Four engines, four different source diets. In the 28 days ending September 2, 2026, the Cheers market baseline ran 32 home-services prompts across 100 metro areas and collected 356,019 citations. Gemini pointed at the contractor's own website in 80.7% of its citations. ChatGPT did that in 48.8%, and sent 44.7% to directories and review platforms instead. Same question, same metros, same week.
That is the gap this article is about: the documentation tells you what each engine can do, and the citation counts tell you what it actually did.
Last verified: September 2, 2026.
Engine documentation
What each AI search engine cites
This is the source model each company documents, not a promise that a specific directory will appear for every local query.
Best for: Web answers with inline citations and a Sources panel
Strengths
- Searches public web sources and may use search partners
- Can use a requested or general location for search
Tradeoffs
- No fixed local-source list is published
- Citations can be incomplete or wrong
Best for: Answers grounded in Google's Search index and ranking systems
Strengths
- Gemini can link public websites or connected content
- AI Mode shows supporting links and can use query fan-out
Tradeoffs
- Not every Gemini response includes sources
- Google does not publish a fixed directory hierarchy
Best for: Live-web answers with source links attached
Strengths
- Searches the live web
- Exposes cited sources and labels some domains
Tradeoffs
- A domain label does not certify each page
- No universal local-source list is published
Best for: Web answers grounded in Bing search results
Strengths
- Shows the Bing query and source list
- Bing can return public web pages and business details
Tradeoffs
- Web search can depend on product and account settings
- A Google result does not establish the Copilot result
The defensible source stack
- Facts you control: The branch website and current business profiles
- Evidence you earn: Customer reviews and relevant third-party coverage
- What to prioritize: The sources visibly cited for the real prompt and market
Source behavior reflects official product documentation reviewed September 2, 2026. No engine guarantees that a specific page, profile, or directory will be cited.
That is the comparison local operators can defend from the engines' own documentation as of September 2, 2026.
ChatGPT Search
OpenAI's help article on searching the web, updated in late August 2026, says ChatGPT search rewrites your question into targeted queries and sometimes hands those to other search providers, sharing an approximate location derived from your IP address so that "restaurants near me" becomes something like "top restaurants San Francisco". It is explicit that nothing about the result is owed to you:
ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.
The same page names actual partners in one narrow case: for restaurant reservations, availability may come from OpenTable, Resy, or Yelp. Outside that, OpenAI does not promise that a Google Business Profile, a Yelp page, or any other directory feeds a local answer. It also tells site owners the one concrete eligibility step, which is allowing OAI-SearchBot to crawl the site.
Gemini and Google AI Mode
Gemini Apps and AI Mode are separate surfaces. Google's Gemini help page is direct about how thin the source trail can be:
Not all responses include related links or sources.
When links do appear, Google says they may include public websites, files you uploaded, or Workspace documents and email if you connected Workspace. A missing Sources button is not evidence that nothing was retrieved.
For AI Mode, Google's Search guide says the feature retrieves from the core Search index and can issue related searches through query fan-out. The response shows supporting web links.
Google's local guidance tells site owners to keep Business Profile information current. That supports treating the profile and the branch website as source truth. It does not establish a fixed citation weight for either one.
Perplexity
Perplexity searches the live web and attaches source links to the answer. Its source-label article, last updated August 7, 2026, describes three domain labels, Government, Academic, and Trusted, and then limits what they mean:
A label describes the website as a whole, not any single article or claim on it.
Perplexity also says most domains carry no label at all, and that the absence of one is not a negative judgment. For a plumbing company, no label is the normal state. The useful local audit is the answer's actual source list, because Perplexity publishes no rule that one review site or directory controls local recommendations.
Microsoft Copilot
Microsoft says Copilot can turn a prompt into a shorter Bing query, ground the response in Bing results, and show the exact query and sources. Bing says its index is built from crawled third-party pages and that business queries may return details such as store hours and location.
For local visibility, make sure Bing can reach the correct branch page and that the public business facts do not conflict. Do not assume that a strong Google result automatically creates the same Copilot result.
What the citation mix actually looks like
Documentation says what an engine can do. Counting citations says what it did.
Yext analyzed more than 6.8 million citations across 1.6 million responses in October 2025 and found the engines disagree about what a good source is. 52.15% of Gemini citations came from brand-owned websites. 48.73% of ChatGPT citations came from third-party sites such as Yelp, TripAdvisor, and MapQuest. On subjective queries, directory citations inside ChatGPT rose to 46.3%, and niche industry sources made up 24% of Perplexity's citations.
Yext citation study
Different source diets
52.15%
Gemini citations from brand-owned websites
Base: Highest website share in the study
48.73%
ChatGPT citations from third-party sites
Yelp, TripAdvisor, MapQuest and similar
46.3%
Directory share of ChatGPT citations
Subjective, buying-decision queries
24%
Perplexity citations from niche industry sources
Subjective queries, vertical-specific content
Yext, October 2025: more than 6.8 million citations across 1.6 million responses from Gemini, ChatGPT, and Perplexity. All categories, not home services alone.
| Finding | Value | Base |
|---|---|---|
| Gemini citations from brand-owned websites | 52.15% | Highest website share in the study |
| ChatGPT citations from third-party sites | 48.73% | Yelp, TripAdvisor, MapQuest and similar |
| Directory share of ChatGPT citations | 46.3% | Subjective, buying-decision queries |
| Perplexity citations from niche industry sources | 24% | Subjective queries, vertical-specific content |
The Cheers market baseline runs a narrower version of that test on home services only: 32 natural-language trade prompts, 100 metro service areas in the United States and Canada, four engines, 28 days ending September 2, 2026, 356,019 citations. Every cited domain is sorted into contractor sites, directories and review platforms, social and forums, media and vendors, or Google itself. Aggregate only, no customer named. The methodology and the wider cuts are in AI visibility statistics.
Gemini sent 80.7% of its citations to a contractor website and Perplexity 72.3%, so on those two surfaces a thin or unreachable location page costs the citation outright. ChatGPT split almost evenly, 48.8% contractor sites against 44.7% directories and review platforms, which means a stale Angi or BBB record is doing real work there. Google AI Overviews and AI Mode is the only engine where google.com itself is a meaningful share, at 11.4%.
The most cited platform domains across all 100 metros were google.com, expertise.com, angi.com, bbb.org, and reddit.com. That is a list worth checking your own records against before deciding a directory does not matter.
Next step
Is AI recommending your business?
Find out how visible you are across ChatGPT, Gemini, Perplexity, and AI Overviews.
What I audit across all four
I save the prompt, engine, location context, businesses named, pages cited, and time checked. Then I group the sources into the business website, platform-managed business records, review or directory pages, publishers, and other third-party evidence.
The provider-tracking guide covers the tools that preserve that record. The citation cleanup guide covers conflicting branch facts once the source audit finds them.
I do not tell a client to build a listing because a generic checklist says an engine likes it. I want to see that source appear for the actual service, market, and competitor set, or know that customers already depend on it. The Cheers AI Visibility Grader runs the first version of that check on one profile across ChatGPT, Gemini, and Perplexity.
Crawler permissions and optional source maps are separate from citation measurement. What Is LLMs.txt? explains what that file can and cannot do.
Sources
Every link below was opened and checked on September 2, 2026.
- Searching the web with ChatGPT. OpenAI Help Center, updated August 2026. Search behavior, query rewriting, search partners, approximate location, reservation providers, ranking, and OAI-SearchBot eligibility.
- View related sources in Gemini Apps. Google Gemini Apps Help. Which sources related links can include and when a response has none.
- Optimizing your website for generative AI features on Google Search. Google Search Central, last updated July 2026. AI Mode retrieval, query fan-out, and business-detail guidance.
- Understanding source labels. Perplexity Help Center, updated August 7, 2026. What the Government, Academic, and Trusted labels do and do not mean.
- Practical tips for using Perplexity. Perplexity Help Center. Live-web search and source links.
- How web search works in Microsoft 365 Copilot. Microsoft Support. Bing query generation, grounding, and the Sources control.
- How Bing delivers search results. Microsoft Support. Bingbot crawling, index construction, and business results such as store hours and location.
- AI Visibility in 2025: How Gemini, ChatGPT, and Perplexity Cite Brands. Yext, October 2025. 6.8 million citations across 1.6 million responses.
- Cheers market baseline, 28 days ending September 2, 2026. 32 home-services prompts, 100 metro areas, four engines, 356,019 citations, aggregate only. Methodology.
Dylan Allen-Arnegard is the CEO and Co-Founder of Cheers. I help multi-location service brands get recommended by AI search and Google, with results tracked by location, employee, and competitor.