Most advice about getting recommended by AI is asserted rather than measured. This guide is built from a dataset we collected ourselves in Q3 2026: 240 brand-free buyer questions across twelve categories, each asked to five assistants, producing 1,188 completed answers. Every figure below traces back to a per-category page you can open and check.
First: assistants often name nobody
Across the categories measured, the share of answers naming at least one specific provider ranged from 38% to 64%. The remainder gave guidance on how to choose, or sent the buyer to a directory or review site.
That gap is the most actionable thing in the dataset. A question where the assistant names nobody is open ground: no competitor holds it either, and becoming the first concrete answer is usually cheaper than displacing an incumbent from a list.
Before asking how to outrank a competitor in an answer, check whether the answer names anyone at all.
Second: concentration varies by category, and predictably
Three of the twelve categories already had what we would call a default answer, meaning one provider appeared in at least a quarter of all answers: project management software at 38%, skincare at 35%, and electric bikes at 28%. All three are product categories with national or global brands.
Every local-service category behaved differently. In dentists, law firms, restaurants, hotels, gyms, home services and real estate, the most-named provider appeared in roughly 7% to 13% of answers. Those markets are not consolidated in the assistant's mind, and a well-documented local business is competing against a long tail rather than against an incumbent.
Third: models behave differently enough to matter
On the same questions, one model named an average of 3.5 providers per answer while another named 2.1. Citation behaviour differed more sharply still: one attached sources to essentially every answer, another to about 85% of them.
The practical consequence is that a single cross-model share-of-voice number hides the thing you would act on. A brand can be invisible in the model that names few options and comfortable in the one that lists many, and the average will describe neither situation.
Fourth: the sources are not who you would guess
Of 11,763 citations recorded, about 15% pointed at the websites of the providers being named, and the rest went elsewhere. reddit.com was cited in all twelve categories, which makes community presence a category-independent requirement rather than a tactic for consumer brands.
In six of the twelve categories, at least a fifth of answers pointed the buyer to a directory or review site rather than naming providers directly. In those categories a complete, well-reviewed directory listing is a precondition for being named at all, and no amount of publishing on your own domain substitutes for it.
What to do with this
Work in this order. Find the questions in your category where assistants currently name nobody and answer them concretely. Check which source types dominate citations in your category and get your facts onto those sources rather than only onto your own site. Then measure per model, not on an average, so you can tell which engine your work moved.
The per-category data, including the full question lists and the domains cited in each, is published alongside this guide.