Gemini, AI Overviews, and AI Mode cite differently

One page kept showing up.

For 19 days, Zumi, an AI Search Intelligence Platform, asked Gemini, Google AI Overviews, and Google AI Mode the same question about buying one. Together, the three engines cited 362 different domains, and 9 appeared in all three. One of those 9, demand-genius.com, was cited in 59 of the 77 answers collected.

One question, one category, 19 days: those counts describe this window, not a general property of any engine. A different question would likely move them.

It is not a famous site. That is the point. In this study, citations had a small stable core. The wording around those citations changed much more.

Key takeaways

A brand cited by one Google engine is not thereby cited by the other two. The three shared 9 of their 362 cited domains, and the site cited most often was not the best-known name in the category. It was the one whose page answered the question.

• Gemini, AI Overviews, and AI Mode shared 9 of 362 cited domains. A reading from one engine does not describe where a brand stands in the other two.
• All 9 shared domains were pages built to answer the question directly. No homepage and no general brand page reached the shared set.
• One domain, demand-genius.com, appeared in 59 of the 77 answers collected. Nothing else came close to that.
• 253 of the 362 domains, 70%, appeared in exactly one answer. Citations rotate, so a single snapshot describes one day rather than a position.
• Each engine keeps its own vocabulary, and that vocabulary moves. AI Overviews used "Answer Engine Optimization" in 5 of its first 9 tracked answers and 0 of its last 9.
• AI Mode cited 276 domains across 26 answers. Gemini cited 38 across the same number. The two are not the same measurement problem.

What was measured

The prompt never changed. Most other things did.

Which pages did all three engines agree on?

9, out of 362.

Every one of the 9 domains belongs to a company selling into this category. For this analysis, every cited domain was placed into one of 15 source types, including analyst firms, review platforms, news publishers, community forums, and academic sources. All 9 shared domains fell into one type: the vendor's own site.

Of the 9 domains, 4 were opened and read directly. The other 5 were counted from the citation list alone.

Demand-Genius's cited page is a landing page titled "AI Search Intelligence." It covers the same buying criteria as the question: sentiment tracking, stakeholder mapping, visibility tracking, gap identification, and revenue connection.

Getperspective.ai's cited page is a blog post that ranks ten platforms against five buying criteria. It includes its own comparison table.

HubSpot's cited page is a blog post titled "AI search strategy: A guide for modern marketing teams." It covers entities, schema, and structured data, then a seven-step implementation and an FAQ. It ranks no platforms and lists no buying criteria.

The Starr Conspiracy's cited page is also a comparison guide, but for general demand-generation software instead of AI search intelligence.

Two of these pages answered the question directly. The other two were cited anyway.

One of the 9 domains, tryprofound.com, belongs to Profound. Profound is still sixth of the 9 by citation count, with 12 citations compared with demand-genius.com's 59. A better-known company did not automatically win more citations. A page built around the question did.

This finding is about the 9 domains, not about Zumi and Profound.

What kind of sources do the engines prefer?

They mostly prefer the same kind.

All three engines built most of their answers from what vendors say about themselves.

That runs opposite to broader B2B AI-citation research. Separate studies put third-party sources, not vendor sites, at roughly 85% of citations for B2B category queries (Valasys/Rampiq, 2026), and found a recommended vendor's own site cited only 11.6% of the time across 40 B2B SaaS categories (DerivateX, 2026). This study covers one AI-search-intelligence-platform question, not the wider B2B category, and a smaller, narrower sample points a different direction than the larger studies do.

Analyst coverage and review-site presence are expensive to build. Across the three engines and 362 domains, those source types earned 0 to 5 citations each.

Did the sources stay the same?

No. When each day's cited domains were compared with the first day, the overlap stayed low: 13% on Gemini, 7% on AI Overviews, and 15% on AI Mode. Splitting the study window in half showed the same pattern.

The 19 days split into two stretches: July 20 to 29, and July 30 to August 7. Each count below is the number of distinct domains an engine cited across that stretch, deduplicated. The last column counts the domains that appear in both stretches, and states them as a share of the second stretch: of everything the engine cited in the later stretch, that much had already been cited in the earlier one.

AI Mode changed the most. It cited more sources later in the window and kept fewer of the sources it started with.

The churn also showed up in the long tail. Of the 362 domains, 253 appeared in exactly one answer. The same pattern held inside each engine: 71% of Gemini's sources were cited once, 77% of AI Overviews' sources were cited once, and 72% of AI Mode's sources were cited once.

So the pattern is a small stable core and a large rotating set of sources. demand-genius.com sat in the core. Almost nothing else did.

Did the story stay the same?

No. The answer wording changed almost as much as the sources did.

Each day's answer text was compared with day one, word for word, with links removed. The overlap stayed low. Comparing each day with the previous day gave similar results.

The two columns are close because each day reads like a fresh answer, not a lightly edited version of the previous one.

The three engines also read the same question in different ways.

Gemini also stayed inside one vocabulary. It used "share of voice" in 20 of 26 answers, the word "citation" in 24 answers, "sentiment" in all 26 answers, and some form of "hallucination" in 9 answers.

The other two engines used a wider, less steady vocabulary.

"Dark Funnel" names buyer research that happens where attribution software cannot follow it. The term mostly belonged to AI Overviews, which used it in 8 of 25 answers. Gemini used it twice; AI Mode used it once.

That split is examined in the dark funnel in AI search. The two optimization acronyms were not fixed inside an engine either.

AI Overviews dropped the term inside three weeks. AI Mode adopted it in the same window. Gemini's use of both terms stayed steady.

Formatting varied by engine too.

Every engine used bullets. Gemini regularly used bold section headers; the other two rarely did.

Demand-Genius shows the clearest split between being cited and being described. Gemini linked to demand-genius.com in 21 of 26 answers, but did not mention it in the surrounding sentence. AI Overviews mentioned it in the prose in 7 of 25 answers. AI Mode mentioned it in 11 of 26 answers.

Citing a page and describing the company behind it are different things. On this question, Gemini mostly did the first.

Region mattered much less than engine. United States and India answers on AI Overviews and AI Mode were close to identical on every measure taken. Gemini cited fewer sources in India, with 2.6 sources per answer compared with 3.6 in the United States, on a base of seven answers.

How different are the three engines?

The overlap stayed small when counted in pairs too.

Gemini works from a short list. AI Mode works from a much wider one. One content plan will not fit both.

What this changes about measurement

• The three Google engines need to be measured separately. They shared 9 of 362 domains. A single "Google AI visibility" number blends three engines that disagree on almost everything.
• A page earns the citation. The company behind it does not. All 9 shared domains were vendor-owned, and every page checked directly was built around the same evaluation criteria as the question. Profound still ranked sixth of the 9.
• Citation counts and answer wording move separately. 70% of cited domains appeared once and did not return. Wording overlapped with day one by 21% to 28%. Tracking cited domains alone will not show whether the surrounding story changed.
• Each engine's vocabulary can change. AI Overviews used "Answer Engine Optimization" in 5 of its first 9 tracked answers and 0 of its last 9. The terms used to describe this category can shift inside three weeks.

Google is not one AI engine. It is three, and each one answers differently.

Method note

This analysis uses one question. A wider study would show whether the vendor-page pattern and vocabulary drift also appear in other categories. The underlying data is available on request.

Zumi tracks which pages get cited and how the surrounding language shifts, across up to nine AI engines. Book a demo to see how a specific category is being cited.