When you track thousands of AI answers across wildly different industries, you stop seeing noise and start seeing tendency. Each model has a citation "personality" — recurring preferences about what kind of source to bring, how confidently to name, and when to hold back. They aren't fixed rules, and they shift with every update. But the patterns exist, and recognizing them changes how you read your own presence.
First, an honest caveat: bias here isn't an accusation. Any system that has to choose what to include in a short answer will, by definition, privilege some signals over others. The useful move isn't to be outraged by it — it's to understand the ruler so you can play better.
Authority Bias: The Weight of Established Sources
The most universal tendency across models is a fondness for high-authority sources. Encyclopedias, reference publications, recognized institutions, and sites with established reputations appear with disproportionate frequency. It makes sense: a model trained not to be wrong prefers to lean on what's least likely to be wrong.
For your brand, that means a mention in a respected outlet weighs far more than ten mentions in obscure corners of the web. The model doesn't count mentions; it weighs by trust.
Recency Bias: Live Search Favors the Current
Models with real-time retrieval carry a recency bias. They tend to surface what's fresh on the web — recent, updated, active content. A page that answered the question perfectly two years ago, but that no one has touched since, loses ground to a newer answer of similar quality.
Models that answer from memory have the opposite bias: knowledge cutoff. They privilege what was true up to their training date and may completely ignore what changed afterward.
Consensus Bias: AI Likes What Repeats
There's a subtle, important pattern: models tend to reflect consensus. If many trustworthy sources say the same thing about a topic or name the same brands, that convergence becomes the "safe" answer the model offers. Minority positions, even correct ones, appear less.
The practical implication is uncomfortable but clear: being associated with a topic in many places, in coherent ways, builds the kind of consensus the AI reproduces. A scattered, contradictory presence muddies the signal.
Format Bias: Clear Structure Gets Rewarded
Repeatedly, well-structured content — with explicit questions and answers, clear hierarchy, direct language — is easier for AI to interpret and reuse. It's not that the model "likes" lists; it's that it can extract information more reliably from an organized page than from a diffuse one. Format becomes bias because it makes retrieval's job easier.
Why These Biases Aren't Tricks to Exploit
Some people read a list like this and think of shortcuts: pile on structure, manufacture recency, fake consensus. It doesn't work, and for a good reason. Model biases are, at heart, proxies for quality. They privilege authority, currency, consensus, and clarity because those things tend to travel with trustworthy information.
When you try to fake the signal without the substance — apparent authority without real reputation, structure without content, repetition without merit — you're optimizing for a ruler the next model update will recalibrate. What survives each update is the substance behind the signal.
How to Use This in Your Favor
Recognizing biases serves one purpose: prioritization. If authority weighs heavily, invest where a mention pays off most — outlets and sources the models respect. If recency matters, keep your best-answer content alive instead of publishing and abandoning it. If consensus counts, pursue positioning coherence everywhere your brand appears. If format helps, write to be interpreted: clear, direct, answering real questions.
And because these biases differ from model to model and change over time, the only way to know which one is helping or hurting you right now is to observe the real behavior, not the theory. That reading — which citation patterns each model applies to your industry, and where you fit in or fall out — is what Genoma turns into something you can track and act on.