There are three slots. Nationally that's a lottery. In your city it isn't.

Start with the number that decides everything: an AI answer names three to five companies.

Not ten blue links. Not a page two. Three to five, in a sentence or a short list, and then the answer ends. There is no scrolling. Whoever is not in that set was not in the conversation.

A June 2026 analysis measured how those slots distribute across 91 companies in three national categories, using ten models and three hundred prompts. The result is not a bell curve. It is a power law:

Category            Leader   Leader share   Median company   Gap
Observability       Datadog        78.3%              ~9%    69 pts
Outbound sales      Apollo         60.5%             ~10%    50 pts
AI coding tools     Copilot        55.3%              ~8%    47 pts

The top name is in nearly every answer. The fifteenth is in almost none. Independent work points the same way — an analysis of AI Overview citations across a large query set found the top domains capturing the overwhelming majority of citations, with the tail getting effectively nothing.

That study’s conclusion for everyone outside the top two is that incremental content work will not close a gap of forty points, and that the answer is to become the definitive reference for one narrow niche instead.

It is right about the arithmetic, and it stops one step before the useful part.

The power law is a function of the question, not of AI

Here is what the national numbers actually measure: what happens when you ask a question with tens of thousands of eligible answers and no constraint.

“What’s the best observability tool?” has a global candidate set. Every company on earth that does the thing is competing for three slots. Of course it concentrates — the model has nothing to narrow on except prominence, and prominence is exactly what the biggest company has the most of. Concentration under those conditions isn’t a property of AI. It’s a property of asking an unconstrained question.

Now add a constraint the customer was always going to add anyway.

“Who’s the best plumber in Tulsa?” The candidate set is not tens of thousands. It is a few dozen. And the model still returns three to five.

Same number of slots. Two orders of magnitude fewer competitors.

That is the entire argument, and it is why a national visibility report and a local one are different products even though they run the same way.

What changes when the candidate set is small

Three things, and they all move in your favour.

The gap between first and fifth is small. With forty candidates and five slots, the difference between being in the answer and being out of it is not a 69-point mountain: a field that size has no room for one. How small it is in your category is a measurement rather than an assumption, which is why the report takes it instead of asserting it.

Prominence stops dominating. The national leader’s advantage is that it is written about constantly, everywhere. Nobody is written about constantly in Tulsa. When no candidate has overwhelming presence, small differences in readability, freshness and third-party description decide the ordering — and those are the things you can change.

Nobody is defending. The national top three employ people whose job is this. In a city category, the companies in the answer are usually there by accident. They have not run the test. They do not know they are in it. They are not maintaining anything.

                                  National     City × category
Candidate set                      10,000+             20 – 60
Slots in an answer                   3 – 5               3 – 5
Leader share                      55 – 78%    not measured yet
Gap, 1st to 5th                    ~50 pts    not measured yet
Anyone actively working on it          yes        almost never

The right-hand column is an argument from structure, not a result. The two cells marked not measured yet are the ones only a measurement can fill, and they stay that way until we have one to publish with the data behind it. The left-hand column is that study’s published figures.

The trap inside the good news

Two of them, and they’re the reason this isn’t free money.

A small candidate set is small for everyone. If a competitor in your city reads this and acts first, the same shallow gap works against you. The window is open because nobody is looking, and that is a temporary condition by definition.

Local questions aren’t always answered locally. Ask about a service that people buy remotely and the model may ignore your city entirely and return national names. Whether your category is genuinely local is an empirical question about how the models behave, not an assumption — and it is the first thing worth checking, because if the answer is no, everything above stops applying and if you opened after 2023’s framing takes over instead.

What to do with this

Find out how many candidates you’re actually competing with. Ask the buying question in your city, across the three search-enabled models, several times. Collect every name that appears. That set — not the census of businesses in your town — is the real field. It is almost always much smaller than people expect, and seeing it written down changes how the problem feels.

Find out where the current occupants come from. Look at what the models cite when they name them. In local categories it is very often a small number of third-party pages: a directory, a regional roundup, a review site, one local news piece. Those are addressable in a way that outranking a company with a decade of press is not.

Pick the constrained question, not the broad one. Best [category] in [city] is winnable. Best [category] is not, and the question your customer asked explains why the constrained version is also the one your customer actually asks.

Then measure the ordering, not just the presence. Being fifth in a five-name answer is not the same as being first, and the two are recorded as the same thing by most tools. Mentioned is not recommended has the split.