What AI Engines Cite: Insights from a 10,000-Prompt Study

What changed in AI search citations?

AI search citations have become a visible part of the answer interface, not merely an underlying ranking input. When an engine produces a synthesized answer, the cited sources are the pages a reader can inspect, trust, or ignore. That changes the question for publishers from whether a page ranks to whether it is selected and cited when an answer is generated.

The clearest before-and-after evidence comes from Pew Research Center's analysis of Google activity in March 2025. On visits where Google displayed an AI summary, users clicked a traditional result link 8% of the time. On visits without an AI summary, they clicked a traditional result link 15% of the time. Users clicked a source cited inside an AI summary on 1% of those visits. This is more than a different page layout because it makes the citation itself a scarce discovery surface.

InnovAit AI's study, published September 25, 2026, examined 10,000 prompts across ChatGPT Search, Perplexity Pro, Gemini, and Google AI Overviews. It reports that informational prompts produced higher citation rates than commercial, transactional, or navigational prompts across the engines it tested. That finding fits a practical reality: an answer engine needs material it can use to answer a question clearly, with enough context to support the response.

The study should be read as a benchmark from its publisher, not as a universal law of retrieval. Its reported results are still useful because they point toward a better operating model: observe the prompts, engines, source domains, and answer appearances that affect your category. One score cannot capture this environment. Citation Share, the percentage of relevant AI answers in a category that cite you, is the more useful headline metric because it measures presence in the answers people actually receive.

1. Treat an AI citation as a selection event inside a generated answer.
2. Separate visibility in ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews.
3. Track question intent, not just a broad keyword bucket.
4. Audit whether cited pages give a direct answer before adding surrounding detail.

What does the 10,000-prompt study say about engine differences?

The study says that engines do not cite content in the same pattern. Its reported informational-prompt citation rates range from 72% for ChatGPT Search to 83% for Gemini, with Perplexity at 78% and Google AI Overviews at 75%. For commercial prompts, the reported range is 60% to 70%. The exact percentages come from InnovAit AI's own dataset, so they should guide testing rather than replace it.

Its directional finding is more important than any single percentage. Informational coverage appears to give engines more material to cite than a page designed only to capture a late-stage commercial query. A category page that explains terms, boundaries, methods, costs, alternatives, and evidence can participate in more answer paths than a thin landing page that repeats a sales claim.

Engine variation also makes generic advice suspect. A domain that appears frequently in one product can be less visible in another. InnovAit reports different domain patterns for YouTube, Wikipedia, Stack Overflow, Reuters, and its own site across the engines it tested. That does not mean a brand should chase those domains or mimic their format. It means the research, documentation, editorial coverage, and direct-answer assets available in a category may differ by engine.

Omnicite calls the operating discipline behind this work Citation Engineering. It is not an attempt to game a model. It is the work of creating authoritative, current coverage that an engine can understand and cite, then measuring whether that work appears in the answers that matter.

1. Use the same prompt set across multiple engines before drawing conclusions.
2. Keep informational, commercial, transactional, and navigational questions separate.
3. Record the cited URL, not only whether a brand name appeared.
4. Compare your citation presence with named competitors on the same prompts.

Why does the before-and-after click data matter to publishers?

The before-and-after click data matters because an answer can satisfy a search without sending a visit. Pew found that 26% of visits to Google pages with an AI summary ended the browsing session, compared with 16% of pages containing only traditional results. That is a material change in how publishers should interpret a top-of-page answer environment.

It would be wrong to conclude that traffic no longer matters. A cited page can still earn visits, qualified readers, and downstream demand. The point is narrower: traditional organic sessions alone do not show whether a company is present in an AI answer. A page can receive less click traffic while still becoming a cited source, or it can retain traffic while disappearing from a high-value answer set.

Pew also found that 88% of the AI summaries in its sample cited three or more sources. That leaves room for more than one publisher, but not unlimited room. The relevant competition is the small group of sources cited for a particular question. That is why category-level monitoring beats a report that shows one average position for thousands of unrelated terms.

For B2B SaaS teams, the risk is being absent when a buyer asks for the best tool, an alternative, or a way to solve a problem. For local and multi-location businesses, the risk is being absent when a user asks for the best service in a city. In both cases, the outcome to measure is answer presence across the question universe, then Citation Share against the competitors that appear beside you.

1. Before: traditional result clicks on pages without an AI summary were 15% of visits in Pew's March 2025 analysis.
2. After: traditional result clicks on pages with an AI summary were 8% of visits.
3. After: clicks on sources cited within an AI summary occurred on 1% of those visits.
4. What to do: measure answer presence and Citation Share alongside traffic, conversions, calls, and bookings.

Who is most affected by this shift?

Companies with an established search program are affected because their existing content may be optimized for rankings without being easy to cite. A page can be technically sound and still fail to answer the question an engine is trying to resolve. The issue is especially sharp for firms that depend on category discovery rather than branded demand.

B2B SaaS and technology growth teams need to know whether answer engines recommend their product or a competitor for category and comparison prompts. Their best evidence comes from citation share on those prompts, share of voice against named competitors, and AI-sourced signups where attribution is available. The goal is not a promise of a fixed citation count. It is a monitored baseline and a repeatable system for improving coverage.

Local, multi-location, and service businesses face a related problem with a geographic dimension. A potential customer may ask an engine for the best service in a city, then never inspect a long list of links. These businesses need pages that clearly establish service coverage, local expertise, proof, and the practical answer to the question. They also need to check whether AI Overviews and other engines cite those pages for the places where demand exists.

Publishers face this shift too. Pew found that Wikipedia, YouTube, and Reddit together made up 15% of sources listed in the AI summaries it examined. That does not create a shortcut for an independent publisher. It demonstrates that the cited-source set can favor formats and sources that are already legible to users and search systems. A publisher's response is to make its own expertise explicit, sourced, current, and complete enough to be a reference.

1. B2B SaaS teams should monitor category, comparison, and alternative prompts.
2. Service businesses should monitor service-plus-location prompts.
3. Publishers should monitor the pages and content formats cited for their specialist topics.
4. Leadership teams should treat citation visibility as a reportable discovery metric, not a side project.

How should a team respond to engine-specific citation patterns?

Teams should respond by building a prompt-led measurement system, then using it to direct editorial work. Start with the questions customers actually ask, including definitions, comparisons, buying criteria, implementation questions, local service questions, and objections. Run that set across the engines that matter to the business. Record each answer, cited URL, competitor appearance, and the date.

Next, inspect the page rather than assuming that more words will solve the problem. A source-worthy page answers its title question near the top, distinguishes facts from claims, links to primary evidence where a factual claim needs support, and covers the follow-up questions a reader would reasonably have. Google Search Central says structured data provides explicit clues about the meaning of a page. That is useful context, but markup does not substitute for clear on-page information or evidence.

InnovAit reports that its citation rates increased as its tested retrieval-window size moved from 250 tokens to 500 and then 1,000 tokens. That is a reported study finding about the systems it assessed, not an instruction to write longer blocks of copy. The sound editorial response is to make each section independently intelligible. Give the question a direct answer, define the scope, provide the evidence, and use descriptive headings so a page has clear units of meaning.

Finally, refresh based on observed decay rather than a fixed content calendar. InnovAit's study reports that Google AI Overviews retained 78% of citations after 30 days and 48% after 90 days in its benchmark. Whether your category follows that exact path must be tested. The actionable point is that citation visibility can move after publication, so measurement and updates belong in the same system.

1. Create a controlled prompt set tied to buying and discovery questions.
2. Capture citations by engine and date in a consistent record.
3. Prioritize gaps where competitors are cited and your company is absent.
4. Publish direct answers supported by real sources, then recheck the same prompt set.
5. Refresh pages when measured answer presence or Citation Share declines.
6. Report citation metrics with business outcomes instead of treating citations as a vanity count.

What should marketers avoid doing?

Marketers should avoid treating a single study as proof that one tactic will produce citations everywhere. The InnovAit study has useful observations, but citation systems change, prompts vary, and each category has its own source landscape. The right response to a benchmark is to test it against a defined question set and publish the result of that test.

Do not turn source formatting into a substitute for substance. Schema can help systems understand page meaning, as Google documents, but it cannot make unsupported claims authoritative. The editorial standard is straightforward: use real dated sources for statistics, preserve the context of the evidence, and do not publish numbers that cannot be checked.

Avoid chasing a citation count without asking where it came from. A citation on a low-intent definition question may be useful, but it does not carry the same commercial weight as a citation on a category comparison or a local recommendation. Citation Count per day measures volume. Answer Presence measures breadth. Share of Voice compares competitors. Citation Share gives the percentage of relevant answers in a category that cite you. Each answers a different question.

The bigger mistake is waiting for a stable rulebook. AI answer products will continue to change. The defensible strategy is quality, coverage, freshness, and measurement at a scale an in-house team may struggle to sustain. Rankings get you found, while citations get you chosen.

1. Do not fabricate evidence, customer results, or citation statistics.
2. Do not promise a prospect a specific number of citations.
3. Do not rely on a single engine or one prompt when reporting visibility.
4. Do not confuse structured data with an editorial answer.
5. Do not call an uncited page authoritative without checking the sources and the answer itself.

Sources

Source: InnovAit AI

InnovAit AI reported results from a 10,000-prompt analysis across ChatGPT Search, Perplexity Pro, Gemini, and Google AI Overviews, including intent, domain, chunk-window, and retention findings. InnovAit AI, 2026-09-25

Source: Pew Research Center

Pew Research Center reported that traditional result links were clicked on 8% of visits with an AI summary and 15% without one in its March 2025 analysis, while cited links in summaries were clicked on 1% of visits with a summary. Pew Research Center, 2025-07-22

Source: Google Search Central

Google Search Central says structured data provides explicit clues about a page's meaning and can help Google understand page content. Google Search Central, 2026-09-30