Obility Editorial · · 4 min read
One study gave the fear its name
Ask the founder of a smaller company what worries them about AI search and the answer is usually some version of this: the assistant will just name the market leader. There is a study people point to. In a September 2025 preprint titled “Generative Engine Optimization: How to Dominate AI Search”, four researchers at the University of Toronto ran an experiment they called big brand bias. They put fifty unbranded prompts about cola and soda to ChatGPT and Perplexity, then sorted every brand named in the answers into a list of twenty major brands, a list of twenty niche or regional ones, and everything else.
The result looks bleak for the small brand. Major brands took 56.3 percent of identified mentions in ChatGPT’s answers and 67.9 percent in Perplexity’s. Niche brands took 12.3 and 5.8 percent. Coca-Cola and Pepsi led the counts in both engines.
The prompts asked for popularity, and the engines supplied it
Read the prompts before the percentages. The examples the paper lists are: most popular cola brand, best selling soda in the world, top cola brands in the US, classic soda brands everyone knows, and most iconic soda brands. Every one of them asks which brand is most popular or best known. The correct answer to those questions is Coca-Cola. An engine that named a craft soda as the best selling soda in the world would simply be wrong.
So the experiment shows that engines answer popularity questions with popular brands, which is what anyone would want them to do. It does not show what happens when a buyer asks a question that a smaller brand is genuinely the right answer to. The authors’ own reading is more careful than the label. They write that, absent explicit constraints, source prominence and model priors pull recommendations toward major labels. For a smaller company, “absent explicit constraints” is the important part. It is also one consumer category, two engines and a preprint, so it says nothing direct about B2B software.
Buyers rarely ask for the most popular tool
The pull toward leaders has a plain cause. A model learns which names appear alongside a category, and a search step retrieves the pages that rank for the question. For a broad question, both point to the companies that are written about most. That is a fair description of a market leader, and no amount of content will change it quickly.
B2B buyers, though, usually arrive with constraints: team size, industry, the system the tool must connect to, the region, the budget, a compliance requirement. “Best payroll software” is a broad question. “Payroll software for a forty person agency paying contractors in three countries” is a different question, answered from different pages: a specialist guide, a comparison written for agencies, a vendor page that states which contractor countries it supports. The leader may have no such page. A smaller company whose product was built for that case can.
Here is an illustration with made-up numbers. A payroll startup tracks 40 buying questions. Ten are broad, such as best payroll software. Thirty carry the constraints its customers actually mention on sales calls. It appears in 1 of the 10 broad answers and 12 of the 30 qualified ones. Reported as one figure, 13 of 40, it looks like a brand losing to bigger names everywhere. Split, it shows the broad questions belong to the leaders, as expected, while the qualified questions, where its buyers really are, are winnable and already partly won. The 18 qualified answers it is missing from are the work list.
Each lost qualified answer has a readable cause
For every qualified question you lose, read the full answer and the sources it cites. The cause is usually one of three. No page of yours states the constraint plainly, so nothing tells the engine you support those countries or that team size. The independent pages the engine retrieves for that question do not mention you. Or a competitor genuinely fits the case better. The first two are content and coverage work. The third is a product or positioning fact, and writing around it produces pages that fail the buyer’s first check.
The Toronto authors reach a similar conclusion in their advice for niche brands: compete by owning a specific, narrow niche through deep expert content and coverage in specialty publications, rather than on the broad question.
Obility’s Answer Engine Insights is built for this split. Prompts are grouped by audience, buying stage and market, and each observation keeps the surrounding answer, the competitors it recommends and the sources it cites. That shows which competitor took each qualified answer and from which pages, so the gap can be assigned to a content review or a source correction. Which engines and markets are covered is something to confirm at a demo.
Set targets on the questions you can win
Do not expect to win the broad question soon. It reflects how much the whole web writes about each brand, and a model that has already learned a category’s leaders keeps them until a newer model is trained. One page aimed at “best payroll software” rarely moves it. Track the broad questions as a lagging signal, but set goals on the qualified ones.
And if you cannot find a qualified question where your product is the best answer, AI search has not created that problem. It has exposed a positioning problem that buyers would have found anyway.