Why Gemini Cites You but ChatGPT Does Not: A 960-Answer Study

A page can appear in one AI service and be absent from another. Keep the results separate by engine before deciding what to change.
A page can appear in one AI service and be absent from another. Keep the results separate by engine before deciding what to change.
A colleague finds the company in Gemini, while a ChatGPT answer recommends competitors. Neither observation automatically invalidates the other. The services may select different sources.
A Prefer study published on September 19, 2026 helps illustrate this issue. Its questions concern AI search, so the findings should not be treated as a benchmark for every industry.[1]
This article explains how to read that study, then provides a process for comparing answers and sources for a business. The product documentation below was checked on November 11, 2026.
What differed across 960 AI answers?
The findings show both differences and shared sources. Results for all cited sites and results for the most frequently cited sites describe different groups, so read them separately.
The sample contains 80 questions sent to four engines three times
On September 13, Prefer sent 80 questions to four engines three times each. The services were ChatGPT, Gemini, Perplexity and Claude.[1]
The total of 960 counts collected answers. It is not a count of users, market size or distinct customer needs. Repeated answers to the same question are part of that total.
The questions cover AI search terminology, optimization and tools. Some name particular vendors; they are not a balanced sample of every kind of consumer purchase or business procurement question. Prefer’s published dataset
Before using the study as context for a business, read the questions it asked. A different subject can make a different set of websites relevant. The useful comparison starts with the question and the conditions under which the answer was obtained.
Sites unique to one engine and sites shared across engines
Of 1,329 cited domains, 966 appeared in only one engine. A domain is the main part of a site's address—for example, agent-signal.ai.[1]
However, 23 of the 25 most-cited domains appeared in at least three engines. Some sources were shared widely; others appeared in just one service. The engines did not operate in completely separate source worlds.[1]
The first result concerns all sites found across the study. The second focuses on the most frequently cited sites. One result does not invalidate the other, because the groups being examined differ.
If a business appears in only one service, these totals do not establish whether its performance is normal or poor. Record which page that service cited and which sources the other services used. That makes the observation useful for a later content review.

What this study does and does not establish
Keep the collection method and counting rules alongside the figures. Results obtained through a different interface or under different conditions need not match a manual check in an app.
API answers differ from checks in the ChatGPT or Gemini interface
The researchers used APIs: interfaces through which software sends questions to AI. They did not collect 960 answers by operating the consumer ChatGPT interface. The repeated runs were close together, rather than daily observations.[1]
Prefer's method specifies the DataForSEO LLM Responses API, with web search requested on each call. Each run was a fresh call without chat history; whether search occurred was based on what the API reported. Prefer’s methodology
The model names recorded for the study were gpt-5.6-luna for ChatGPT, gemini-3.8-flash for Gemini, sonar-pro for Perplexity and claude-sonnet-5 for Claude. These identify historical measurement conditions. They are not a list of models currently available in the consumer interfaces.
For an internal comparison, record whether an answer came from the web interface, an app or an API. Preserve each observation with its own conditions rather than merging different collection methods without explanation.
URLs, sites and brand mentions require different counts
A citation points to a source. A brand mention is a name appearing in the answer. Record these separately: a company can be mentioned without a link to its website.
Prefer defines a source as a unique cited URL within an answer. Its site counts generally use the host name with www removed. Subdomains are merged only for selected large platforms whose subdomains represent language, mobile or regional versions. Prefer’s published dataset
A company's developer site and support site can therefore count separately. Treating every reported site as a consolidated company domain would change the meaning of the totals.
The ranked lists count answers that cited a site at least once, rather than the total number of links. For a company's own records, keep brand mentions, cited URLs and site names in separate fields. The original evidence can then be regrouped without losing the counting rule.
A single-day study does not set an industry benchmark or prove an improvement
The study does not establish that a particular business underperforms in ChatGPT, or that three daily checks are sufficient. Its practical lesson is to preserve differences instead of hiding them inside one combined score.
Variation between closely spaced runs and changes across days after a page edit are different questions. Prefer's figures are not a before-and-after experiment on that business's content.[1]
The public JSON contains questions, methods and aggregate results. It does not provide all 960 answer texts with a complete answer-by-answer citation record. AgentSignal has not independently rescored every answer or verified the accuracy of every citation.
An internal presentation should therefore identify Prefer as the research publisher and retain the AI-search topic, API collection method and single-day scope. Decisions about a company's own pages need a separate record of questions relevant to its customers.
Check search and citations in ChatGPT, Gemini, Perplexity and Claude
Separate whether search ran, whether a source appeared in the answer, and whether that source supports the claim. Current documentation explains what can be checked now; it does not establish why an engine selected a particular site during Prefer's earlier study.
ChatGPT: inspect search activity and citation links
In ChatGPT on the web, search results and citations appear in the conversation when web search is used. Workspace settings can limit availability, so inspect the actual answer before recording that a search occurred. OpenAI’s web search guide
In the API, a web_search_call records a search-related action, while citation annotations identify sources attached to answer text. The sources field provides the wider list of URLs consulted, which can differ from the inline citations. OpenAI API documentation
This distinction matters when reading a report. Finding a company URL in consulted sources is different from finding it cited in the answer. For a manual interface check, preserve the search activity and links actually shown rather than claiming to have inspected API fields that were not visible.
Gemini: compare the Sources panel with the answer
When Gemini provides related links, open Sources within or beneath the answer. Google's help page describes the related links appearing in a side panel. Gemini’s source guidance
Not every answer has a Sources button. Its absence means that Gemini did not provide links for that response; it does not establish every source or process the system did or did not use.
Open each relevant link and identify the statement it supports. A panel may include related material, so do not treat every listed page as evidence for every claim in the answer.
Google Cloud also documents API metadata that connects generated content to source information. That is a developer-facing mechanism, not proof that the Gemini consumer app uses an identical internal implementation. Google Cloud documentation
Perplexity: distinguish search results from citations in an answer
For Perplexity, distinguish pages returned by search from the sources cited in the generated response. Current Agent API documentation describes result IDs and URLs, with citation markers referring to those result IDs. Perplexity Web Search
The documentation notes that inline markers depend on the prompt. Inspect the relationship between the ID, URL and supported statement rather than counting markers alone. More sources do not, by themselves, establish greater accuracy.
Prefer's study recorded sonar-pro. Perplexity now also documents a migration from Sonar to the Agent API, so the present response format should not be imposed on the earlier study. Sonar migration overview
Preserve the collection interface alongside each answer. If only a source list can be inspected, record that limit rather than reporting that every source was used in the final text.
Claude: check whether search ran and what the citations support
Anthropic's API documentation explains that, when web search is enabled, Claude determines when to search from the prompt. It receives search results, may repeat the process and then produces a response with cited sources. Anthropic’s web search documentation
Requesting search and observing that search ran are separate facts. Where an execution record is available, preserve it; where citations appear, compare the linked pages with the associated claims. Do not fill gaps with assumptions about hidden processing.
API documentation also does not establish every setting in the consumer Claude interface. A report should state the interface used and the search or citation information that was actually visible.
The purpose of comparing the four services is to understand the explanations and links customers may receive. A source count alone cannot answer that question.
1. Choose a real buying question
The following is our proposed workflow, not a treatment proven by the study. Begin with a question related to work the business wants to win.
Connect customer requirements to the question
Collect questions from sales calls and enquiries. Start with a question that searches for providers without naming the business. This helps examine discovery by someone who does not already know it.
For a corporate training provider, an illustrative question is:
How should I choose beginner-friendly AI training for 20 employees, including support after the course?
Replace the conditions with those of actual customers. Compared with a broad request for recommended training, this question specifies the audience, group size and follow-up support needed.
If cost is central to a real enquiry, record that requirement as well. Choose wording that reflects customer needs rather than adding favorable terms merely to make the company more likely to appear.
Keep changed questions separate from earlier citation records
Save the wording instead of recreating it for each check. If the question changes from lowest price to strongest ongoing support, treat it as a new question. Different recommendations would not be surprising.
Preserve the full text and the date it came into use. Record changes to language, target region and purpose when those conditions affect the comparison.
Overwriting the old question makes it harder to distinguish a change in the page from a change in the request. Keep a shared location from which people can copy the agreed wording, while retaining earlier versions.
Questions can still be improved as customer needs become clearer. The important point is to label the change rather than presenting the old and new observations as identical conditions.
2. Preserve the answer and its links
Submit the saved question in a new conversation and preserve the full answer with its sources. Check the linked page and the reason for its inclusion, as well as whether the company name appears.
Open the same question and its sources in ChatGPT and Gemini
For an initial manual comparison, open ChatGPT and Gemini. Start a new conversation in each and submit the saved question. Keep follow-up conversations separate from these checks.
Save the time, exact question and full answer together. Include the model or feature where visible; do not invent settings that the interface does not expose.
Next, open citation links in ChatGPT or Sources in Gemini. Copy the destination title and URL, then look for the explanation relevant to the question. Where no source is shown, record that observation.
The guide to checking ChatGPT search citations (in Japanese) provides a related walkthrough. Save the first answer before asking follow-up questions, since follow-ups change the conversation context.
Use a blank record for names, URLs and collection failures
Keep one record per answer. Use unknown where a field cannot be verified. The blank cells below are a template, not observed results.
| Record | What it helps establish | Notes |
|---|---|---|
| Time, full question and version | Whether conditions match | |
| Service, interface and visible model or feature | How the answer was obtained | |
| Full-answer location | Why a provider or product was selected | |
| Company name and known aliases | Whether the brand was mentioned | |
| Company citation URL and destination | Which pages a reader could visit | |
| Search status | Whether search was observed or remains unknown | |
| Collection status | Whether an answer was obtained or collection failed |
This is an initial comparison. Tracking change requires repeated collection under consistent conditions and a history of results. Ongoing measurement tools reduce that manual work.
AgentSignal's AI citation view lets teams examine results by AI service for registered questions. Read the captured answers and sources as well as the percentage. They provide clues about information that may need improvement.
3. Read the answers where the company’s page is missing
The answer can help identify what to inspect next. Reading its cited pages does not reveal the engine's complete internal reason for omitting another page.
Check whether the cited page supports the answer and its conditions
Open the cited pages and examine how they answer the question. The purpose is to identify missing information, not reproduce a competitor's article.
Compare the statement in the answer, the attached citation and the explanation on the destination page. Keep the following notes alongside the answer being reviewed.
| Check | Where to look |
|---|---|
| Reason for a recommendation | The sentences around the provider or product name |
| Evidence for that reason | The source page's explanation, pricing and eligibility |
| Conditions and exceptions | Audience, support period and extra charges |
| Unverified information | Unavailable pages or explanations that cannot be found |
If support for a claim cannot be found, leave it unverified instead of assuming the answer is correct. The comparison of source checking in Perplexity and ChatGPT (in Japanese) offers further context for choosing a research tool.
Separate missing explanations from services the company does not offer
In the training example, useful conditions might include who provides post-course support, how long it lasts and whether there is an additional fee. If the company already offers that support but has not explained it, add verified details.
If the company does not provide the support that mattered to the recommendation, wording alone will not make it eligible. Consider the offer or a different customer need. Absence is not automatically a writing problem.
For Google's AI search features, readable important content and useful internal links remain part of the basics. Eligibility does not guarantee inclusion.[2]
That guidance concerns AI Overviews and AI Mode in Google Search. Do not present it as a universal eligibility rule for every AI service, including the Gemini app. The immediate editorial goal is a page on which a prospective customer can understand the service and its conditions.
4. Compare each engine and question after the change
Record the URL, changes and publication date. Continue the same questions and examine results by engine. Keep the underlying answers and citations accessible beside the percentages.
Separate successful answers from failed collection attempts
A URL citation rate can count successful answers containing a link to the company's page as a share of all successfully collected answers. Use the same engine, question and comparison period for both parts of that calculation.
URL citation rate = successful answers citing a company page ÷ successfully collected answers × 100
As an illustrative calculation, links in three of ten successfully collected answers produce a 30% URL citation rate for those checks. This is not 30% of the market seeing the business.
Keep failed executions separate from valid answers without a citation. A missing link in a collected answer and an answer that could not be obtained provide different evidence.
The explanation of AI citation rates and denominators covers the definition in more detail. Align the tool's counting rule with the internal report before comparing results.
Record page edits and measurement changes together
Keep page edits and changes to questions or AI features in the same history. Mark periods in which the measurement conditions changed so they remain visible during comparison.
Mentions may rise while links do not. Links may rise without additional enquiries. Separate brand mentions, citations, visits and enquiries to identify the next improvement.
If links increase, inspect the landing page for clear service conditions and a way to enquire. If citations remain unchanged, reopen answers to the same question and check whether the added information actually addresses the need.
Begin with one question related to work the business wants to win. Record which engine recommended which page and why. That makes the comparison useful for improving information, rather than simply watching a score.
The initial check is complete when the answer and links are saved and the unknown fields are explicit. Use that record as the starting point for the next check under the same conditions.
FAQ
- Q. Is checking one engine enough?
- Choose services relevant to your customers and preserve engine-level results rather than relying only on an aggregate.
- Q. Does absence imply poor writing?
- No. Relevance, service conditions and whether the page can be retrieved also need examination.
- Q. Was this a daily tracking study?
- No. Repeated runs took place close together and do not establish day-to-day variation.
Sources
- [1] How AI engines search and cite: 960 answers measured (Prefer) — accessed 2026-11-11
- [2] AI features and your website (Google Search Central) — accessed 2026-11-11
- [3] AI answer study: September 2026 dataset (Prefer) — accessed 2026-11-11
- [4] Web search (OpenAI) — accessed 2026-11-11
- [5] Web search API guide (OpenAI) — accessed 2026-11-11
- [6] View related sources in Gemini Apps (Google) — accessed 2026-11-11
- [7] Grounding with Google Search (Google Cloud) — accessed 2026-11-11
- [8] Web Search (Perplexity) — accessed 2026-11-11
- [9] Migrate from Sonar to the Agent API (Perplexity) — accessed 2026-11-11
- [10] Web search tool (Anthropic) — accessed 2026-11-11
About the author
Shogo MizushimaCEO of kairos Inc. / AgentSignal Developer
Develops AgentSignal, a tool for measuring AI crawler visits and AI-referred traffic, and diagnosing AIO readiness. Writes about measurement and practical improvements for AI search using observed data.
Related articles

Case studies and research
37.5% More Search Visits: A Manufacturer’s SEO and AI Case
A specialist manufacturer reported 37.5% more organic search sessions after combined SEO and AI-search work. The useful lesson is how its content helped buyers choose, not a promise that AI optimization alone delivers that growth.
Published

Case studies and research
Claude and ElevenLabs: connect MCP for images, voice and video
Set up ElevenLabs MCP in Claude, prepare images, footage, narration and music, then finish in Studio. Includes connection commands and output checks.
Published

Case studies and research
Claude and HeyGen: set up MCP to create avatar explainers
Connect Claude and HeyGen through MCP, authenticate, prepare a script and create an avatar explainer. Includes voice, caption and MP4 checks.
Published