The first time you see your own company named inside an AI answer, the reaction is usually suspicion. You assume you phrased the prompt in a way that fished for it. So you rephrase, open a clean session, try a different engine, and it happens again. Somewhere along the line you stop treating it as a fluke.

Then something odder follows. Over the next few weeks the same name starts turning up in answers to questions you were not thinking about when it first appeared. Adjacent questions. Broader ones. Questions where you would have expected to be nowhere near the shortlist.

That progression is the interesting part, and it does not look like the way rankings used to move.

The first mention costs far more than the second

Getting named the first time is slow, unglamorous work. Something has to exist in enough places, described consistently enough, that an engine assembling an answer has a reason to reach for you rather than for one of the obvious incumbents. Usually it takes a piece of writing specific enough to be worth attributing, published somewhere that gets read by people who write other things.

What surprises people is how much cheaper the next ones are. Once a name is in circulation as an answer to one question, it starts appearing as a candidate for neighbouring questions without anyone doing additional work to put it there.

Which makes the early phase feel worse than it is. You spend a quarter producing material and checking prompts and seeing nothing, and the temptation is to conclude the whole exercise was mistimed. The curve simply has a long flat section at the beginning, and the flat section looks identical to failure while you are standing in it.

There is a good write-up on the AI citation snowball effect that puts a name to something I had spent a year describing badly in client calls. The naming helps, because the pattern is easy to misread as ordinary momentum when the mechanism underneath it is different.

Why the second one comes cheaper

Very little of this is mysterious once you follow where the material goes. A claim shows up in one credible place. A roundup picks it up because roundups are assembled by people reading the same sources you are. A forum thread quotes the roundup. A newsletter summarises the thread. Six months later the claim exists in five places, each one citing something upstream, and the name attached to it has travelled with it every time.

Nothing about that requires a search engine. It is how professional opinion has always propagated. What changed is that the propagation now feeds something that reads all of it and produces a single synthesised answer, so the accumulated agreement of those five places arrives at the reader as one confident sentence.

That compression is what makes the compounding feel faster than it used to. In the old arrangement, a reader had to encounter your name in three separate places over three separate sessions before it registered as familiar. Now the three encounters get folded into one paragraph and handed over together, and familiarity that would have taken months to build gets delivered in a single reading. The flip side is that absence compresses the same way. A name missing from the upstream material is missing from every answer assembled out of it, all at once, with nothing gradual about the exclusion.

The practical consequence is that a first mention earns its keep through repeatability. Traffic from it barely registers and was never really the point. A mention that cannot be lifted out without its original context attached goes nowhere, however prestigious the place it appeared, and a lot of hard-won coverage sits in that category.

You can test any given mention for this in about a minute. Copy the sentence your name appears in, paste it somewhere with nothing around it, and read it cold. If it still says something, it will travel. If it needs the two paragraphs above it to make sense, it will sit where it is.

The descriptor travels with you, and you did not choose it

Teams get caught out by the next bit. Whatever label was attached to you in the mentions that spread is the label you inherit, and it turns out to be remarkably hard to shake off later.

You see it when an engine describes your company accurately but in a category you have been trying to move out of for two years. Or with a qualifier somebody used once, offhand, in a comparison piece. The description was reasonable at the time. It then got copied by everyone summarising that piece, and the summaries got read.

Which means the framing in your early mentions matters more than their volume. A handful of mentions that describe you the way you want to be understood compounds into something useful. A larger pile that describes you as a cheaper version of something else compounds into a problem you will be arguing with for a long while.

Not every mention is built to travel

The kinds of mention behave differently, and it is worth being honest about which ones do work:

The last row is the one to aim at and the hardest to engineer. If your phrasing becomes the way people describe the thing, your name rides along with the phrasing for as long as it stays in use.

What makes a mention portable

Portability comes down to whether a paragraph survives being cut out of its surroundings. A few things reliably help:

• A claim stated plainly, in one sentence, without three qualifiers wrapped around it.
• Something specific enough that paraphrasing it loses information, which is what makes attribution worth doing.
• A subject named with the words other people use for that subject, so the fragment matches what someone is asking.
• No dependence on a preceding paragraph. Lifted text arrives alone.

The inverse is equally consistent. Hedged, carefully balanced writing is almost impossible to cite, because there is nothing in it anybody needs to credit to a source. Balanced writing has its place. That place is not the paragraph you hope gets quoted.

Where the snowball stalls

Momentum stops for reasons that are usually visible in hindsight:

• The mentions accumulate around one piece of work and nothing new arrives, so the association ages with it.
• The name appears in circulation while the site itself says something different, and the two versions of you cancel each other out.
• Everything sits inside one community. Snowballs need contact with other surfaces, and a single forum, however friendly, is one surface.
• The specific claim that earned the early mentions gets quietly retired from your own materials, so anyone checking finds nothing behind the citation.

That last one happens more than you would think, usually after a rebrand or a site migration, and it is the cheapest of the four to avoid.

The pattern across all of them is a supply problem rather than a distribution one. Snowballs stop when nothing new is being added at the front, and no amount of promotion fixes an empty pipeline. Teams tend to reach for more outreach at exactly the moment they needed one more genuinely specific claim.

Watching it move without traffic numbers to lean on

The awkward operational truth is that none of this arrives in your analytics in any legible form. Somebody reads an answer, remembers a name, searches it three days later on a different device, and lands as direct traffic with no story attached.

So the measurement is manual and it is mostly bookkeeping. A fixed set of prompts, run on a regular schedule, with what came back written down: which names appeared, in what order, described how, and citing what. Keep the raw answers rather than a summary, because the phrasing is the signal and a summary destroys it. Run the same prompts on more than one engine, since they disagree in ways that tell you something.

A fair amount has been written on earning and tracking AI citations, and the part that consistently holds up is the least exciting part: record what you observed, date it, and keep the record long enough that the trend becomes visible. Nobody sets this up until the third time somebody in a leadership meeting asks whether the company shows up in AI answers and the room has nothing to offer but anecdotes.

Six weeks of notes is enough to start seeing shape. A quarter of notes is enough to tell the difference between something moving and something you wanted to be moving.

What the accumulation actually buys

The endpoint is not a position on a list. It is that when somebody asks a question in your category, your name is already part of the furniture of the answer, mentioned in passing, as though its presence needed no explanation.

You can tell when a competitor has reached that state, because answers include them without justifying the inclusion. Nobody argues for them. They are simply there, the way certain names are simply there in any conversation among people who work in a field.

Getting to that point is a long accumulation of small, specific, repeatable statements attached to your name in places other people read. The first one is expensive. Every one after it is slightly less so, which is the whole reason the thing snowballs at all.