The source ecosystem behind AI answers
A cross-brand study of search results, citations, and the evidence behind buyer questions
An AI answer can name a brand without citing its own pages. It can surface one page in search and cite another. It can use competitor pages, documentation, and other external sources to answer the same buyer question. We studied those differences across Viziquo’s audit records to understand what teams can observe—and what they should investigate before deciding what to change.
143
audited brands
72,223
answer-citation records
167,788
exported search-result records
Viziquo Research
Published September 29, 2026
These inventory totals span 184 audit runs. The findings below use the latest available run for each brand and provider and focus on organic questions with recorded sources. Citation findings draw on all eight captured provider surfaces; the search/citation comparison draws on the five with search exports. Provider coverage is uneven, so the results describe this audited corpus rather than an average AI provider.
EXECUTIVE SUMMARY
A citation gap is a question to investigate, not a diagnosis
Across eight captured provider surfaces, an audited brand’s own page was cited on 314 of 2,591 citation-bearing organic provider-question observations (12.1%). Provider coverage was uneven, so pooled results describe this dataset rather than an equal-provider benchmark. The median brand-provider panel had an owned citation on 7.1% of its eligible questions. That is a limited owned-source footprint in this corpus, but a citation record alone cannot establish whether the brand appeared in the answer or why a page was chosen.
Three distinctions help teams use the evidence:
1. Search presence and citation presence differ.
Across the five surfaces with search exports, an owned page appeared in search without an owned citation on 58 of 2,484 source-bearing organic observations. That recorded state is a prompt to compare pages and answer claims; it is not proof that a page was read and rejected.
2. Buyer intent changes the competing evidence.
Within 157 brand-provider panels containing both question types, configured-competitor-owned pages were cited 29.9 percentage points more often on vendor-selection questions than on diagnostic questions, on average. The panels are unevenly distributed across providers, so this is an observed contrast in the audited set.
3. A recurring host is not necessarily a single dominant page.
In 170 eligible brand-provider panels, the most recurrent external host reached a median 35.7% of questions with an external citation; the most recurrent exact URL reached 20.0%.
The takeaway: Start with the exact buyer question and answer. Then examine recorded search results where available, citations, ownership, and competing evidence together. These figures identify where to investigate; they do not show which publishing or SEO action caused an answer.
WHAT WE MEASURED
What we measured
The study uses source records from Viziquo AI visibility audits created from June 2 through September 27, 2026. The full inventory describes scale and coverage. The findings use the latest available run for each brand and provider, and organic questions with recorded sources.
| Study layer | Count | Meaning |
|---|---|---|
| Distinct audited brands | 143 | One brand appeared under two IDs and is counted once by brand URL. |
| Audit runs in the full exports | 184 | Repeated runs are present in the inventory. |
| Source-bearing provider-question observations in the full exports | 8,692 | A brand, run, provider, scenario, and conversation with at least one source record. |
| Citation URL records in the full exports | 72,223 | Across eight captured provider surfaces: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overview, Grok, Microsoft Bing AI Answers (Copilot Search), and Perplexity. |
| Search-result URL records in the full exports | 167,788 | Across five captured provider surfaces. |
| Latest-run organic observations with a citation or search result | 2,607 | The narrower population used for the findings. |
| Latest-run organic observations with at least one citation | 2,591 | The denominator for the citation findings across all eight surfaces. |
The files omit conversations with no exported search result or citation. The 2,591 observations are not 2,591 unique buyer questions: the same scenario can be tested for multiple brands or providers. Percentages below are question-presence measures, not shares of raw URL rows.
HOW TO READ THE EVIDENCE
How to read the evidence
For an important buyer question, examine four connected observations:
Buyer question and provider
The task and context, and which assistant produced the answer and its source records.
Exported search results
Which URLs were recorded in that provider’s search layer, when it exposes one.
Answer and citations
Which URLs were attached to that provider’s answer.
Brand outcome
Whether the brand was absent, mentioned, or recommended, and how competitors were described.
Those observations do not always move together.
Every source also needs two labels. Ownership is relative to the audited brand: brand-owned, configured-competitor-owned, or neither. Source role describes the publishing surface, such as official documentation, a review directory, a community, media, or research. A competitor’s documentation page can be both competitor-owned and documentation. An external source may still be an unconfigured competitor, so “external” does not automatically mean independent.
FINDINGS
Four findings and an exploratory note
The findings below describe the audited questions and captured source records. Each chart has its own population and denominator. Read the accompanying takeaway as a next question to investigate, not as a claim that a page type or channel caused an AI answer.
FINDING 01 · OWNED CITATIONS
Brand-owned pages appeared in a small share of citation sets
Across all eight captured citation surfaces, 314 of 2,591 citation-bearing organic provider-question observations included at least one brand-owned URL (12.1%). One provider supplies 1,919 of those 2,591 observations (74.1%) and 143 of 177 brand-provider panels. This pooled figure gives every observation one vote. Looking instead at each panel, the median panel rate was 7.1%; the middle half ran from 0% to 20%, and the mean panel rate was 13.2%.
Brand-owned citation participation
Distribution of owned-citation rates across 177 brand-provider panels. Each panel’s denominator is its own citation-bearing organic observations.
Median 7.1%
0% · 25th percentile
75th percentile 20%
100%
MEAN PANEL
13.2% of each panel’s eligible questions
| Measure | Result |
|---|---|
| Observations with a brand-owned citation | 314 of 2,591 · 12.1% |
| Median brand-provider panel rate | 7.1% |
| 25th–75th percentile of panel rates | 0%–20% |
| Mean brand-provider panel rate | 13.2% |
As a sensitivity check, in the other seven provider panels combined, 78 of 672 citation-bearing organic observations included an owned URL (11.6%). Those seven panels cover few brands, so this is not a separate market benchmark.
WHAT THIS MEANS
For many audited brand-provider combinations, a brand’s own pages rarely supplied citations on these organic questions. A brand may still be named or recommended using someone else’s sources, and an owned citation may support only a minor detail.
WHAT TO EXAMINE NEXT
For questions without an owned citation, was the brand named in the answer? What claims did the cited pages support, and did the brand have a relevant page or evidence for those claims?
Distribution of owned-citation rates across 177 brand-provider panels. Each panel’s denominator is its own citation-bearing organic observations. “Organic” means the tested question did not name the audited brand or a configured competitor.
FINDING 02 · SEARCH AND CITATIONS
An owned page could appear in search without being cited
The five provider surfaces with search exports produced 2,484 latest-run organic observations with a search result or citation. We classified each observation by whether any brand-owned URL appeared in its recorded search and citation sets.
Where brand-owned pages appeared in exported source records
Five provider surfaces with search exports. 2,484 latest-run organic observations with a search result or citation. Categories are mutually exclusive and sum to 2,484. This is not a search-to-citation conversion funnel.
No owned URL in either exported set
2,121 · 85.4%
2,121 of 2,484 observations
Owned URL in search, none in citations
58 · 2.3%
58 of 2,484 observations
At least one owned URL cited
305 · 12.3%
305 of 2,484 observations
| Owned-source state in the exports | Observations | Share of 2,484 observations |
|---|---|---|
| No owned URL in either exported set | 2,121 | 85.4% |
| Owned URL in search, none in citations | 58 | 2.3% |
| At least one owned URL cited | 305 | 12.3% |
The 58 search-only observations show an owned page in the recorded search set but no owned page attached as a citation. They do not show that a provider read, evaluated, or rejected that page. The 2,121 observations without an owned URL in either exported set do not show whether the answer named the brand. An owned citation likewise does not establish a meaningful recommendation.
WHAT THIS MEANS
Search records, citations, and answer wording answer different questions. A team needs all three to tell a discovery gap from a citation gap or a weak answer mention.
WHAT TO EXAMINE NEXT
Where an owned page appeared only in search, compare its substance with the pages cited for the answer’s actual claims. Where no owned URL appeared, read the answer before concluding that the brand was absent.
Three mutually exclusive export states across 2,484 source-bearing organic provider-question observations on five surfaces. This is not a search-to-citation conversion funnel. The five exports have different recording behavior; on two surfaces, search and citation URL lists largely or entirely mirror one another.
FINDING 03 · BUYER INTENT
Vendor-choice questions contained more configured-competitor citations
On vendor-selection questions, configured-competitor-owned URLs appeared in a mean 74.7% of citation sets when each brand-provider panel was weighted equally. On diagnostic questions, the corresponding mean was 43.1%. More directly, within the 157 panels that had both question types, the vendor-selection rate was 29.9 percentage points higher on average.
Configured competitors in cited evidence
Mean panel-specific question presence of a configured-competitor citation, with each brand-provider panel weighted equally. The 29.9-point contrast is the mean of within-panel differences across 157 panels that contain both question types, not the subtraction of these two means.
Vendor selection
74.7%
421 citation-bearing observations · 176 panels
Diagnostic
43.1%
534 citation-bearing observations · 158 panels
Vendor selection
Citation-bearing observations: 421
Brand-provider panels: 176
Mean panel-specific question presence: 74.7%
Diagnostic
Citation-bearing observations: 534
Brand-provider panels: 158
Mean panel-specific question presence: 43.1%
Configured-competitor question presence by buyer intent, averaged across brand-provider panels. Source-role categories are not mutually exclusive.
| Buyer task | Citation-bearing observations | Brand-provider panels | Mean panel-specific question presence |
|---|---|---|---|
| Vendor selection | 421 | 176 | 74.7% |
| Diagnostic | 534 | 158 | 43.1% |
The within-panel contrast was positive in the aggregate on each of the eight captured surfaces, but the provider panels are small outside the largest one. Of the 157 paired panels, 125 come from the largest provider panel. Across the remaining 32 paired panels, the mean difference was 11.2 points. This makes the direction worth investigating without treating 29.9 points as a universal provider effect.
Reviewed external official documentation also appeared more often on tactical/how-to citation sets than on vendor-selection sets: mean panel-specific presence was 38.6% versus 12.1%. Treat that as a secondary, incomplete-classification observation. A large share of external hosts is still unverified, particularly in the smaller provider panels, so the figures do not rank source-type preferences.
WHAT THIS MEANS
The question’s job shapes the evidence a team should inspect. Vendor-choice gaps call for examining competitor alternatives, decision criteria, and how the answer compares them. Tactical gaps call for examining the instructions and specifications cited in that answer.
WHAT TO EXAMINE NEXT
On missed vendor-choice questions, which configured competitors appeared, what did the answer say about them, and which cited pages supported the comparison?
Configured-competitor question presence by buyer intent, averaged across brand-provider panels. The paired 29.9-point contrast is the mean of within-panel differences, not the subtraction of the two overall means. Source-role categories are not mutually exclusive.
FINDING 04 · SOURCE CONCENTRATION
A recurring host and a recurring page are different signals
We measured reach once per question rather than counting every citation row. Among the 170 brand-provider panels with at least five organic questions containing an external citation, the most recurrent exact external URL reached a median 20.0% of eligible questions. The most recurrent external host reached 35.7%. At least one of each panel’s five most recurrent external hosts appeared on a median 72.9% of its eligible questions.
How concentrated were external citations within a panel’s questions?
Median reach across 170 brand-provider panels with at least five organic questions containing an external citation. The top-five value is a union, not the sum of five host shares.
Most recurrent exact external URL
20.0%
Median question reach
Most recurrent external host
35.7%
Median question reach
At least one of the five most recurrent external hosts
72.9%
Median union of question reach, not an additive total
| Concentration measure | Median share of eligible questions across 170 panels |
|---|---|
| Most recurrent exact external URL | 20.0% |
| Most recurrent external host | 35.7% |
| At least one of the five most recurrent external hosts | 72.9% |
The middle half of panels had a top-host reach of 27.3%–44.4%. A host can recur through several pages, and a high row count can come from repeated appearances within a narrow set of questions. Neither pattern proves source authority or a partnership opportunity by itself.
WHAT THIS MEANS
Investigate whether a publisher answers many relevant buyer questions or whether one URL is repeatedly attached to a small cluster of answers. Those are different editorial and distribution signals.
WHAT TO EXAMINE NEXT
Which questions did the recurring host appear on, what did its cited pages contribute, and was the publisher relevant to the audited brand or category?
Median question reach across 170 eligible brand-provider panels. “External” means neither brand-owned nor configured competitor-owned; it can include an unconfigured competitor. Hosts are exact captured hosts, with subdomains unmerged. The top-five value is a union, not the sum of five percentages.
EXPLORATORY NOTE
Matched questions often had different citation sets across surfaces
In the latest-run data, 182 organic brand-scenario units had citations from at least two captured provider surfaces. Comparing every available pair of surfaces within each unit produced 1,606 pair comparisons across all eight surfaces. 1,069 comparisons (66.6%) shared no exact cited URL; 778 (48.4%) shared no exact cited host.
Matched scenarios often shared no exact citation
Exploratory pairwise comparisons from 182 organic brand-scenario units across nine brands. 1,606 pair comparisons are not independent tests.
No exact cited URL shared
1,069 · 66.6%
1,069 of 1,606 pair comparisons
No exact cited host shared
778 · 48.4%
778 of 1,606 pair comparisons
| Exact-overlap test | Pair comparisons with no overlap | Share of 1,606 comparisons |
|---|---|---|
| No exact cited URL shared | 1,069 | 66.6% |
| No exact cited host shared | 778 | 48.4% |
These 1,606 comparisons are not independent tests: one scenario can contribute many provider pairs. Only nine scenarios from two brands had citations on all eight surfaces. Differences also depend on citation count, URL normalization, and export behavior. The evidence supports a narrow conclusion: for these matched questions, a citation map from one surface often did not transfer unchanged to another. It does not rank provider quality or estimate a future citation probability.
WHAT THIS MEANS
Teams should preserve provider-specific evidence for important buyer questions. A single citation list may not describe how the same question is sourced elsewhere.
WHAT TO EXAMINE NEXT
Compare answer claims and citation contexts on a stable set of buyer questions across the surfaces that matter to your audience.
Exploratory pairwise exact-overlap comparisons from 182 matched organic scenarios across nine brands. The 1,606 pairs are not independent tests.
WHAT TO INVESTIGATE NEXT
What to investigate next
Use this sequence for each important unbranded buyer question. Preserve the exact question, provider, date, and run conditions so a later test can be compared meaningfully.
| Observed state | First diagnostic question | Evidence to inspect |
|---|---|---|
| Brand absent from answer; no owned URL in either export | What evidence did the answer use, and was the question aligned with a page we actually publish? | Answer text, cited claims, question-page fit, and discoverability of relevant pages. |
| Brand mentioned; no owned citation | Which external pages supported the description, and was the framing accurate? | Answer wording, competitor and other external citations, and the brand’s own evidence for those claims. |
| Owned URL in search; no owned citation | How does the found page compare with the cited pages for this specific answer? | Search-export URL, cited URLs, page contents, freshness, and the claim each source supports. |
| Owned URL cited | Did the citation support a meaningful brand mention or recommendation, or only a peripheral detail? | Citation context and the answer’s wording. |
| Configured competitors recur | Which buyer tasks and decision criteria brought them in? | Repeated questions, competitor-owned pages, answer claims, and source roles. |
These are investigation paths, not explanations of why a model chose a source. Keep exploratory prompts separate from a stable recurring question set. After any content, distribution, or technical change, compare repeat runs under documented conditions before attributing an answer change to that work.
METHOD AND LIMITATIONS
Method and limitations
Dataset and primary analysis
Viziquo analyzed source exports from 184 audit runs covering 143 distinct brands between June 2 and September 27, 2026. The complete inventory has 72,223 citation URL records across eight captured provider surfaces (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overview, Grok, Microsoft Bing AI Answers (Copilot Search), and Perplexity) and 167,788 search-result URL records across five. We selected the latest source-bearing run for each brand and provider, kept organic questions, grouped URL rows into brand-provider-conversation-scenario observations, and measured whether a source was present at least once per observation. This yielded 2,607 observations with any exported source and 2,591 with a citation. The five-surface search analysis used 2,484 observations with a search result or citation.
For panel summaries, we calculated a rate within each brand-provider combination, then reported the median or mean across those combinations. This prevents a brand with more question records from automatically setting the panel average. It does not balance providers: one provider contributes 143 of 177 brand-provider panels and 1,919 of 2,591 citation-bearing organic observations (74.1%). Each of the other seven contributes only 2–9 brands. The combined values describe the observed audit corpus, not a representative estimate of all brands, prompts, or providers. Only nine latest-run organic scenarios from two brands have citation coverage on all eight surfaces, which is too small for a balanced eight-provider benchmark.
Definitions and limits
Brand-owned and configured-competitor-owned labels are relative to the audited brand. “External” means neither flag is set, not necessarily independent. We reviewed 100 high-volume previously unclassified exact hosts, but verified operator-family labels cover only 53.6% of citation rows and 57.4% of search-result rows in the full inventory. Source-role results such as “official documentation” are therefore incomplete observed frequencies, not exhaustive source-category shares. Search layers also differ by surface; some do not have separate exports, and two of the five exported search lists closely mirror citations. The files omit conversations with no exported source. They cannot reveal internal retrieval decisions, establish causality, or show whether an answer meaningfully mentioned a brand without checking answer text.
Download the complete research deck
Get the charts, methodology, and findings in a presentation-ready format.
Download the PDFFind out what ChatGPT says about your brand.
Share your website. We’ll test real buyer questions specific to your brand in ChatGPT and deliver a reviewed analysis showing where your brand appears, how it is framed, which competitors and sources shape the answers, and what to do next.
Full Viziquo analyses include ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overview, Grok, Microsoft Bing AI Answers (Copilot Search), and Perplexity.
Free · No credit card needed · Delivered within 3 business days