Anatomy of an AI Answer: Two Experts Walk In, One Gets Named

Imagine two experts, equally experienced, equally good at the work. A buyer types a question into an assistant, and four seconds later one of them is named and the other does not exist. Nothing unjust happened inside the machine. Something legible happened, six times in a row, and it is worth watching in slow motion.

What happens in the seconds after someone asks?

Call them Expert A and Expert B. The question is ordinary: "Who should a mid-sized firm hire to sort out its supply chain reporting?" The assistant does not experience this as a sentence; it parses it into intent and constraints. A recommendation is being requested. The subject is supply chain reporting. The buyer is a mid-sized firm, which implies budgets, tooling, seniority. Already the question has become a specification, and the specification decides everything downstream.

This is stage one, and it eliminates nobody yet. But note what it establishes: the machine is not looking for the best person. It is looking for the person best matched to a decomposed, machine-written brief. Experts who describe themselves in the language buyers actually use enter the next stage with their relevance pre-declared. Experts who describe themselves as "passionate about transformation" enter it as static.

It is worth pausing on how unromantic this is. No committee convenes, no taste is exercised, nobody remembers the brilliant workshop you ran. A specification meets a corpus, and everything you have ever published either matches the brief or does not.

Where does the engine go looking?

Stage two is retrieval. The assistant queries its index, or the live web, for material matching the specification, and here geography matters: not where the experts are, but where their evidence lives. We mapped this terrain in where AI looks before recommending anyone: professional publications, directories, review platforms, industry press, personal sites it can actually crawl.

Expert A has published explainers on supply chain reporting, holds a current directory profile, and was quoted in two trade articles. Expert B's expertise lives in client work under NDA, a LinkedIn feed, and a conference talk that exists only as a photograph. Retrieval returns eleven documents touching Expert A and two touching Expert B. Nobody has judged anyone yet; the machine has simply gathered what exists. But the match is already lopsided, because retrieval cannot weigh evidence that was never written down in a place it reads.

Why do passages beat pages?

Stage three is where most content quietly dies. The engine does not evaluate whole pages; it works with chunks, self-contained passages a few hundred words long, and scores each for relevance, clarity, freshness and authority. A brilliant insight buried in paragraph fourteen of a meandering post scores as noise. A crisp passage that defines a problem, states a claim and supports it with a number scores as gold.

This is not speculation. Princeton's GEO research, which tested nine optimisation tactics across 10,000 queries, found that adding citations, quotations and statistics lifted a source's visibility in generative answers by up to 40%, with fluent, authoritative prose close behind. The engine keeps what it can lift and defend. Expert A writes in liftable units. Expert B's two documents are a bio page of adjectives and a post with no dates, no numbers and no quotable claim. Eleven documents against two just became nine usable chunks against zero.

Writing for this stage is a craft of its own, and a learnable one. Give each important claim its own paragraph with a beginning and an end. Put the number beside the claim it supports rather than three screens away. Name your sources in the text, date what you publish, and resist burying your best sentence under a warm-up. None of this dumbs anything down; it packages depth so that extraction preserves it. The engine is a quoting machine, and quotable is a property you can build on purpose.

How does the engine decide who is actually who?

Stage four is entity resolution, the quiet identity check we explored in Entity, Not Ego. The chunks mention names; the engine must now decide which names denote one consistent, trustworthy person. Expert A is the same name, same title, same specialty everywhere, so nine chunks fuse into one strong candidate. If Expert B had survived to this stage, an old bio naming a different niche and an abandoned profile with a stale employer would have split the signal into weak fragments.

Then comes stage five, corroboration. Self-description is cheap, so the engine weighs what independent sources agree on. Directory, trade press, review platform: do they tell one story? For Expert A they do. This is also the stage where the machine's own caution shows: naming a person in a recommendation is a small act of reputational risk for the assistant, and it prefers candidates whose facts are confirmed by parties with nothing to gain, a dynamic we traced in how AI decides what is true about a person.

Notice what corroboration does to the economics of exaggeration. A lone website can claim anything, which is exactly why the engine discounts it. Three independent sources repeating the same modest facts outweigh one page of magnificent self-description, because agreement is expensive to fake and adjectives are free. The experts who win this stage tend to be the ones who invested in being verifiable rather than impressive.

When two people share the same name

Entity resolution gets much harder when the corpus contains two, three, or a dozen real people who share an identical or near-identical name. Imagine two people named David Mehta: one runs a boutique tax advisory in Mumbai, the other is a mechanical engineer in Toronto who occasionally blogs about vintage motorcycles. A generic web search barely disambiguates them, but an AI engine trying to answer who David Mehta the tax advisor actually is has to do something closer to detective work. It looks for co-occurring signals, the words that keep showing up next to the name, terms like tax, advisory, Mumbai, GST, repeated across multiple independent documents. The mechanical engineer's motorcycle blog shares none of those co-occurring terms, so the engine's confidence that this cluster of mentions belongs to the tax advisor stays high, while the motorcycle blogger's mentions form their own separate, unconnected cluster.

Where this breaks down is when someone's own public footprint mixes signals that don't obviously belong together, a common problem for people with genuinely varied careers. If the tax advisor David Mehta also happens to have written two guest posts about running marathons, the engine has to decide whether that's the same person having a hobby or evidence of a second cluster it should keep separate. This is exactly where structured identifiers help. A Wikidata entry, a consistent professional schema markup, or a single canonical bio linked from every property acts like a name tag the engine can trust more than free-text guessing, a mechanism explored further in how Wikipedia and Wikidata earn AI's trust. Without one, the engine defaults to caution, and caution usually means favoring the more heavily corroborated cluster and quietly setting aside the thinner, ambiguous one, even when that thinner cluster was the actual person being asked about.

With several corroborating sources, which one actually gets quoted?

Corroboration is not winner-take-all in the sense of one source deciding the outcome, but synthesis, stage six, still has to choose which one or two sources actually appear as the visible citation, since most answers do not list every supporting document, only the strongest handful. When three independent sources agree on the same fact about a person, say a directory listing, a trade press quote, and a conference bio, the engine tends to favor the source that states the fact most directly and most recently over the one that merely implies it. A trade press quote from four months ago that says plainly who someone is and what they specialize in will usually beat a two-year-old conference bio saying the same thing in a roundabout way, even though both are technically corroborating. Recency and directness of phrasing act as quiet tiebreakers among sources that already agree with each other.

The sources that get quietly passed over are not being punished. They are just less efficient to lift, which matters when a model is choosing the shortest, clearest path to a defensible sentence. This is also where the machine's own caution shows again, similar to the dynamic behind some sources getting favored more broadly across many queries, a pattern examined in whether AI models play favorites. A model would rather cite the plainest, most recent confirmation of a fact than assemble a citation from three partial fragments, even if the fragments taken together are just as true.

Figure: the same six-stage pipeline walked through as a single ordered sequence, start to finish.

What tips the final naming?

Stage six is synthesis: the model writes the answer, citing as it goes. By now the contest is usually settled, but when two candidates both survive, the tiebreakers are mundane. Whose passages answer more of the specification's sub-parts? Whose facts are fresher? Whose identity resolved with more confidence? The margin can be one good published explainer or one recent third-party mention. From the outside it looks like fortune. From the inside it is bookkeeping.

Recency deserves special mention among the tiebreakers. Retrieval systems weigh freshness, and a candidate whose latest citable evidence is four years old is presumed, gently and wrongly, to have stopped. An expert in their prime with a dormant paper trail can lose to a slightly junior rival who published last quarter. The machine is not measuring ability; it is measuring the most recent proof of ability, and the two drift apart in silence.

Why does the loser never find out?

Here is the property that makes this pipeline strategically dangerous: it produces no rejection letter. In classic search, Expert B could at least see themselves on page three and feel the gap. In an AI answer there is no page three. There is one synthesised reply, read by a buyer who never clicks anything further, which is increasingly the norm now that 69% of Google searches end without a click. Expert B loses the match without ever learning a match took place, which is why so many capable people experience the pattern we described in why AI recommends your competitor as inexplicable bad luck.

And the match keeps being replayed. Tracking by the Semrush AI Visibility Index shows 40 to 60% of cited sources rotate month over month. Today's winner defends the belt every time anyone asks. Which is bad news for complacency and genuinely good news for Expert B, because the pipeline holds no grudges either.

The remedy for this species of invisibility is measurement you run yourself. Ask the assistants the questions your buyers would ask, monthly, and keep the answers. It is the only scoreboard this game publishes, and almost nobody in the losing bracket has ever looked at it.

What would Expert B change tomorrow?

Each stage names its own remedy, so the to-do list writes itself:

1. Describe yourself in specification language. The problems you solve, for whom, at what scale, in the words buyers type.
2. Move evidence to crawlable ground. Publish the thinking trapped in decks, DMs and gated feeds; retrieval cannot cite a memory.
3. Write liftable passages. Claims with numbers, dates, sources and quotes, stated cleanly enough to survive extraction.
4. Resolve your entity. One name, one title, one bio, everywhere the engines read.
5. Earn corroboration. A directory listing, a trade quote, a review stream: independent voices that repeat your story.
6. Retest monthly. Ask the assistants the buyer's question and watch your progress through the stages.

If you would like the current state of your own pipeline mapped, stage by stage, that is exactly what our services do, and the rest of the journal unpacks each stage in more depth.