Measured data
Which sources do AI engines cite?
When an assistant recommends a company, it is retelling a source. Knowing which sources it reads in your field is the whole game: that is where presence turns into a mention. Short answer from our measurements — maps and directories, far more than brand websites, and in wildly different amounts depending on the engine. Here is the real picture.
First: about half of answers cite nothing at all
Before asking which sources, note that only part of answers are sourced in the first place. Across our measured answers, roughly 47% carried any citable source; the rest were produced from the model's memory with nothing to point to. That split is the first fork: an engine answering from memory can't be influenced by a listing, because it isn't reading one. Everything below is about the sourced half — the answers you can actually move.
The sources that feed answers: maps and directories
In the sourced answers, the platforms that come up most are not brands' own sites — they are mapping and listing services. Aggregated across the measured niches, the sources cited most often were:
- 2GIS — the single most-cited platform across niches.
- Google (incl. Maps) and Yandex Maps — the mapping layer.
- Zoon and category directories such as ProDoctorov, Yandex Uslugi — vertical listings for a field.
- Niche portals — a local car-rental aggregator, a regional real-estate board, a city medical catalogue — which often outrank the global names inside their own niche.
A brand's own website, by contrast, appears among the cited sources in only about 32.7% of answers. Two answers in three are assembled without your site at all. That's why rewriting your homepage closes, at best, a third of the task — the rest lives on the platforms the model reads.
The number almost no tool shows you: engines don't cite alike
This is the finding that changes what you do. How heavily an engine leans on directories is not a property of "AI" — it's a property of each engine. In a single measured niche, the share of an engine's answers that rested on a directory source ran like this:
| Engine | Answers leaning on a directory | Interval |
|---|---|---|
| Yandex (search answer) | 86.5% | 83.0–89.4 |
| Perplexity | 53.1% | 48.1–58.1 |
| GigaChat | 4.6% | 2.9–7.2 |
| OpenAI (from memory) | 0.0% | 0–1.6 |
Read that top to bottom: for a customer on Yandex, being in the right directories decides almost everything; for one on a memory-only engine, no listing reaches the answer at all. The practical rule — optimise for the engine your buyers actually use, not for "AI" as a single thing.
And the leading source changes by niche
There is no universal source list to "get into". The platform that feeds the most answers is different in each field: in medical niches it's vertical catalogues like ProDoctorov and city medical boards; in real estate, a regional property portal; in car rental, local aggregators alongside 2GIS. The only way to know which platforms feed your field is to measure that field — which is exactly what we publish, by niche, as open data.
Provenance: these shares are measured on buyer questions in the Russian-speaking market, where almost no one else measures Alice and Yandex. The mechanism — web-searching engines cite maps and directories, memory engines cite nothing — is the same across markets, but the exact platforms are local. Open data, by niche and engine: /barometr.json.
What this means for getting named
- Find the sources first, then act. Don't guess — a car-rental playbook won't help a clinic. The platforms that feed answers in your niche are measurable, and your report lists them.
- Be present and clearly described there. A listing the model reads is the entrance to its answer; a description that ties you to the category, in a customer's words, is what gets you named rather than just listed.
- Match the engine to your audience. If your buyers ask Yandex, directories are nearly the whole battle; if they use a memory-only assistant, no source reaches the answer and the lever is elsewhere.
Being in a source is the entrance to the answer — and not a guaranteed place in it; no one honestly promises a position. We don't sell timelines or a flattering score — we show which platforms feed answers in your niche, how each engine uses them, and whether you're there.
See which sources decide your niche
Free, no sign-up: your buyers' questions across several engines, the platforms their answers are built from, how much each engine leans on directories, and whether your brand appears in those sources.
Run a free checkCommon questions
- Which sources do AI assistants cite most?
- Maps and directories — 2GIS, Google and Yandex Maps, Zoon, vertical catalogues like ProDoctorov — far more than brand sites. A company's own website is among the sources in only about a third of answers, and the leading source changes by niche.
- Do all engines cite the same sources?
- No. Reliance on directories ran from 86.5% of answers on Yandex's search and 53.1% on Perplexity down to 4.6% on GigaChat and 0% on a memory-only engine, in one measured niche. Web-searching engines cite listings; memory engines cite almost nothing.
- If I'm in the directories, will the AI name me?
- Presence in the sources is the entrance to the answer, not a guaranteed place in it. Being listed where the model looks is necessary; a clear description against the category is what gets you named. We don't sell a position — we measure whether you appear.