Bias in Search: Visibility, Perception, and Control

You can’t eliminate bias in search, but you can manage its impact on your brand and reputation.

Bias in search isn’t always negative. It’s easy to frame it as something sinister, but bias shows up for structural reasons, behavioral reasons, and sometimes as a deliberate choice. The real task for marketers and communicators is recognizing when it’s happening, and what that means for visibility, perception, and control.

Two recent pieces got me thinking more deeply about this. The first is Dejan’s exploration of Selection Rate (SR), which highlights how AI systems favor certain sources over others. The second is Bill Hartzer’s upcoming book Brands on the Ballot, which introduces the concept of non-neutral branding in today’s polarized market. Put together, these show how bias isn’t just baked into algorithms, it’s also unavoidable in how brands are interpreted by audiences.

Selection Rate and Primary Bias

Selection Rate can be thought of as the percentage of times a source is chosen out of the available options (selections ÷ options × 100). It’s not a formal standard, but a useful way to illustrate primary bias in AI retrieval. Dejan points out that when an AI system is asked a question, it often pulls from multiple grounding sources. But not all sources are selected equally. Over time, some get picked again and again, while others barely show up.

That’s primary bias at work.

For marketers, the implication is clear: if your content is rarely chosen as a grounding source, you’re effectively invisible inside that AI’s output ecosystem. If it’s selected frequently, you gain authority and visibility. High SR becomes a self-reinforcing signal.

This isn’t just theoretical. Tools like Perplexity, Bing Copilot, and Gemini surface both answers and their sources. Frequent citation enhances your brand's visibility and perceived authority. Researchers even coined a term for how this feedback loop can lock in dominance: neural howlround. In an LLM, certain highly weighted inputs can become entrenched, creating response patterns that are resistant to correction, even when new training data or live prompts are introduced.

This concept isn’t new. In traditional search, higher-ranked pages earn more clicks. Those clicks send engagement signals back into the system, which can help sustain ranking position. It’s the same feedback loop, just through a different lens. SR doesn’t create bias, it reveals it, and whether you benefit depends on how well you’ve structured your presence to be retrieved in the first place.

Branding and the Reality of Interpretation

Brands on the Ballot frames this as non-neutral branding: companies can’t avoid being interpreted. Every decision, big or small, is read as a signal. That’s bias at the level of perception.

We see this constantly. When Nike featured Colin Kaepernick, some people doubled down on loyalty while others publicly cut ties. When Bud Light partnered with a trans influencer, backlash dominated national news. Disney’s disputes with Florida politicians over cultural policy became a corporate identity story overnight.

None of these were just “marketing campaigns.” Each was read as a cultural stance. Even decisions that seem operational (which platforms you advertise on, which sponsorships you accept, which suppliers you choose) are interpreted as signals of alignment.

Neutrality doesn’t land as neutral anymore, which means PR and marketing teams alike need to plan for interpretation as part of their day-to-day reality.

Directed Bias as a Useful Lens

Marketers already practice deliberate exclusion through ICP targeting and positioning. You decide who you want to reach and, by extension, who you don’t. That’s not new.

But when you view those choices through the lens of bias, it sharpens the point: positioning is bias with intent. It’s not hidden. It’s not accidental. It’s a deliberate narrowing of focus.

That’s where the idea of directed bias comes in. You can think of it as another way to describe ICP targeting or market positioning. It’s not a doctrine, just a lens. The value in naming it this way is that it connects what marketers already do to the broader conversation about how search and AI systems encode bias.

Bias in Traditional Search

Bias isn’t confined to branding or AI. We’ve known for years that search rankings can shape behavior.

A 2024 PLOS study showed that simply altering the order of results can shift opinions by as much as 30 percent. People trust higher-ranked results more, even when the underlying information is the same.

Filter bubbles amplify this effect. By tailoring results based on history, search engines reinforce existing views and limit exposure to alternatives.

Beyond those behavioral biases lie structural ones. Search engines reward freshness, meaning sites crawled and updated more frequently often gain an edge in visibility, especially for time-sensitive queries. Country-code top-level domains (ccTLDs) like .fr or .jp can signal regional relevance, giving them preference in localized searches. And then there’s popularity and brand bias: established or trusted brands are often favored in rankings, even when their content isn’t necessarily stronger, which makes it harder for smaller or newer competitors to break through.

For marketing and PR professionals, the lesson is the same: input bias (what data is available about you) and process bias (how systems rank and present it) directly shape what audiences believe to be true.

Bias in LLM Outputs

Large Language Models introduce new layers of bias.

Training data is rarely balanced. Some groups, voices, or perspectives can be over-represented while others are missing. That shapes the answers these systems give. Prompt design adds another layer: confirmation bias and availability bias can creep in depending on how the question is asked.

Recent research shows just how messy this can get.

• MIT researchers found that even the order of documents fed into an LLM can change the outcome.

• A 2024 Nature paper catalogued the different types of bias showing up in LLMs, from representation gaps to cultural framing.

• A PNAS study confirmed that even after fairness tuning, implicit biases still persist.

• LiveScience reported that newer chatbots tend to oversimplify scientific studies, glossing over critical details.

These aren’t fringe findings. They show that bias in AI isn’t an edge case, it’s the default. For marketers and communicators, the point isn’t to master the science, it’s to understand that outputs can misrepresent you if you’re not shaping what gets pulled in the first place.

Pulling the Threads Together

Selection Rate shows us bias at work inside AI retrieval systems. Branding shows us how bias works in the marketplace of perception. Directed bias is a way to connect those realities, reminding us that not all bias is accidental. Sometimes it’s chosen.

The key isn’t to pretend bias doesn’t exist; of course it does. It’s to recognize whether it’s happening to you passively, or whether you’re applying it actively and strategically. Both marketers and PR specialists have a role here: one in building retrievable assets, the other in shaping narrative resilience. (PS: an AI cannot really replace a human for this work.)

Practical Guidance

So what should you do with this?

Understand where bias is exposed. In search, bias is revealed through studies, audits, and SEO testing. In AI, it’s uncovered by researchers probing outputs with structured prompts. In branding, it’s revealed in customer reaction. The key is knowing that bias always shows itself somewhere and if you’re not looking for it, you’re missing critical signals about how you’re being perceived or retrieved.

Recognize who hides bias. Search engines and LLM providers don’t always disclose how selections are weighted. Companies often claim neutrality even when their choices say otherwise. Hiding bias doesn’t make it go away, it makes it harder to address and creates more risk when it eventually surfaces. If you aren’t transparent about your stance, someone else may define it for you.

Treat bias as clarity. You don’t need to frame your positioning as “our directed bias.” But you should acknowledge that when you pick an ICP, craft messaging, or optimize content for AI retrieval, you’re making deliberate choices about inclusion and exclusion. Clarity means accepting those choices, measuring their impact, and owning the direction you’ve set. That’s the difference between bias shaping you and you shaping bias.

Apply discipline to your AI footprint. Just as you shape brand positioning with intent, you need to decide how you want to appear in AI systems. That means publishing content in ways that are retrievable, structured with trust markers, and aligned with your desired stance. If you don’t manage this actively, AI will still make choices about you, they just won’t be choices you controlled.

A Final Danger To Consider

Bias isn’t really a villain. Hidden bias is.

In search engines, in AI systems, and in the marketplace, bias is the default. The mistake isn’t having it. The mistake is letting it shape outcomes without realizing it’s there. You can either define your bias with intent or leave it to chance. One path gives you control. The other leaves your brand and business at the mercy of how others decide to interpret you.

And here’s a thought that occurred to me while working through this: what if bias itself could be turned into an attack vector? I’m certain this isn’t a fresh idea, but let’s walk through it anyway. Imagine a competitor seeding enough content to frame your company in a certain light, so that when an LLM compresses those inputs into an answer, their version of you is what shows up. They wouldn’t even need to name you directly. Just describe you well enough that the system makes the connection. No need to cross any legal lines here either, as today’s LLMs are really good at guessing a brand when you just describe their logo or a well-known trait in common language.

The unsettling part is how plausible that feels. LLMs don’t fact-check in the traditional sense; they compress patterns from the data available to them. If the patterns are skewed because someone has been deliberately shaping the narrative, the outputs can reflect that skew. In effect, your competitor’s “version” of your brand could become the “default” description users see when they ask the system about you.

Now imagine this happening at scale. A whisper campaign online doesn’t need to trend to have impact. It just needs to exist in enough places, in enough variations, that an AI model treats it as consensus. Once it’s baked into responses, users may have a hard time finding your side of the story.

I don’t know if that’s an actual near-term risk or just an edge-case thought experiment, but it’s worth asking: would you be prepared if someone tried to redefine your business that way?