The study was conducted by an italian digital marketing consultant.
Search used to be a list of blue links. Today, the first thing many people see is an AI-written answer: Google’s AI Overview on top of the SERP, Perplexity’s single-paragraph reply, ChatGPT answering with web search turned on. The link list is still there, but for a growing share of queries it is no longer the starting point. It is the footnote.
In this new layout, a business that is not named in the answer is not just lower on the page. It is out of the shortlist.
To see how often this really happens, I ran a controlled study on 100 local commercial queries across one Italian region (Veneto) during October. Three engines, 577 total responses, every source they cited stored and counted. The headline number: in 94 of the 100 queries, Google served an AI Overview. The full dataset is public on Zenodo (CC BY 4.0, DOI 10.5281/zenodo.23676358) if you want to inspect it, replicate it in your own market, or cite it.
What follows is what the data says, how the three engines actually pick who to mention, and the practical checklist I now use with clients.
How we studied it: 100 queries, three engines, one region
I built a set of 100 queries that match how people type into an assistant, not into a keyword tool: “best fish restaurant in Venice”, “emergency plumber in Treviso”, “how much does an accountant cost for a sole trader in Padua”. The set is spread across 10 business categories (restaurants, hotels, hairdressers, dentists, lawyers, accountants, plumbers, car workshops, real estate agencies, photographers) and 5 cities of different sizes (Venice, Padua, Verona, Treviso, San Donà di Piave). It covers four intents: finding the best option, checking the price, asking for a recommendation, checking availability.
Each query was sent to three engines:
- ChatGPT with web search enabled, via the OpenAI Responses API.
- Perplexity via its Agent API with web search.
- Google AI Overview via SerpApi, Italian locale.
Perplexity ran three repetitions per query. ChatGPT ran three repetitions on 39 queries (budget cap on the rest). AI Overview was captured once per query. Total: 577 responses, every citation stored.
Two honest limits: this is a single-month snapshot, and the responses come from the engines’ developer-facing endpoints, which can differ slightly from the user-facing apps. Everything else is in the dataset.
How often does Google show an AI Overview?
Out of 100 local commercial queries, 94 got an AI Overview from Google. Six did not. ChatGPT and Perplexity, as expected, answered all of them.
This is a very different picture from the one you get on broad keyword samples. SeoZoom’s December 2025 study on roughly 30 million Italian keywords measured AI Overview coverage at around 54%, with retail and travel categories well below 10%. The two numbers are not comparable: SeoZoom samples the keyword universe, I sample a specific slice (local, commercial, natural-language). The gap is probably driven by phrasing. Questions that contain “best”, “recommended”, “how much does X cost” seem to pull AI Overviews far more aggressively than short, transactional keywords. If your clients ask Google in sentences, they are already in AI-Overview territory most of the time.
A few patterns inside the 94:
- Category matters. Lawyers, dentists, real estate agents, photographers and plumbers got an AI Overview on 10 of 10 queries. Hairdressers were the lowest at 8 of 10.
- Intent matters a bit less than expected. “Best”, “price” and “recommendation” queries got an AI Overview 24 times out of 25. Availability queries (“open on Sunday evening”) got it 22 out of 25. The AI is slightly more cautious when the correct answer depends on live, changing data.
- Small towns are not skipped. San Donà di Piave, the smallest town in the set, matched the big cities with 19 AI Overviews out of 20 queries. Local AI search is not a metro-only phenomenon.
Explore: How to Use AI to Improve Your Marketing Strategy: A 2026 Step-by-Step Playbook
Which sites do ChatGPT, Perplexity and Google AI cite most?
Four sites are cited by all three engines: ProntoPro, TripAdvisor, Treatwell and TheFork. If your clients operate in categories these platforms cover, a complete, up-to-date profile there is non-negotiable. This is the single most useful finding in the study for anyone doing SEO.
Beyond that small overlap, the three engines pull from very different pools:
- Perplexity leans on directories and aggregators. ProntoPro was cited 239 times, PagineGialle 183, Treatwell 182, TripAdvisor 159.
- Google AI Overview pulls heavily from social. Facebook, YouTube and Instagram are among its most-cited sources, and they are not top-cited by the other two.
- ChatGPT prefers institutional and sector sources. The Michelin Guide, regional health authority websites, bar association sites, the Chamber of Commerce, Normattiva (the official Italian legal database).
Per category, the dominant sources are remarkably stable:
| Category | Top-cited sources (out of 10 queries) |
|---|---|
| Restaurants | TripAdvisor 10/10, TheFork 9/10 |
| Hairdressers | Treatwell 10/10, Fresha 10/10 |
| Dentists | MioDottore 10/10, Doctolib 7/10, DocDental 7/10 |
| Hotels | Hotels.com 10/10, TripAdvisor 10/10, Booking 9/10 |
| Plumbers | ProntoPro 9/10, Cronoshare 9/10, PagineGialle 9/10 |
| Photographers | ProntoPro 10/10, Matrimonio.com 6/10 |
| Real estate | Immobiliare.it 10/10, Dove.it 10/10, Idealista 9/10 |
| Car workshops | PagineGialle 7/10, PagineBianche 7/10 |
| Accountants | ProntoPro 8/10, PagineBianche 7/10, Addlance 7/10 |
| Lawyers | iustlab.org 7/10, avvocatoflash.it 6/10, Normattiva 6/10 |
In regulated categories, official bodies carry weight. For real estate, the Chamber of Commerce appeared in 10/10 queries. For lawyers, the Ministry of Justice site and Normattiva sat next to the legal directories. For dentists, regional health authority websites showed up on urgent and pediatric queries. Where information is formally verifiable, the AI tends to lean on whoever certifies it.
Discover; SEO Tips for Small Businesses in 2026: 12 Proven Ways to Get Found on Google and AI Search
Do AI answers change if you ask twice?
Yes. Ask the same question twice and you get overlapping, but not identical, source lists. On the queries I repeated:
- ChatGPT: the two source lists had on average 40% overlap (Jaccard similarity, across 39 queries, three runs each).
- Perplexity: 51% overlap on the same measure.
A citation is not a position you win and keep. It is a slot you re-earn on every query. Clients used to the stability of organic rankings need to hear this plainly: AI-cited is a probability distribution, not a rank.
How do AI engines pick which businesses to name?
Across the 577 responses, the pattern is consistent enough to describe as a five-step pipeline:
- Parse the intent. “Best fish restaurant in Venice” is read as: recommendation, specific city, specific cuisine.
- Reformulate as web queries. ChatGPT ran at least one web search for every one of its 177 responses, three on average, up to fourteen on complex queries. It rewrites your question into the search the engine would actually need.
- Collect many sources. Perplexity pulled 15 sources per query on average. Google AI Overview pulled around 7.
- Filter down to a handful. ChatGPT cited about 4 sources per response. The businesses it names come from those pages, not from training memory.
- Compose the answer with footnotes. The citations are not decoration. They are the raw material of the reply.
The practical takeaway is less dramatic than the AI-ranking discourse suggests. There is no secret prompt that gets you cited. The engines read a small set of trusted sources per category, pick businesses that appear on them with clear, consistent data, and summarize. The leverage is upstream: on which sources list you, how completely, and how coherently with your own site.
How to get cited by AI: a practical checklist
Based on the five-step pipeline and the per-category source tables above, here is the working checklist:
- Let the right crawlers reach you. The cheapest fix, and the one most often missed. Every AI answer starts with a crawler request; if your robots.txt or a firewall plugin turns it away, nothing you write can be cited. Check that OpenAI’s OAI-SearchBot, PerplexityBot and Googlebot are not blocked (a blanket “block all AI bots” rule copied from a forum is the usual culprit). Note the split: blocking GPTBot or Google-Extended only opts you out of model training, not out of being cited in search answers. Those are separate decisions.
- Audit the AI-cited sources in your category. Pick five queries a client cares about, run them on ChatGPT with web search, on Perplexity and on Google. List the sites cited. That is your real target list, not a generic directory roundup.
- Claim and complete the profile on every cited source. Treatwell for hairdressers, MioDottore for dentists, Immobiliare.it for real estate. Full description, correct categories, services, photos, hours. “Complete” is the keyword; sparse profiles get ignored.
- Keep NAP (name, address, phone) identical everywhere. Site, Google Business Profile, directories, social. Mismatches are the single most common reason an AI refuses to confidently name a business.
- Do not neglect Google Business Profile. It is the backbone of every AI Overview on a local query. Correct category, complete service list, recent photos, reviews answered.
- Review generation, continuous. Not just Google reviews. Platform-specific reviews (TripAdvisor for restaurants, Treatwell for salons) are what the matching aggregator ranks on, and that ranking is what the AI quotes.
- Make your site readable to a machine. Clear service pages, specific locations, FAQ sections that match how people actually phrase the question. “Dentist offering pediatric emergency care in Padua” ranks for the question the AI is reformulating.
- Target sector-specific authority sources. Michelin for high-end restaurants, bar associations for lawyers, health authority sites for medical. ChatGPT in particular rewards these.
- Treat social as a citation surface for Google. AI Overview pulls from Facebook, YouTube and Instagram more than most SEO conversations acknowledge. A business Facebook page with current information is a Google-AI asset.
- Expect variance. A client cited today may not be cited tomorrow. Report on frequency over time, not on a single pass. The 40–51% overlap numbers in this study are the baseline.
- Measure. Pick a monthly cadence: run the same five queries, log citations, flag drops. The tooling to do this at scale is immature; a spreadsheet and 30 minutes a month beats no measurement at all.
Also Read: Website vs Social Media 2026: Which Comes First?
Limits, and an invitation to replicate
Three things I cannot claim from this data:
- Whether a citation drives traffic or revenue. I measured where sources come from, not what happens downstream.
- Whether the same patterns hold outside Italian and outside the Veneto. The method generalizes; the per-category source lists do not.
- Whether the single-month snapshot is representative. AI Overviews change weekly. A second edition is on my list.
If you work on a different market, the dataset is on Zenodo under CC BY 4.0: full query set, per-engine citation counts, per-category source breakdowns, methodology. Replicating this on 100 UK, German, Spanish or US queries is a weekend of work and a genuinely useful addition to the public record. If you do, I would like to read it, and I will link to it.
