make it clear within the team that a fluently worded recommendation is not evidence – especially in areas where no one has the depth of expertise to assess the result.
Assessment: two studies, one pattern
Both analyses report on different systems and different subject areas, and both are limited in their scope: one looks at a single provider in one market segment, the other at one use case on a specific political occasion.Individual findings cannot be generalised from them.What is emerging is more of a pattern: answer systems pass on the weaknesses of their source selection to users without making them visible – and where the selection can be influenced, it is influenced.
For web and software projects, this means two things.First, visibility work is shifting from pure placement in results lists to the question of whether your own content is machine-understandable and substantively robust enough to be cited – substance and clean structure are the same tool here.Second, every procurement decision needs a review step that does not consist of an AI answer.
Anyone who takes both seriously gains visibility without becoming dependent on a channel whose rules will continue to change.