Who AI trusts for local recommendations: an OmniCite study

By Karim Nabhan, founder of OmniCite. Data: Oct 10 to Nov 1, 2026, with a Nov 3 update.

Key findings

We asked four AI engines 83 local questions about creators, food, beauty, skin care, family outings and wellness in Calgary, Seattle and Toronto, and recorded every source they cited. Across 686 answers from Oct 10 to Nov 1, 2026, the engines cited 832 different websites. The data comes from Samoony, OmniCite's AI visibility product for local creators.

• Specific questions break the directory monopoly. Every broad "top creators in [city]" answer cited an influencer directory (61 of 61). For local, specific questions, almost none did (1 of 137). Local publications, business sites and booking pages took their place.
• The engines rarely agree. When two engines answered the same question, they shared no cited website in 46% of cases (224 of 485 pairs). On average they overlapped on 8% of their sources.
• Engines search in short, specific phrases. Their behind-the-scenes searches averaged 6 to 7 words and carried the asker's constraints: neighbourhood, budget, access needs, audience. None of 456 mentioned follower counts.
• Small creators are invisible until someone asks for them by name. Five local creators with small followings were named in 6 of 61 broad answers, 7 of 249 category answers and 0 of 137 specific local answers, but in 23 of 36 answers when someone asked for them by name.

1. The more specific the question, the less directories matter

Influencer directories such as FeedSpot, influData, Embold and Modash rank people by follower count. They own the broad question, and they disappear as the question gets specific. The table shows the share of answers that cited at least one source of each type.

The most cited websites overall were FeedSpot (305 answers), Instagram (164), influData (135), Embold (113) and Modash (96). On specific local questions the leaders were Google listings, Reddit, Daily Hive, ParentMap, Leavenworth's tourism site, Calgary Locale, Avenue Calgary and Seattle's Child, and most of the remaining sources were the businesses' own websites.

Reading: a business or creator does not need a big following to be cited on a specific question. It needs a page that answers that question.

2. The engines rarely agree on sources

When two engines answered the same question in the same run, they shared no cited website in 224 of 485 pairs (46%). Their average source overlap was 8%. Each engine has its own favourite places to look:

Reading: being visible on one engine says little about the others. A Google AI Mode answer leans on Instagram profiles, while Copilot and ChatGPT lean on lists and articles. ChatGPT gave an answer with no citation at all in 11 of 121 answers (9%).

3. How engines actually search

AI engines do not search your question as typed. They rewrite it into a few short web searches first, called fan-out. We captured 456 of these searches from ChatGPT and Gemini: 191 while testing 54 real-style questions, and 265 behind our tracked questions.

• They averaged 7 words on the real-style questions (6 on tracked ones) and kept the asker's constraints: "hot pot Calgary under $30 per person Chinatown NW", "wheelchair accessible pumpkin patch Seattle smooth paths", "Calgary bridal makeup artist mature skin soft glam".
• On the real-style questions, 14% put a business or person's name in quotes. Engines first find candidates, then search each name to check it.
• 10% named a neighbourhood and 8% a price or budget. Only 9% used "best" or "top".
• None mentioned follower counts.
• 35 searches added the year 2023 to questions asked in October and November 2026, more than twice as often as they added 2026 (14). Pages that show a current date may match better. This is an inference, not a measurement.

Reading: the searches behind an AI answer look like the questions real customers ask, not like rankings. A page that states the same specifics (area, audience, budget, access, hours) is what those searches find.

4. Small creators are invisible until someone asks for them

We tracked five local creators with small followings: a food reviewer, a makeup artist and a dermatologist in Calgary, a mobility therapist in Toronto, and a family guide in Seattle.

When engines were asked about a creator by name, they relied on the creator's own YouTube, TikTok and Instagram profiles. Where no clear profile page existed, some answers fell back on people-search sites such as IDCrawl, RocketReach and Social Catfish.

Reading: engines know these creators exist, but nothing connects them to the questions their audience asks. Samoony began publishing pages for these creators on Oct 27. None had been cited by Oct 28; the first citations came on Nov 1.

Update, Nov 3. samoony.com was cited for the first time on Nov 1, in 5 Gemini answers. In the 3 answers where Gemini read a creator's page, it named that creator twice. Across all tracked answers since Oct 20, the pages that led to a small creator being named were list pages and pages about that creator, such as a city guide, an article or a local business blog. Individual directory profiles were never cited.

What this means for local businesses and creators

1. Stop chasing the broad question. "Top influencers in Calgary" is answered from follower-ranked directories on every engine. Paying for a spot there rarely changes the outcome: paid cards usually sit outside the numbered ranking, and on at least one major directory the paid card is added by script, so AI engines reading the page never see it.
2. Write pages that state the specifics. Neighbourhood, who it is for, price, hours, access, and what makes it different. Those are the words the engines' own searches use.
3. Show up on the local sites engines already trust. On specific questions, local publications were cited in 39% to 42% of answers. A mention there carries more weight than another directory listing.
4. Check each engine separately. Engines share few sources, so a business can lead on Google AI Mode and be absent from ChatGPT.
5. Keep pages current. Engines often add a year to their searches, so pages that show when they were last checked may match more often.

Method and limits

• Data: 686 AI answers recorded by Samoony's tracking between Oct 10 and Nov 1, 2026, plus a Nov 3 update. Engines: Google AI Mode, Bing Copilot, ChatGPT and Gemini, each queried through its public search or API interface. Answers came from monthly tracked questions for five creators plus creator-list lookups for city guides. Failed engine calls were excluded.
• Location: questions were asked as a searcher in the creator's city, as of Oct 25. Earlier Bing Copilot answers were effectively asked from the United States.
• Citations: a website counts once per answer, however many of its pages were cited. Source types were assigned by hand from the top cited domains, and the remaining domains count as "other". Answers from internal test accounts are excluded throughout. Seattle and Toronto publications were added to the local publication group for this update.
• Fan-out: 456 searches from ChatGPT and Gemini only: 191 captured while researching six creators, and 265 from tracked questions. Google AI Mode and Copilot do not expose their searches.
• Limits: three cities, 83 distinct questions and 23 days. Calgary accounts for 77% of answers, and Gemini contributed 67. Treat the percentages as direction, not precision. Each question was asked once per engine per run, and AI answers vary from run to run.