Dan Cristian

When someone asks ChatGPT "which online bookstore should I use for children's books?", they get one answer, not ten blue links.

So we measured who is inside that answer.

We put 18 real purchase questions to five AI engines — ChatGPT, Google Gemini, Perplexity, Google AI Mode and Google AI Overviews — in Romanian, from Romania, across two complete weekly runs. No question contained a brand name. Every brand that appears was named by the engine on its own.

The result: 180 analysed answers, 247 brand mentions, 21 tracked brands, 422 distinct cited domains.

The headline finding is the one we did not expect.

58% of answers name nobody

Out of 180 answers, only 75 (42%) mention any of the 21 major brands in the market. The other 105 respond with book titles, authors and small sites — but never say where to buy.

And of those 75, only 19 recommend a single brand. The rest share the stage between 3.3 brands on average.

Six of the eighteen questions gathered three or fewer brand mentions in total, across every engine, in both weeks:

Question territory Total brand mentions
Business books for entrepreneurs 1
Books for a teenager who barely reads 1
Motivational vs. psychology 1
Easy English reading 2
Cooking 3
Nutrition and health 3

A third of the market's purchase territory currently belongs to nobody.

Being mentioned is not the same as being recommended

Here is the part that changes how you should measure this.

The market leader, Libris, has 37 mentions — 20.6% of all answers. But the interesting number is where those mentions land:

Position 1 Positions 2–3 Positions 4–7 Position 8+
Libris (37 appearances) 24 (65%) 8 (22%) 5 (13%) 0
BookZone (24 appearances) 3 (12%) 12 (50%) 8 (33%) 1 (4%)

Libris doesn't just appear most often — it appears first two times out of three.

Weight by position (1st = 100%, 2nd = 50%, 3rd = 33%…) and the hierarchy rewrites itself. Editura Trei is 6th by raw mentions but 3rd by weighted visibility: it appears rarely, but when it does, it appears at the top. Few territories, owned.

If you are tracking AI visibility with a mention counter, you are measuring the wrong thing.

Each engine is a different channel

Mention rate per brand, as a share of each engine's 36 answers:

Brand ChatGPT Gemini Perplexity AI Mode AI Overviews
Libris 13.9% 30.6% 16.7% 13.9% 27.8%
Litera 13.9% 25.0% 5.6% 27.8% 11.1%
Cărturești 11.1% 25.0% 8.3% 16.7% 19.4%
BookZone 8.3% 8.3% 11.1% 22.2% 16.7%
Humanitas 5.6% 19.4% 5.6% 5.6% 0%

Two things worth pulling out:

Engines differ wildly in generosity. Gemini produced 78 mentions across its 36 answers — 2.2 brands per answer. Perplexity produced 30 — 0.8 per answer. Statistically, one recommendation in Perplexity is worth more than two and a half in Gemini, because there are far fewer places on the podium.

Same company, opposite verdicts. Humanitas sits at 19.4% on Gemini and 0% on Google AI Overviews. Same brand, same content, two Google products, completely different retrieval logic.

Ghost brands: cited but never named

This is the finding with the most direct practical value.

Being cited (your site feeds the answer) and being named (the answer recommends you) are different currencies. Cross-reference them and two species appear:

Brand Cited as source Named as brand Diagnosis
Târgul Cărții 18 4 ghost
Librarul 14 1 ghost
eMAG 15 9 under-named
Editura Trei 5 17 memory
Nemira 1 10 memory
Humanitas 0 13 memory

Ghost brands get crawled, their listings and prices become raw material for the answer — and then the engine recommends somebody else. Their content is credible; their entity is not. In practice this points at weak entity signals: incomplete structured data, no consolidated identity (sameAs, profiles, reviews), a generic name that is hard to bind to a domain.

If your site is cited but your name is missing from recommendations, the fix is not more content. It's entity work — and it's the fastest win in this whole field.

Memory brands are the mirror image. Humanitas is recommended 13 times with zero citations of humanitas.ro in the entire window. It lives purely off the models' parametric memory — decades of cultural presence. Flattering, but fragile: wherever the answer gets built from crawlable sources and your site isn't among them, you drop out of the conversation.

Reddit is the market's second source

422 distinct domains were cited across the 180 answers. The top of that list:

Domain Answers citing it
carturesti.ro 46
reddit.com 34
litera.ro 34
libris.ro 34
bookzone.ro 29
targulcartii.ro 18

Reddit sits above Facebook (11), YouTube (6) and every media publication in the set. When somebody asks an AI which bookstore to use, part of the answer comes from readers arguing with each other in threads no brand controls.

There is also a shadow market in that list — clb.ro, librex.ro, librariadelfin.ro, booknation.ro — sites with almost no public recognition that show up constantly. Models don't cite by fame. They cite what they can crawl and parse.

The rankings move weekly

Same 18 questions, same 5 engines, seven days apart:

Brand 20 Jul 27 Jul Change
Editura Corint 2 8 ×4
elefant.ro 9 13 +44%
Humanitas 5 8 +60%
Libris 21 16 −24%
Editura Trei 12 5 −58%

elefant.ro went from 7.0% to 10.9% share of voice in one week — a jump that in organic Google would take months.

The right word for this isn't chaos, it's plasticity. Answers regenerate at every query, from re-crawled sources and constantly updated models. For today's leaders that means the advantage needs maintenance. For everyone else it means the window to enter is permanently open — unlike classic SEO, where positions 1–3 are set in concrete by years of accumulated authority.

Reproduce it

The per-engine rates are n/36 rounded to one decimal, so the underlying counts reconstruct exactly:

import csv

rows = list(csv.DictReader(open("piata-carte-ai-2026-07.csv", encoding="utf-8")))
engines = ["chatgpt_pct", "gemini_pct", "perplexity_pct",
           "google_ai_mode_pct", "google_ai_overviews_pct"]

by_engine = {e: sum(round(float(r[e]) * 36 / 100) for r in rows) for e in engines}
print(by_engine)                # 33, 78, 30, 56, 50
print(sum(by_engine.values()))  # 247

Enter fullscreen mode Exit fullscreen mode

247 also equals the sum of mentiuni_total, and the sum of the two weekly runs (128 + 119), and the sum of mentions across all 18 questions. Four independent axes, one number.

One trap: the dataset uses two denominators. Mention rates are against 180 answers (per-engine: 36). Citation percentages are against 172 — only the answers that actually contain citations. Eight answers were generated with no links at all.

Get the data

Everything above comes from three CSV files, published CC BY 4.0 — free to use, including commercially, with attribution:

All four carry byte-identical files. Take them and check us.

Limits, stated up front

Two runs over two weeks — enough to see volatility, not enough for trends. 18 questions, not the full purchase universe. 21 tracked brands, not the whole market; smaller players appear only in the citation data. Roughly 10 answers per question, so single-question deltas are noise. Mentions are not conversions. Models update continuously — these figures describe engine behaviour between 18 and 31 July 2026, not a permanent truth.

All brands named here were named by the AI engines, not by us. This is market research, not an evaluation of anyone's products or service.

Full study, with all 18 questions and interactive charts: websem.ro

Episode 1 of the series asked ten LLMs a single identical question about luxury jewelry and got 29 different brands back, with 18% agreement between models — DOI 10.5281/zenodo.21724399.