What if a brand’s success in ChatGPT answers did not depend on what a company actually delivers – but on how its name sounds? That is exactly what a recent study by Peec AI suggested: A company name that is perceived 11 percent more positively by people already leads to a 0.8 percent more positive description by language models. Regardless of how well the company actually performs.
This phenomenon has a name: semantic drift – a shift in reputation caused by an AI’s tendency to interpret context only approximately rather than understand it precisely. It does not show up in clicks or rankings, but in small, barely noticeable shifts in what an AI believes it knows about a brand.
For the rare extreme case – an AI answer that portrays a brand completely incorrectly or even as the exact opposite of what it is – the term synthetic crisis can be used: a reputational crisis caused by an AI’s inability to interpret context correctly. Such cases are loud, visible, and quickly noticed. We recently experienced an example of such a synthetic crisis ourselves; more on that here.
Semantic drift is the far more common, quieter version of the same underlying problem. It does not show up in citation rates or visibility charts. It only becomes visible when someone actually reads what an AI says instead of merely counting how often it says it. That is precisely why this shift in meaning has a much greater impact on the everyday reality of most brands.
Since Generative Engine Optimization (GEO) has become a standard part of many marketing strategies, three metrics in particular have dominated the discussion: how often a brand appears in AI-generated answers, which sources are cited, and how positive or negative the tone is. All three are blind to exactly the kind of shift at issue here.
Frequency depends heavily on which questions a monitoring tool tests in the first place – usually assumptions made by a marketing or SEO team, rather than what people actually type in. The cited sources, in turn, show only a fraction of what truly shapes an answer. A much larger portion comes from general training material that never appears as a visible source, but can still fuel semantic drift.
And the sentiment score – as illustrated by the name study above – can already be distorted by factors that have nothing to do with a company’s actual performance. Tone can be measured. Whether a statement is factually correct is an entirely different question.
One of the largest independent studies on the subject to date shows that this is far from a marginal phenomenon. Together with the BBC, the European Broadcasting Union had more than 3,000 answers from ChatGPT, Copilot, Gemini, and Perplexity evaluated by specialist journalists across 18 countries and 14 languages.
The result: 45 percent of all answers contained at least one significant comprehension issue. 31 percent showed serious sourcing problems – missing, misleading, or incorrectly attributed sources. 20 percent contained major accuracy errors such as outdated information or fabricated details.
One aspect is overemphasized, another is omitted entirely, while a long-outdated piece of information persists far longer than current communications would suggest. None of this shows up in a visibility dashboard.
An experiment involving our own agency makes this shift tangible. At our request, Claude, Gemini, and ChatGPT each produced a comparison between us and a comparable agency. The results contradicted one another fundamentally: Claude described us as “factual and strategic,” while Gemini called us “creative, narrative, and emotionally driven” – completely opposing assessments of the same brand, supplied without being asked.
When it came to classifying our target audience, several models identified a positioning that was indeed accurate when the agency was founded in 2009: a focus on young companies. We deliberately moved away from that positioning ten years ago – no blog post, press release, or social media post still mentions it. Reputational history does not disappear through silence alone – it requires active counter-signals over years.
That is why it is not enough to simply count how often a brand appears. It is worth looking more closely at which nuances in a brand’s positioning shift over time: Which terms suddenly become associated with a brand even though they were never communicated that way? Which terms are missing even though they should be central? Companies that ask these questions regularly can identify semantic drift before it becomes entrenched.
The more important step comes next: defining the positioning – across owned channels as well as through third parties – much more sharply than most companies do today. Every single word deserves scrutiny: Does this term really create the intended association? And is it substantiated, or merely claimed?
Three approaches have proven effective:Drei Ansatzpunkte haben sich als wirksam erwiesen:
Focus instead of breadth. Companies that position themselves too broadly in thematic terms – including through content outside their core business that is primarily intended to generate additional organic traffic – increase the risk of being semantically misclassified. Every additional thematic association creates a potential attribution risk.
Visibility of what is real. Authentic insights on websites, social media, and especially review platforms are harder to distort than pure marketing claims and are heavily used by language models.
Earned Content. Corroborating data points from independent third parties – editorial teams, specialist media, and platforms with no vested interest – carry more weight than a company’s own claims.
“Reputation is no longer a quantitative question, but a semantic one. Anyone who only looks at rising bars in monitoring dashboards misses the actual risk: not whether a brand is mentioned often enough, but whether what is being said about it is still accurate at all. Reputation is infrastructure – not a one-off campaign for optimizing visibility.”
Miriam Schwellnus, Expert for GEO, PR, and Brand Storytelling at Mashup Communications
A shift in reputation caused by an AI’s tendency to interpret context only approximately. It does not become visible in metrics, but only when AI-generated answers are actually read – and it affects virtually every brand with a digital history.
A synthetic crisis is the rare extreme case: an AI answer that portrays a brand completely incorrectly or even as the opposite of what it is. Semantic drift is the far more common, quieter version – not outright misinformation, but a gradual loss of precision in tone and attribution.
In a study of more than 3,000 AI-generated answers across 18 countries and 14 languages, 45 percent contained at least one significant comprehension issue, 31 percent showed sourcing problems, and 20 percent contained major accuracy errors – often in the form of exactly the kind of imprecision that characterizes semantic drift.
A large part of what shapes AI-generated answers comes from general training material that cannot be corrected in a targeted way. A more effective approach is the continuous development, over years, of both owned content and content validated by independent third parties.
Outrage has long been more than just a risk for brands: it can become a…
Discover the key aspects of successful collaboration with PR agencies, the use of storytelling and…
Social Media Storytelling in Football: How Professional and Amateur Clubs Grow Their Audience and Build…
In pop culture, artificial intelligences are always portrayed as villains. By emphasizing the human-machine interface…
As a PR and brand storytelling agency, we help companies find their role as mentors…
Nice for today, but not for the next five years on my profile: Instagram Stories…