How AI recommends wall art: we analysed 248 AI answers
We asked ChatGPT, Perplexity, Gemini and Google AI Overviews where to buy wall art — 248 times, across four countries. The engines barely agree: only 3% of cited sources are shared by all four, and 27% of buying answers recommend no brand at all. Here is who AI actually recommends, and why.
More and more people start a purchase by asking an AI assistant, not a search box. So we ran the obvious experiment: we asked the four engines most buyers now use where to get wall art, 248 times, and logged every brand they named and every source they cited. The picture that came back is far messier — and far more open — than classic search ever was.
The study in one paragraph
In July 2026 we ran a frozen panel of 62 real buying questions ("where can I get a photo printed on metal?", "best metal print companies", "Displate alternatives"…) through the four AI engines most people now ask — ChatGPT, Perplexity, Gemini and Google AI Overviews — in the UK, Poland, Germany and France. That produced 248 answers, of which 227 passed quality control. We logged every brand mentioned and every source cited: 523 distinct domains in total. We run a metal-print studio, so we published the leaderboard with our own brand-name questions excluded — the numbers below are what the engines say when nobody asks about us specifically.
Finding 1 — "AI search" is four different search engines wearing one name
Of the 523 domains the engines cited, only 16 — three percent — were cited by all four engines. Seventy percent appeared in exactly one engine's answers and nowhere else.
Classic search never worked like this. Google and Bing disagreed at the margins; these four systems build their answers from largely separate source sets. For anyone who sells online, the practical consequence is blunt: being visible to ChatGPT says almost nothing about whether Perplexity, Gemini or Google's AI Overviews know you exist. There are four doors now, and they open with different keys.
Four engines, four largely separate source sets: 70% of cited domains appear in just one engine's answers.
Finding 2 — Perplexity outsources brand discovery to communities
We classified every citation by source type. 17.7% of Perplexity's citations point at community content — Reddit threads, YouTube videos, forums. For ChatGPT the figure is 2.2%. That is an eight-fold gap: Gemini sits at 5.7%, Google AI Overviews at 13%.
In other words, when a buyer asks Perplexity what to hang on their wall, a meaningful slice of the answer is written by Redditors and YouTubers, not by any brand or publisher. It is the engine where companies have the least direct control over their own story — and where a single genuinely helpful community thread can out-recommend a marketing budget.
Perplexity leans on community content eight times more than ChatGPT (n = 227).
Finding 3 — communities out-cite every retailer
Across the non-brand panel, YouTube collected 82 citations and Reddit 65 — 147 combined, the largest third-party citation block in the study. The top retailer (whitewall.com) collected 90. No other commercial site came close.
The pattern repeats in miniature for niche publishers: photographytalk.com, a photography blog, earned 16 citations — appearing in AI source sets alongside platforms a thousand times its size. The engines are not weighing traffic; they are weighing topical authority, however small the site that carries it.
Citation counts across the non-brand panel (own-brand prompts excluded). Community platforms (YouTube, Reddit) sit right below the top retailer.
Finding 4 — a quarter of the market is simply unclaimed
27% of all buying answers named no tracked brand at all. On Perplexity it was 35%; even on the most brand-forward engine (ChatGPT, 21%) roughly one answer in five recommends categories, materials and criteria — but no company.
For every brand in this vertical, that is the real headline. The recommendation slot in a quarter of AI buying answers is standing empty, waiting for whoever the engines learn to trust first.
A quarter of AI buying answers recommend no brand at all — an empty recommendation slot (n = 227).
Finding 5 — visibility is per-engine and per-market
The same brand's share of voice swings wildly between engines and countries. In our tracked set, WhiteWall leads mentions overall (78 across non-brand prompts) but converts only 77% of its mentions into citations; Displate converts 53%. The strongest conversion in the panel — 93% — belongs to the smallest brand tracked (ours: 29 mentions, 27 cited), which we read as a note about how engines cite rather than about us: engines cite the sources they actually read, and they read pages built to be read.
Per-market splits diverge again: the leaderboards in Poland, Germany, France and the UK share their top block (YouTube, Reddit, WhiteWall) but the mid-field is dominated by local printers a UK reader has never heard of — labophotos.fr, optimalprint.pl, cewe.de. AI recommendation, like politics, is local.
Conversion reflects how engines cite — they link the pages they actually read. Mention bases differ by brand.
What this means if you sell anything
- Treat the four engines as four channels. Measure each separately; a monthly "ask the engines your own buying questions" panel costs a few dollars in API calls.
- Earn community presence honestly. Perplexity's 8× community weighting means one helpful, disclosed answer in the right thread outworks a page of ads. (Astroturfing gets recognised — by moderators and, increasingly, by the engines.)
- Write pages engines can quote. Direct answers, honest comparisons, real specifications. Our conversion stat exists because extractable pages get cited when mentioned.
- Claim the white space. A quarter of buying answers name nobody. The brands that fill those slots first will look, to the next wave of buyers, like the default.
We run the metal-print studio behind this study — the smallest brand in the panel, and the one whose mentions turned into citations most often. If you would rather see the prints than the data, start here.
Ready to Create Your Metal Print?
Transform your favorite photo into a stunning premium metal print. Free EU shipping over £115.
Methodology
62 buying prompts (17 category, 17 transactional, 13 competitor, 11 brand, 4 informational), authored natively in English, Polish, German and French and frozen with a content hash before collection. Engines: OpenAI Responses with web search, Perplexity Sonar, Gemini with Google Search grounding, Google AI Overviews via SERP capture (the only leg with exact national geo-targeting; the API legs are grounded proxies of the consumer surfaces). 248 answers collected 2026-07-02; 21 quarantined by quality gates (vendor errors, blocked fetches — never silently counted as data); 227 analysed. Every citation URL canonicalised (redirect-resolved, tracking-stripped, deduplicated to registrable domain). Brand mentions matched with word-boundary, alias-aware matching, audited against false positives. The headline leaderboard excludes our own brand-name prompts; full-panel numbers are available on request, along with the prompt list and per-engine tables.
Frequently Asked Questions
ChatGPT named a tracked brand in 79% of buying answers — the most brand-forward engine in our panel. Perplexity was the least (65%), leaning instead on Reddit threads and YouTube reviews.
Rarely. Only 3% of the 523 sources cited in our study were used by all four engines, and 70% appeared in just one engine's answers.
Earn mentions on the sources engines already cite — community threads, YouTube reviews, niche topical blogs — and publish direct, quotable answers on your own site. In our data, mentions converted to citations best for extractable, specification-rich pages.
Perplexity's retrieval favours recent community discussion: 17.7% of its citations in our panel were community sources, versus 2.2% for ChatGPT.
62 frozen buying questions across four AI engines and four countries in July 2026: 248 answers, 227 analysed after quality control, every brand mention and cited source logged and classified. Methodology and prompt list available on request.
