Do AI Models Play Favorites? How Familiarity Bias Decides Who Gets Named

Ask a model to recommend an expert in your field and a strange thing happens. The same three or four names surface, again and again, across sessions and sometimes across rival models. Is the machine biased? Yes. But the bias has a shape, and shapes can be learned.

Do AI models play favorites?

Let us take the question at face value, because people genuinely type it. If by favourites you mean a curated roster, some backstage arrangement where certain consultants paid for placement, then no. You cannot buy your way into a ChatGPT answer, and no major model ships with a list of preferred professionals. But if you mean a systematic tendency to name some people over others with similar credentials, then yes, emphatically. The tendency is called familiarity bias, and it is not a bug in the machinery. It is the machinery.

A language model is, at heart, an engine of expectation. It has read a colossal sample of the written world and learned which words tend to follow which. When you ask it who the leading voice on supply-chain finance is, it does not consult a ledger of merit. It produces the name that most strongly co-occurs with those concepts in everything it has absorbed. Familiarity is not one input among many. For the model's parametric memory, familiarity is close to the whole game.

Where does familiarity bias actually come from?

Three places, and it pays to keep them distinct. The first is training data. If your name appeared frequently and consistently near your specialism in the books, articles, forums and directories the model learned from, the association is baked into its weights. The second is retrieval. Modern assistants search the live web before answering, and they draw on a narrow set of sources they have learned to trust: established publications, review platforms, active communities. If those sources keep mentioning you, the retrieved evidence keeps confirming you. We mapped that terrain in where AI looks before recommending anyone. The third is consensus. When several independent sources agree, the model treats agreement as reliability and names you with confidence. When they disagree, it hedges, or names somebody clearer.

The five biases that decide who gets named

Familiarity is the headline, but it travels with four companions. Together they form a reasonably complete account of why one qualified person gets recommended and another does not.

Why does ChatGPT keep recommending the same people?

Because the sources it consults keep recommending the same people. This is the part most professionals miss. The model is not stubborn; it is downstream. When a handful of names dominate the industry publications, the review platforms, the conference line-ups and the community threads, the model's answer is a faithful compression of that record. Complaining to the machine is complaining to a mirror. The reflection changes when the room does.

There is also a safety instinct at work. A model penalised for wrong answers learns to prefer defensible ones, and the most defensible recommendation is the one with the deepest paper trail. Naming the obvious candidate is rarely punished. Naming a brilliant unknown might be. If it stings to watch a rival's name come back for a query you should own, we wrote the post-mortem for exactly that feeling in why AI recommends your competitor instead of you.

Is the favouritism permanent? The churn says no

Here is the finding that should reorganise your morale. Semrush's AI Visibility Index, as reported in Similarweb's roundup of generative AI statistics, found that 40 to 60 percent of the sources cited in AI answers rotate month over month. Read that again. In any given month, roughly half the citations underpinning AI answers are replaced. The favourites are not carved into the weights like initials in wet cement. A large share of who-gets-named is decided at retrieval time, freshly, from whatever the engine finds and trusts this month.

That churn cuts both ways, of course. It means incumbency is softer than it looks, and it means any position you win must be maintained. But for anyone currently outside the answer, it is the single most encouraging number in the field: the door is rehung every few weeks.

Do all the models share the same favourites?

Partially, and the overlap is itself revealing. Ask ChatGPT, Gemini, Claude and Perplexity the same who-should-I-hire question and you will typically find a shared core, the names with the deepest public records, surrounded by a fringe that differs per engine. The core overlaps because the models drank from overlapping wells: the same books, the same major publications, much of the same public web. The fringes differ because each system retrieves differently, trusts sources differently, and updates on its own calendar. Perplexity, which shows its citations, will happily reveal exactly which pages produced its shortlist; the others make you infer. We compared their temperaments properly in Perplexity vs ChatGPT vs Gemini.

The practical consequence is that you cannot audit one engine and declare yourself visible. A name can be a fixture in one model's answers and absent from another's, particularly for people whose reputation lives in one ecosystem, one country or one platform. The professionals who appear in every engine's core are the ones whose evidence is spread across many independent, widely-trusted sources, which no single retrieval pipeline can miss. Breadth of documentation, not volume on a single channel, is what turns a one-engine favourite into an everywhere favourite.

Can a lesser-known name break in?

Yes, and the mechanism is precisely the biases in the table above, worked deliberately. What a newcomer cannot manufacture is decades of fame. What a newcomer absolutely can manufacture is consistency, evidence and presence in the right rooms. A big audience, incidentally, is not the lever; we have shown before that a big following will not make AI name you. Follower counts live on platforms the models barely weight. Citations, bylines, reviews and structured facts live where the models actually read.

Watch what the machine considers a strong signal: a specialist who answers one category of question in depth, whose name is spelled and titled identically everywhere, who appears in two or three trusted publications, and whose site states plainly what they do. That profile can start being named within months, not decades, especially for specific queries where the famous incumbents are only vaguely relevant. Specificity is the newcomer's crowbar. The narrower the question, the weaker the celebrity's grip on it.

Why does this feel so familiar? Because it is how people work too

Strip away the servers and the training runs, and familiarity bias is not really an alien machine quirk. It is the same shortcut a person's brain uses every day, dressed in different clothes. Think about how you settle on a dentist, a mechanic, or a family doctor. You rarely conduct a formal audit of every practitioner in town. You go with the name that has come up more than once, from more than one person, in a way that felt consistent each time. Psychologists describe the underlying pattern plainly: repeated, effortless exposure to a name breeds a comfortable sense that it must be reliable, even when no one ever ran a side-by-side comparison. A language model does something structurally similar at enormous scale. It has effectively "heard" some names mentioned constantly, next to the right topic, in sources that agree with each other, and that repetition reads, to the model, the way repeated recommendations read to you at a dinner party. This is closely related to the dynamic we explore in unknown expert versus celebrity in AI answers, where fame and mere familiarity get tangled together.

Imagine two equally qualified estate planning attorneys, hypothetically, call them Renu and Marcus. Both have the same years of experience, the same caliber of client work, and genuinely comparable expertise. Renu has been quoted in three regional legal publications, keeps a maintained bio page, and has been a repeat guest on two small industry podcasts, always under the same spelling of her name, the exact machine-readable habit we cover in the comeback of the byline. Marcus does equally excellent work but has never been quoted anywhere, has no consistent bio across the two directories that list him, and has no public track record a search engine or an AI model can find. Ask an AI assistant to recommend an estate planning attorney and it will lean toward Renu, not because her legal work is superior, but because Renu is the name the model has effectively heard more consistently and from more independent directions. Marcus's competence is real. It is simply invisible to a system that can only work with what has been written down and repeated.

The uncomfortable part of this is that it can look unfair from the outside, and in a narrow sense it is: merit and familiarity are not automatically the same thing. The encouraging part is that familiarity, unlike raw talent, is buildable on a schedule. It does not require becoming famous. It requires becoming consistently and verifiably mentioned, in the same way, in places that talk to each other, which is the entity-building work described in entity, not ego. That is a project a person can actually run, which is more than can be said for waiting to be discovered.

The three-stage build-up from a single mention to becoming an AI system's default answer for a given question.

How do you become the familiar name? Honestly

There is no trick, which is the best news available, because tricks get patched and fundamentals compound. The programme looks like this. Pick the narrow question you want to own. Publish the deepest, best-sourced answers to it that exist, under your byline. Make your identity consistent across every profile so the machine can attach each new signal to one entity rather than three blurry ones. Earn independent corroboration: publications, podcasts, reviews, community mentions. Then re-test monthly with the routine from Googling yourself versus asking ChatGPT, because the churn that let you in will happily let you back out. If you want a structured version of that programme run for you, that is quite literally the service we built.

What should you avoid? Shortcuts that simulate familiarity without earning it. Mass-produced guest posts on sites no engine respects, reciprocal mention schemes, biography pages that quietly inflate. These fail twice over: the models weight the sources you planted them in at close to zero, and any contradiction between your claims and the third-party record teaches the machine to hedge on your name, which is the opposite of the confidence you were trying to buy. Familiarity built from thin material is a scaffold in a storm. The monthly churn takes it down as quickly as it went up, whereas familiarity built from real citations keeps standing precisely because independent sources keep restating it without being asked.

One more honest note. Familiarity bias favours the documented over the merely excellent, and that will sometimes feel unjust, because it sometimes is. But unlike the biases of a human gatekeeper, this one publishes its criteria in its behaviour, holds no grudges, and re-reads the evidence every month. As gatekeepers go, you could do considerably worse.