When an AI Answer Shows Sources: What Query Studies Can—and Cannot—Tell You

Query studies can show which searches displayed AI answers in a particular sample. They do not, on their own, reveal every search a model performed, the information it used internally, or why it recommended a company. Keep those questions separate when planning generative engine optimization (GEO). Use observed answers to measure exposure, source inspection to assess support, and clearly labeled research models to explore possible buyer questions. A useful monitoring panel represents real buyer decisions; it should not be redesigned merely to increase the chance of seeing a citation.

Four events that require different evidence

A marketing report can say “AI visibility improved” while describing several different events. An AI Overview may appear more often, a company's page may be cited, the company may be named, or a recommendation may be justified. Each needs its own record.

These events can overlap without becoming equivalent. A comparison article can be cited while its publisher is not recommended. An answer can name a company without linking its website. An unreadable collection is neither a demonstrated mention nor a demonstrated brand miss.

This distinction matters before studying query patterns. Otherwise, a team may use a research finding about the first event to promise the fourth.

What Pew actually observed about question shape

Pew's July 22, 2025 analysis examined 68,879 unique Google queries from 900 U.S. adults' March browsing. Results were collected April 7–17, 2025. AI summaries appeared for 12,593 queries, approximately 18%.

Pew limits the findings to “how search results appeared at the time they were collected.”

Do not combine the rows into a new result

The rows do not give the joint rate for queries that are *both* question-shaped and ten words long. The ratios `60 / 8 = 7.5` and `53 / 8 ≈ 6.6` do not measure the effect of rewriting a query; the study did not randomly assign those changes. Keep monitoring questions relevant to the buyer's decision instead of padding their wording.

A percentage can answer the reverse of your question

Ahrefs' April 17, 2025 CTR article describes a selected sample of 300,000 keywords: 150,000 with an AI Overview and 150,000 informational keywords without one. A methodological sidebar reports that 99.2% of AIO-triggering keywords were informational.

The sidebar concerns the composition of queries *with* an Overview. It does not say that 99.2% of all informational queries display one. The deliberately balanced comparison sample also does not establish a 50% market-wide display rate.

In probability terms, “informational given an Overview” and “Overview given informational” have different denominators. To estimate the second, you need the relevant population of informational queries, including those without an Overview, and an appropriate sampling design.

For a brand team, write the denominator in plain English before using any percentage. “Share of captured answers that name the brand” differs from “share of all scheduled checks that produced a readable answer.” A high rate in a selected set cannot automatically describe the questions your market actually asks.

Google's fan-out explanation is broader than a hidden-query log

Google Search Central says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources. It also says the two features may use different models and techniques, so their answers and links can vary.

That official description supports planning for related buyer questions. It does not disclose the complete internal search sequence for an individual answer. A source panel is visible output; it is not automatically a full execution trace.

For example, a manufacturer page may help explain inspection methods, materials or an application. Its relevance should be assessed against those buyer needs. Calling each paragraph a “verified fan-out target” would require evidence the team usually does not possess.

Use fan-out as a research concept, with provenance. Separate a provider's general explanation, a researcher's generated related questions, and an actual recorded tool trace. They are useful for different tasks.

What Moz's 50,000 prompts measure

Moz's August 6, 2026 study reports 50,000 generated fan-out prompts across 1,000 subtopics and 20 verticals, plus 5,333 Gemini API grounding queries. It used generated brand and competitor hints. In its generated prompt set, 12.8% contained a target-brand or competitor mention; entity and comparison types accounted for 97% of those brand mentions.

The 97% is about brand mentions in that modeled prompt dataset. It is not the share of real consumer recommendations caused by those two question types. Nor is the dataset an export of Google's private AI Mode fan-out history.

Its 8.2-word average and 0.67 mean alignment describe the generation design and relevance measure. They are not target lengths or scores that a factory page must achieve.

The practical use is exploratory: generate candidate questions, check whether they represent distinct decisions, and review the resulting coverage. Keep generated questions labeled as such. A model supplied with brand hints cannot establish unsolicited brand discovery in an unbranded consumer test.

Missing citations do not prove “training memory only”

In an ordinary consumer answer, the absence of a visible citation is an observation about what was displayed. Unless the available evidence exposes the relevant tool execution, it does not establish that no retrieval occurred, that all information came from training, or that recent information could not affect the answer.

Use a conservative label: “no visible source link in this captured answer.” If the answer is incomplete, record that separately. Do not fill an unavailable execution history with a confident architecture story.

Google's Gemini API grounding documentation illustrates a more explicit observation surface. Its current API response description includes search-call and search-result steps, with executed queries and citation annotations. Those are documented API artifacts, not a claim that an analyst can inspect the same fields in the Gemini consumer app or Google AI Overview.

API experiments can investigate a configured application. Consumer-web monitoring answers a different question about the experience buyers can access. Keep the two evidence sets separate, including model, tool configuration, account and location context. The retrieval-setting guide explains why those conditions belong in the record.

Topic research is useful without pretending demand is exact

SISTRIX's Prompt Research documentation describes grouping questions into topics. It also explains that its AI topic-volume model distributes estimated aggregate demand using Google Search demand as a proxy, with cross-platform assumptions. It says the volume is not suitable for isolating demand exclusively on ChatGPT or Gemini.

Use such a model as a planning input under its stated method. Do not relabel its topic volume as an exact count of actual consumer questions submitted to a particular AI platform.

Keyword and question research can complement each other. Search terms help identify subjects and vocabulary; interviews, RFQs and support conversations reveal the decisions people need to make. Generated variants can expand coverage, but a large number of similar phrasings does not create a large number of independent buyer intents.

Build a panel around decisions that can be acted on

For an exporter, discovery, technical qualification, comparison and purchasing conditions are different jobs. Preserve that distinction when selecting questions.

1. Name the decision. A procurement engineer checking a machining envelope needs different evidence from a marketer choosing monitoring software. Write down what a useful answer should help the reader decide.
2. Separate discovery from identity checks. A question that already names the brand can test its description. It cannot establish that the system independently discovered the brand.
3. Keep exact question versions. If wording changes, record a new version rather than replacing historical negatives. Similar variants remain one intent where they ask the same decision.
4. Record the actual surface. Google AI Overview, Google AI Mode, Gemini App and an API are separate observations. English wording and a U.S. scenario do not prove a U.S. network exit.
5. Preserve outcomes and missingness. Retain complete answers, visible URLs, negative cases, AIO-not-triggered observations and failures. Review the entity and recommendation reason before accepting a positive result.
6. Recheck after a real intervention. Record what changed and when it became public. A later difference is an observation; attributing it causally requires a stronger design.

The panel should remain useful when the brand is absent. That absence can reveal a content gap, an unclear offer or a need for stronger evidence. It should not automatically trigger a rewrite of the question until the brand appears.

A fictional example with two legitimate denominators

Suppose a team schedules 30 Google checks for a defined panel. Six cannot be read. Of the 24 readable search-result observations, 18 show an AI Overview. Two of those 18 Overviews contain a company recommendation whose entity and individual reason pass the team's review.

The display rate is `18 / 24 = 75%` among readable search-result observations. The qualified-recommendation rate is `2 / 18 ≈ 11.1%` among captured AI Overviews. Report the six missing checks separately; neither percentage turns them into valid brand-negative answers.

The two rates answer different questions. A change from 18 to 20 displayed Overviews would not, by itself, show that more Overviews recommended the company. Likewise, two recommendations under one intent would not establish success across two independent buyer decisions.

These numbers are fictional arithmetic, not Xindar's monitoring results or a forecast of Google behavior.

Connect content work to evidence the buyer can inspect

Answer the practical question completely: define the offered scope, show the relevant facts, name the source and keep material conditions beside the claim. If the page cannot be accessed or its technical statement is unsupported, a theory about hidden fan-outs does not solve the buyer's problem.

Google's official guidance requires a supporting page to be indexed and eligible for a Search snippet; it also says eligibility does not guarantee crawling, indexing or serving. Technical eligibility, a visible citation and a qualified service recommendation remain separate gates.

Semvi by Xindar is Xindar's monitoring offering. Its product team has confirmed ChatGPT and Gemini consumer-web sampling and complete-answer/citation-URL export; actual account behavior and plan scope require their own verification. Those confirmations do not establish Google Search, AI Overview or AI Mode coverage. Xindar's English website and content work is a separate implementation scope. The monitoring brief provides questions to settle before purchase.

Questions from marketing teams

Should we make every monitoring question at least ten words long?

Use the words the decision needs. These historical categories establish neither a universal minimum nor a measured benefit from padding a query.

Does the 97% fan-out result mean entity and comparison pages are enough?

It describes brand mentions in Moz's generated dataset, with its design and hints. It does not measure all consumer demand or recommendations. Technical qualification, service conditions and other buyer decisions still need appropriate information.

Can a source link prove that the model read the whole page?

It establishes the visible link. Review whether the cited passage supports the associated claim; do not infer complete internal reading from the presence of a URL.

Should API results be combined with consumer answers to improve the sample size?

Report them separately when they use different products, models, configurations or access routes. A larger mixed number can obscure the experience you intended to measure.

Source and revision note

Reviewed November 4, 2026. This revision preserves the topic of query selection while correcting distinctions between display observations, conditional percentages, generated fan-out research and consumer recommendation evidence. Historical research figures retain their study scope. No universal retrieval trigger, training-versus-retrieval timetable or guaranteed citation outcome is claimed.