Popular categories are remembered; specialist ones are read
A popular category is answered largely from memory and a specialist one from retrieval. Researchers at the University of Toronto tested this on GPT-4o with 216 comparison queries, half about well-known entities such as "Nike or Adidas" and half about niche ones such as "Aeropress or Chemex", in a study at the EDBT/ICDT 2026 workshops. When they shuffled the order of the retrieved snippets, rankings for popular entities barely moved (a mean rank deviation of 2.30) while niche rankings shifted almost twice as much (4.15). When they swapped two entity names inside the snippets, the niche rankings followed the text (4.63) and the popular ones held (2.60).
The consistency figures say the same thing from the other side: Kendall's tau between runs was 0.911 for popular entities and 0.556 for niche ones. The model treats the retrieved pages as confirmation for a brand it already knows, and as the actual evidence for one it does not. About 16 percent of entities in the popular rankings appeared in no retrieved snippet at all; the model added them from memory.
The specialist pool is narrower, and that is usable
The pool of sources an assistant reads for a specialist query is narrower than for a general one, and it is shared across engines. In the same study, the share of unique domains cited fell from 74.2 percent on general queries to 68.6 percent on niche ones: the engines concentrated on particular review sites and discussion pages for niche topics. Overlap with Google's top ten rose by three to four points on niche queries for Claude, Gemini and Perplexity, though it stayed near zero for GPT-4o.
For a specialist firm this is a map. In a mainstream category the sources are too many to influence; in a specialist one, a handful of trade directories, review sites, association pages and forums do most of the deciding, and they can be listed by running the queries and reading what is cited. The free analysis records exactly that: who is cited for your category instead of you.
What the mechanism changes in the work
The mechanism changes three things about where the effort goes.
First, the firm's own pages matter more than in mainstream SEO, because they are read as evidence rather than as confirmation. Pages with numbers, definitions and comparisons were absorbed 55 to 62 percent more strongly across 18,151 citations in Zhang, He and Yao's measurement, and a specialist firm usually has that material already: standards, tolerances, lead times, certifications, the comparison with the alternative method. It sits in the data sheet and not on the page.
Second, the entity has to be unambiguous, because a model with no memory of the firm cannot repair a missing address or a name spelled two ways. In Ahrefs' analysis of 75,000 brands, mentions of the brand across the web correlated more strongly with AI visibility than Domain Rating. In a specialist category those mentions are concentrated in the same narrow pool the engines read, which is why a listing in the right directory can outweigh a link from a general site.
Third, freshness carries more weight. AI engines cite content two to three times younger than Google for the same queries in the Toronto data, and they read the age from datePublished, dateModified and the time element. A specialist page last touched in 2022 is old to the engine even if nothing about the process has changed.
A niche audit in five checks
| Check | What to look at | What a fail looks like |
|---|---|---|
| Who is cited | Run the category queries on four engines, list the cited domains | Your firm absent; the same three directories present every time |
| Are you in the pool | Whether your firm appears on those cited pages at all | Listed nowhere the engines read |
| Own-page evidence | Numbers, definitions, comparisons on the decisive pages | Adjectives, no standards, no figures |
| Entity data | Organization data, identical name and address everywhere, resolving profiles | Two spellings, a dead profile link, no address |
| Freshness and access | Machine-readable dates; live fetch as GPTBot | No date at all; robots allows, server blocks |
The order matters. Being cited starts with being in the pool, so the second check comes before the editing; a perfectly evidenced page nobody retrieves does nothing. Our audit article explains how these checks are scored.
Where the advantage stops
The advantage stops where the category stops being specialist. Once a firm is well known, the model's memory takes over and retrieved pages move the ranking less; the Toronto data is explicit that this is a matter of degree, and a business that grows from niche to household name will feel the mechanism change. It also stops at the measurement: only 34 to 42 percent of cited sources stay the same from one day to the next in the St. Gallen study, and niche entities showed the lower consistency in Toronto, so a single query proves nothing. Seven runs per prompt is the floor the research sets before a rate means anything.
Frequently asked questions
What counts as a specialist category?
One where the assistant has no reliable prior about the businesses in it: regional services, B2B components, professional niches, anything a general-interest corpus covers thinly. The practical test is the Toronto one: if rankings change when the retrieved pages change, the category is being read rather than remembered.
Is niche SEO in AI search easier than mainstream?
Different rather than easier. The number of pages that decide the answer is smaller, which makes the work targetable; the same pages decide for your competitors, which makes it contested. Effort per page is higher because evidence, not placement, is what moves the ranking.
Do backlinks still matter in a specialist category?
For Google discovery, yes, and Google's AI features build on that. For being named by ChatGPT or Perplexity, a mention on a page the engines already cite for your category outweighs a link from an unrelated site. Both are useful; they are bought with different work.
How do I find the sources the engines read for my category?
Run the buying questions your customers ask, in their language and from their market, on each engine several times, and record the cited domains. The pattern is usually visible within a dozen runs. The free analysis does this and lists the competitors and sources named instead of you.
Does a small specialist site need a large amount of content?
No. It needs the decisive pages to carry evidence: what is offered, to which standard, in what timescale, compared with what alternative. Length without those does not help; Zhang's data shows Q&A formatting alone slightly below average for absorption.
How is this different from the small business article?
The small business article is about size: a firm the model does not remember. This one is about the category: a topic where the engines read a narrow, shared pool of sources. A large firm in a specialist category and a small firm in a general one face different halves of the same mechanism.
See what the engines read for your category
The free analysis runs your category on ChatGPT, Perplexity, Claude and Gemini from your market, lists who is cited instead of you, and reads your key pages for the evidence a specialist query needs.
