BazofIA: when AI turns marketing into trusted advice

You are about to spend nearly €1,700 on an artificial intelligence course. Before paying, you ask a chatbot a simple question: “Is this a good school?” It responds with a convincing summary: practical programme, satisfied students, university partner, certificate and experienced instructors. It may even supply links. What you cannot tell from the confident prose is whether the system checked the claims or merely rearranged what the seller wanted you to read.
Learning Heroes advertises IA Heroes Pro with the phrase “certificado oficial”, or “official certificate”, and a currently displayed promotional price of €1,697. Western Europe University lists the same programme as a 60-ECTS título propio, a university-issued qualification distinct from an official Spanish university degree. Such qualifications may have professional value, but anyone seeking an officially recognised university degree needs to know exactly what they will receive. A responsible recommendation would reconcile the sales page with the academic classification.
The chatbot that learned to recommend
Businesses that once fought for the first page of Google now want to appear in the answers to “What should I buy?” and “Where should I study?” This is known as generative engine optimisation, or GEO. A 2024 KDD study measured visibility gains of up to 40% in its experimental benchmark, depending on the method and query.
Helping a search system find genuine facts can benefit readers. Manufacturing the appearance of independent authority is different. A school calls itself excellent; a commercial blog repeats the assertion; a directory cites the same wording; another site collects those mentions. A chatbot retrieving the pages may mistake repetition for corroboration. Three pages do not necessarily amount to three independent sources.
- 1 · A school publishes: “Excellent programme, official certificate, satisfied students”.
Source: seller-controlled material. - 2 · Websites repeat: blogs and directories echo or link to the claims.
Risk: multiple mentions may come from one original source. - 3 · An AI summarises: combines retrieved results, may favour familiar brands and overlook discrepancies.
Risks tested in SafeGEO and brand-bias audits. - 4 · A buyer hears: “This school is recommended”.
Checks needed: exact qualification, awarding body, total fees and verified outcomes.
Illustrative infographic. A hypothetical risk pathway, not an allegation about a particular school’s conduct. Source: SafeGEO (EMNLP 2026).
An experiment that made flawed products look recommendable
In October 2026, Qianfeng Wen and six colleagues published SafeGEO at EMNLP. They assembled 600 recommendation cases and 22 ways of manipulating material that agents might retrieve. In some conditions, the chance that a flawed target product entered the recommendations increased by as much as 83.2 percentage points. Defensive techniques reduced the effect but did not fully remove it.
The attack operated on information read by the agent, not the model’s internal parameters. The final answer could sound impartial while reflecting material prepared to favour an inferior option. This demonstrates a vulnerability in controlled experiments; it is not evidence that Learning Heroes, or another particular school, has run such a campaign.
Familiar names have a head start
Manipulation need not be deliberate. In a November 2026 preprint, Giorgio Monaco, Jacopo Lenti and Francesco Pierri studied whether generative search systems favoured familiar brands. Dominant brands in their sample appeared 1.4 to 2.8 times as often as the average brand. In controlled comparisons, replacing a known name with an unfamiliar one while keeping product attributes unchanged reduced ranking performance. The preprint has not yet undergone peer review.
A school can be well known because it advertises heavily, has a large community or produces vast amounts of content. Visibility is measurable. Teaching quality is a different question. An assistant that defaults to popular names risks recommending fame instead of independently assessed learning outcomes.
What the reader sees, and what the AI actually read
Links alone are no guarantee of a properly checked recommendation. Research published in Data & Policy examined nearly 14,000 search-assisted answers using 2025 logs. It found gaps between relevant pages retrieved and the sources ultimately credited, with considerable variation across systems. This is a study about attribution, not an assessment of schools. It helps explain why displaying citations is not the same as auditing a recommendation.
The distinction students need before paying
IA Heroes Pro promotes lessons, tutorials, projects and community access. Western Europe University categorises the programme as a título propio. Students might value its teaching and still expect exact information about its certificate. The term “official certificate” should be accompanied by an unambiguous account of who awards it and its legal academic recognition.
Judging a school would require more: assessed work, real teaching time, instructors, contract terms, completion rates and evidence of learning or employment outcomes collected using a reliable method. An authentic testimonial can describe one graduate’s experience. It cannot stand in for results from a representative cohort.
Iberestudios has previously investigated Learning Heroes’ marketing. This report asks another question: what happens when AI transforms commercial claims into advice? The research cited here does not establish that the company deliberately manipulated AI agents. The documented case is a difference in the academic terminology used by two primary sources.
The tired customer and the effortless answer
After a long working day, dozens of emails and a traffic jam, having a chatbot reduce ten schools to three names sounds attractive. Research associates unpredictable and congested commutes with stress, while environmental noise can affect sleep and cognitive health. The WHO and ILO have also documented health risks associated with very long working hours. See the commuting review, WHO noise guidance and WHO/ILO findings.
Whether tired consumers scrutinise a particular chatbot recommendation less closely is a separate hypothesis; those studies do not establish that causal link. Music research offers a useful comparison: Milliman’s 1982 field study found that background tempo could alter shopping pace and sales. It did not show that a particular musical frequency disables critical thinking.
When a choice involves money and months of your life, the important checks remain important even when you are short of time. AI can reduce the work of finding evidence. It should not replace the evidence.
How to check a chatbot’s recommendation
Ask it to distinguish the school’s own claims from documents issued by a university or regulator. Ask for the exact qualification, awarding institution and total cost. Compare alternatives using the same criteria and insist on seeing what is missing. If an answer repeats “leading”, “official” or “transformative” without independent documentation, it has not yet helped you choose.
The best use of AI will not be the most confident recommendation. It will be an answer that shows where its evidence comes from and what might change the conclusion. The seller wants you to stop comparing. You have every reason to keep looking.
Featured photograph: Christin Hume / Unsplash. Illustration of someone reading information online.
Leer este reportaje en español.
Sources and documentation
- Wen et al., SafeGEO, EMNLP 2026
- Aggarwal et al., GEO, KDD 2024
- Monaco, Lenti & Pierri, brand bias, 2026 preprint
- Strauss et al., The attribution crisis, Data & Policy 2026
- Learning Heroes, IA Heroes Pro sales page
- Western Europe University, specialised programmes
- Review of commuting and wellbeing
- Milliman, music and shopper behaviour, 1982