Nicolas Sitter
September 2026AI Search Studies

AI Search for Gelato in Rome (2026):the density law breaks

TL;DR: I asked three AI engines where to eat gelato in Rome 24 different ways, in English and in Italian, from a US address and an Italian one — all 288 captures landed, a complete grid. For five straight cities, one number predicted how often ChatGPT cites a business’s own website: how many businesses own one. Rome’s gelaterie own plenty (41.1% of the registry) and ChatGPT cited them 6% of the time — a quarter of what the curve predicted. 70% of its citations went to guides about gelato instead, a layer where sites that only rank gelaterias out-cite the gelaterias themselves. The consensus pick is Gelateria del Teatro, named in 97 of 288 answers. Nine domains are cited by all three engines and seven of them are guides — the gelaterie get two.

Published September 16, 2026 · data captured September 16, 2026
6%
ChatGPT own-site share at 41.1% registry density — off the curve
97/288
answers naming Gelateria del Teatro, across all three engines
70%
of ChatGPT’s citations go to the guide layer
Read the Report

Executive Summary

Six cities in, this series had one finding solid enough to call a law. Rome is the city where it stops working — and the reason turns out to be more interesting than the law was.

The pattern, built across Mexico City, Seoul, Istanbul, Buenos Aires and Helsinki, said: the share of a vertical that owns a web domain predicts the share of ChatGPT’s citations that land on venue-owned sites, in strict rank order. Rome’s registry measures 41.1% own-domain density — second only to Helsinki — which put the prediction somewhere between Buenos Aires’ 25.3% and Helsinki’s 45.5%. The measurement came back 6% (26 of 433 citations).

Where did the rest go? 70% of ChatGPT’s citations landed on the layer that writes about gelato: Rome city magazines, Gambero Rosso’s Tre Coni rankings, global food press, tour-operator blogs — and a small industry of sites whose entire purpose is ranking gelaterias. There are at least two of them. gelatomaps.com alone drew 39 citations — more than any actual gelateria’s own website, in a city that has no shortage of actual gelaterias. Mexico City showed a guide takeover once before, but its taquerías barely had websites to skip. Rome’s gelaterie have 190 of them, and the engines read the guides anyway. Owning a domain looks necessary but no longer sufficient: where a century-old food canon carries a professional press, that press wins the retrieval.

By the numbers: 288 of 288 captures landed, carrying 1,500 cited URLs and 1,409 extracted venue mentions, 95.2% resolved against a 547-row registry — the best resolution rate this series has managed. The absences: No absences: every one of the 288 fired captures landed, which is rarer in this series than it sounds and means nothing here rests on a partial engine. Two engines were fired and are not in the study at all — see the methodology for which and why.

Terminology, as on every page in this series: citations are the URLs an engine attaches under its answer, mentions are the venue names inside the answer text. Source analysis uses the former, the leaderboard the latter.

Section 1

Six cities on the curve, and one of them falls off

Every previous city ranked the same on both bars: more venues with domains, more ChatGPT citations to venue domains. Rome ranks second on the grey bar and fourth on the pink one.

density-law-six-cities-rome

The registry behind the grey bar: of 462 real gelaterie, 257 (55.6%) list some web presence and 190 (41.1%) an own-domain site — Rome’s gelato scene is properly webbed, from giolitti.it — a domain carrying a shop founded in 1900 — to the one-page sites of neighborhood spots. On the five-city curve, that supply level pays out around a quarter to a third of ChatGPT’s citations.

Rome got 6%. Not because ChatGPT went quiet — its 433 citations are the second-largest column in the study — and not from a classification quirk: I re-audited every domain in its unclassified tail by hand after two gelato guide sites initially masqueraded as venue sites, and the corrected number is what you see. The citations exist. They just point one layer up.

What the break teachesThe law was really measuring what the web offers the retrieval step. In CDMX and Seoul, venues without domains left the engines nothing to cite but guides. Rome shows the inverse case: when a vertical is famous enough to sustain a professional guide press, that press crowds venue sites out of the citations even though they exist. Supply sets the ceiling; competition decides how far below it you land.
Section 2

The machines read the critics

303 of ChatGPT’s 433 citations (70%) land on editorial, guides, food press, blogs and tours. Gambero Rosso — whose Tre Coni award decides which gelaterie hang a plaque by the door — collects 58 citations across three engines. And a niche the series hadn’t met yet: 65 citations go to the dedicated gelato-ranking sites — gelatomaps (39 on its .com alone) and con-gelato.it (22) — sites that exist only to rank gelato.

source-mix-by-platform-gelato-rome

Columns are largest-remainder rounded to sum to exactly 100.

2 of 9
The domains all three engines agree on

Nine domains are cited by every engine in the study. Seven of them write about gelato — Dissapore, Gambero Rosso, the two ranking sites, a Roman food blog, a tour operator, a city guide. Two belong to gelaterie: giolitti.it (15) and gelateriadeigracchi.it (4). The consensus layer of this city’s gelato web is, almost entirely, other people writing about it.

25%
Gemini reads somebody’s blog

A quarter of Gemini’s 348 citations go to personal food and travel blogs — valeriacastiello.com, mominitaly.com, kissfromitaly.com, a long tail of individual voices. Helsinki first showed this Gemini habit; Rome confirms it at scale. Its gelateria-website share is 4%, even lower than ChatGPT’s.

0
Reddit stays gone

Rome makes it three consecutive studies with ChatGPT’s Reddit citations at or near zero: 0 of 433 here, after 0 in Buenos Aires and 0.4% in Helsinki. Study-wide, reddit.com drew 15 citations from three engines — none of them ChatGPT, whose 17–20% Reddit era earlier in this series is now archaeology.

Section 3

The consensus scoop: Gelateria del Teatro

A cobblestone side street off Via dei Coronari, sage-and-honey flavors, and the study’s broadest agreement: named in 97 of 288 answers (33.7%), by all three engines, in both languages. Behind it, Otaleg at 93 and Fatamorgana Monti at 88 — and only then Giolitti, the 1900 institution the guidebooks would have led with, fourth at 85.

gelato-leaderboard-answer-mentions

Which engine scoops for which gelateria

For each gelateria, the percentage of an engine’s answers that name it, raw counts beneath. The podium is shared; the favorites are not. ChatGPT leans del Teatro and Come il Latte, Gemini leans Fatamorgana and Otaleg, and AI Mode spreads itself across Giolitti and Otaleg.

#GelateriaAI Mode96 capturesChatGPT96 capturesGemini96 captures
1Gelateria del Teatro26%(25)46.9%(45)28.1%(27)
2Otaleg33.3%(32)26%(25)37.5%(36)
3Fatamorgana Monti26%(25)19.8%(19)45.8%(44)
4Giolitti33.3%(32)28.1%(27)27.1%(26)
5Frigidarium31.2%(30)17.7%(17)18.8%(18)
6Come il Latte13.5%(13)34.4%(33)19.8%(19)
7Gelateria dei Gracchi10.4%(10)20.8%(20)30.2%(29)
8Gelateria La Romana19.8%(19)15.6%(15)9.4%(9)
9gelateria dell'angeletto15.6%(15)18.8%(18)7.3%(7)
10Gelateria Fassi10.4%(10)16.7%(16)12.5%(12)
11Neve di Latte12.5%(12)3.1%(3)18.8%(18)
12Fiordiluna10.4%(10)13.5%(13)9.4%(9)

Every column is 96 captures, so the percentages compare directly — no engine here is running on half a grid.

Top 12 gelaterie by answer mentions — click a marker for per-engine counts

1Gelateria del Teatro2Otaleg3Fatamorgana Monti4Giolitti5Frigidarium6Come il Latte7Gelateria dei Gracchi8Gelateria La Romana9gelateria dell'angeletto10Gelateria Fassi11Neve di Latte12Fiordiluna
Section 4

The mention board crowns a workshop; the citation board crowns a chain

Scored by answer mentions, Gelateria del Teatro wins with 97 and a citation score of 47. Scored by citations, the winner is Gelateria La Romana — score 127, mentions rank #8. La Romana is a Rimini-born chain with eight Rome branches on one domain, and that’s the whole trick: every branch’s map presence and every gelateriaromana.com citation aggregates into a single brand total no single-shop artisan can match.

The two boards have now disagreed in four of the six cities that measure both. Rome’s version has a sharper edge than most, because the disagreement is structural: a citation-counted dashboard would tell you Rome’s gelato champion is a national franchise, while what the engines actually say to users, 97 times out of 288, is a two-room workshop near Piazza Navona.

A separate scoring hazard surfaced during the run: 70-plus registry venues list facebook.com as their “website”, and an unguarded citation counter merged them into one phantom brand. The guard this series added after its Marseille incident intercepted it; it reaches no board on this page.

#1 by answer mentions
Gelateria del Teatro
97 of 288 answers · all 4 engines · cite score 47
#1 by citation score
Gelateria La Romana
cite score 127 · 8 branches, one domain · 43 mentions (#8)
Why publish both boardsMentions measure what a user hears; citations measure whose pages the retrieval step read. When a chain’s domain does the aggregating, the citation board inherits the chain’s org chart. Anyone selling “AI visibility scores” off citation counts alone would rank La Romana above del Teatro — and miss what every engine actually recommends. The score would be real. It would simply be answering a question nobody asked.
Section 5

English and Italian agree on the shops — and split on the flavors

The control prompt overlaps 43% between languages — the same number Buenos Aires measured, well clear of the Istanbul/CDMX zeros. Then the pattern gets granular: ask about a place and the languages converge; ask about a flavor and they diverge to almost nothing.

en-vs-local-control-overlap-series-rome
EN vs IT top-5 overlap by prompt (ChatGPT × IT proxy, map-widget top-5) — 21 comparable templates
Prompt templateEN vs IT top-5 overlap
grattachecca (bleed control)100% (5/5)
district: Monti67% (4/6)
most famous gelaterias67% (4/6)
control (best gelato)43% (3/7)
avoiding tourist traps43% (3/7)
district: Pantheon43% (3/7)
district: Trastevere43% (3/7)
oldest historic gelaterias43% (3/7)
kids persona43% (3/7)
where Romans actually go43% (3/7)
best desserts (bleed control)25% (2/8)
district: Prati/Vatican25% (2/8)
gourmet / foodie persona25% (2/8)
cheap but good25% (2/8)
gluten-free for celiacs25% (2/8)
tourist persona11% (1/9)
artisanal gelato11% (1/9)
chocolate gelato11% (1/9)
sorbet & fruit11% (1/9)
vegan / dairy-free11% (1/9)
pistachio gelato0% (0/10)

Overlap = shared venues ÷ distinct venues across both top-5 lists (Jaccard). Top-5 read from ChatGPT’s map carousel, present in 90 of 96 ChatGPT captures. Three templates (late-night, recent openings, spotting-artisanal advice) returned no comparable pair and are not listed. Median across the 21: 25%.

The cleanest agreement in the study is also its most Roman: ask for grattachecca — the hand-shaved ice sold from riverside kiosks — and English and Italian return the identical five kiosks: Sora Maria, Sora Mirella, Alla Fonte d’Oro, Chiosco Testaccio, Er Chioschetto. A hundred-year-old street tradition with almost no web presence of its own turns out to be the one thing both languages’ corpora describe identically. Five kiosks, two languages, no disagreement — the steadiest result in the study belongs to the vendors least likely to have a marketing budget.

The flavor prompts sit at the other end: pistachio overlaps 0%, chocolate, vegan, sorbet and artisanal 11% each. Everyone agrees on where gelato matters in Rome; which counter does the best pistachio is apparently a question each language answers from its own experts. The tourist-persona prompt (11%) splits too — English answers cluster near the sights, Italian answers send visitors a rione further out.

The coupling is the strongest since TokyoItalian prompts put 29.9% of their citations on .it domains, English prompts 14.7% — a 2.03× coupling, ahead of Buenos Aires’ 1.33× and behind only Tokyo’s 5×. Italy’s food web is big enough that the engines can stay domestic when asked in Italian; they just don’t have to.
Section 6

The twelve domains that answer gelato questions

google.com sits out of this table: its 509 citations are AI Mode referencing itself. What remains reads like a food-media masthead — two national magazines, a city magazine, two dedicated gelato-ranking sites — before the first and only venue domain appears at #10.

Most-cited non-Google domains across 1,500 citations (normalized), 2026-09-16
DomainEnginesCitationsWhat it is
dissapore.com3/368Italy’s irreverent food magazine — the most-cited real-web domain in the study
gamberorosso.it3/351The Italian food authority whose Tre Coni ranking is the gelato world’s Michelin
romeing.it1/347English-language Rome city magazine — and ChatGPT’s private favourite; no other engine cites it
gelatomaps.com3/339A site that exists solely to rank gelaterias — and out-cites every gelateria in Rome
theinfatuation.com1/326US restaurant-guide platform
con-gelato.it3/322A second dedicated gelato guide. Two of these exist. Both out-cite the shops
tripadvisor.com2/322The reviews everyone says they ignore
tripadvisor.it2/321And its Italian twin, cited separately
walksofitaly.com2/320A tour operator’s content arm
puntarellarossa.it3/316A Roman food blog all three engines read
lacucinaitaliana.it2/316Italy’s oldest food monthly
giolitti.it3/315One of only two gelateria domains every engine cites — a 1900 shop’s own site

Engine counts out of the 3 in the study; a domain cited by one engine reads 1/3. Gambero Rosso’s domains total 58 citations including its international edition; the table row is gamberorosso.it alone. Thirteen rows here rather than twelve: giolitti.it and puntarellarossa.it tie at 15and 16, and cutting one to round the table would have hidden the only gelateria domain in the consensus set.

Section 7

Where the machines put gelato on the map

The registry runs from the centro storico’s tourist corridor out through Prati, Monti and Trastevere to the beach gelaterie of Ostia — 462 mapped venues in all. The top 12 cluster tightly around the historic center, del Teatro and Frigidarium within a ten-minute walk of each other. One test this city can’t run: the per-rione accuracy check, because Rome’s Google Maps rows carry an empty neighborhood field — the same seed-granularity artifact Amsterdam and Berlin hit, disclosed here rather than charted as a finding.

All 462 mapped registry gelaterie across Rome

Section 8

What the instruments did this run

Two engines were fired and left out

Copilot cannot be queried from an Italian IP at all — a standing restriction on which countries each engine can be reached from, confirmed with a direct probe before being written down as a limit. It therefore ran only the US half of the grid, 48 of 96. Perplexity landed 14 of 96. Neither is in this study: a study whose central question is whether the language you ask in changes the answer cannot be run on an engine that can only be asked from one country, and half a grid will not compare with three whole ones. Both sets of captures remain in the database; they are simply not this article.

AI Mode’s mirror cracked open — slightly

70.8% of AI Mode’s 719 citations are google.com self-references — down from the flat 100% Buenos Aires measured after the September searchviewer change (decoded here). A partial thaw, not a recovery: the metric stays retired, the number is reported for the record.

Thirteen answers recommended nothing

13 of 288 captures name no venue: 6 AI Mode, 5 Gemini and 2 ChatGPT — mostly the how-to-spot-real-gelato advice prompts answered with criteria instead of names, which is arguably the correct reading of the question. ChatGPT skipped web search in only 3 of 96 captures, and its map carousel rendered in 90 of 96, delivering 1,228 map entities.

If you run a gelateria in Rome

1

In Rome, the guides are the citation. An own domain still keeps you aggregatable as a brand and still carries your hours and your address, but in this city it buys almost none of what the engines quote: 6% of ChatGPT’s citations, 4% of Gemini’s. The pages they actually read are Dissapore, Gambero Rosso, Romeing and the gelato-ranking sites — coverage there is the citation.

2

The Tre Coni is machine-readable prestige. 58 citations across three engines make Gambero Rosso the closest thing this study found to a direct lever: the venues its rankings name are the venues the engines name.

3

A chain’s domain out-scores your shop on citation dashboards — know which metric you’re being sold. La Romana’s eight branches make it the citation-counted #1 while ranking seventh in what engines actually say. If an agency pitches you a visibility score, ask which board it’s reading.

4

Flavor pages are the open ground. The languages disagree most on pistachio, chocolate and vegan — exactly the prompts where a well-documented specialty could still move an answer. The “best gelato in Rome” canon is set; “best pistachio” is not.

Conclusion

Rome was picked to be the sixth point on a curve and became the first counterexample instead. The gelaterie did what the series’ own advice says — nearly half own a domain — and ChatGPT read Gambero Rosso anyway. What five cities made look like a supply problem turns out to have a second condition attached: venue websites win citations where nothing more authoritative writes about the vertical. Rome’s gelato has been written about, professionally and in two languages, since before the web existed.

The rest of the picture is the series’ recurring cast in Italian dress: Gemini off in the blogs, Reddit gone from ChatGPT a third study running, one workshop (del Teatro) carrying the consensus while a chain (La Romana) carries the citation count. And one genuinely Roman result: the only prompt where both languages agree completely is about shaved ice sold from kiosks that barely have websites at all.

Next in the series: a vertical where no professional guide press exists, to test whether the density law holds when nobody else is writing.

Methodology

The grid

24 prompt templates (control, flavors and styles, dietary, personas, districts, price, canon, plus grattachecca and dessert entity-bleed controls) × 2 languages (English, Italian) × 4 engines (ChatGPT, Gemini, Google AI Mode) × 2 proxy countries (US, IT), fired 2026-09-16. 288 of 288 captures landed — every engine ran every cell. Two further engines were fired and excluded: Copilot, which cannot be queried from an Italian IP and so managed only the US half of the grid (48 of 96), and Perplexity, which landed 14 of 96. Neither is in the numerator or the denominator of anything on this page.

The registry

500 Google Maps rows from a gelato-specific search (“gelateria”, English locale), classified by a category map: gelato and ice-cream shops kept real, alongside dessert and chocolate shops per this series’ established precedent — the pull is gelato-scoped, and Google files Venchi under Chocolate shop. Bars, pasticcerie, cafés and frozen-yogurt shops classified out but kept resolvable. Three name-targeted recovery passes added 47 rows: canon venues the 500-row cap missed — Giolitti among them, which is to say the seed for a study of Roman gelato initially omitted the most famous gelateria in Rome, a useful reminder of what a bounded pull does to a long-tail city — plus venues the answers surfaced, the grattachecca kiosks and the dessert-probe pastry shops. Final registry: 547 rows, 462 real gelaterie, website layer recomputed at publish time.

Extraction and resolution

Venues named in each prose answer were pulled out by a language model — a different extraction model than this series usually runs, under the same rule set, with every one of the 1,409 extracted names mechanically verified to appear verbatim in its source answer (zero failures). Names resolved to the registry in strict precedence: exact normalized match, fuzzy (≥0.86), Italian-diacritic folding, a hand-checked 44-entry alias table, then unique token containment — 95.2% of mention rows resolved, the series’ best. Duplicate Google listings of one venue (Otaleg’s two, Günther’s three) merge canonically. Deliberately unresolved: bare “Crema” (ambiguous across four registry rows), “Giolitti a Testaccio” (a map entry with no live listing — possibly a ghost branch), and supermarket product lines.

Counting rules

Citations are per-capture deduplicated URLs; source buckets are keyword-and-domain based, with venue identity decided by registry website domains plus a gelato-keyword heuristic that explicitly excludes gelato-named guide sites — two of them (gelatomaps.com, con-gelato.it) initially read as venue websites and would have tripled the own-site share had they stayed. Social platforms and marketplace pages never count as a venue’s identity domain. The leaderboard counts each venue once per capture; the citation score counts ChatGPT map presence plus citation-domain matches. Source-mix columns are largest-remainder rounded to sum to exactly 100. All headline stats are published as summary.csv.

Frequently Asked Questions

Playground study