AI Mode vs AI Overviews: Why You Should Measure Them Separately

Being cited in Google AI Overviews does not guarantee that the same page will be cited in AI Mode. Treat them as separate measurement surfaces, while keeping a shared foundation of useful content and technical SEO.

That is a more practical conclusion than building two versions of every page.

What did the research compare?

Ahrefs’ December 2025 study used September 2025 US data. It examined 540,000 query pairs for citation analysis and 730,000 pairs for content similarity. The distinction matters: the larger number was not the sample size for every result in the study.

The reported cited-URL overlap was 13.7%. Yet average semantic similarity was 0.86 on the study’s cosine-similarity measure, and 89.7% of response pairs scored above 0.8.

In other words, answers could convey similar meaning while drawing on different supporting pages. The 0.86 score is not the percentage of statements that were correct or a measure of agreement by human reviewers.

Why might sources differ?

Google explains that both experiences may use query fan-out: multiple related searches across subtopics and sources. It also says the products may use different models and techniques, so responses and links can vary.

This gives publishers a reason to provide useful supporting detail across a topic, not merely repeat a target phrase on one landing page.

For example, a service page can explain the commercial offer while a guide explains migration, a comparison examines alternatives and a case study documents a result. Those pages serve different information needs without becoming near-duplicate landing pages.

What does the overlap figure not prove?

Ahrefs compared single generations for the matched queries. The authors explicitly note that changing citations across generations may mean the underlying source pools overlap more than the snapshot suggests.

The 13.7% figure is therefore not a permanent law of Google search. Nor does it establish that separate content teams or technical schemas are required.

It supports a narrower conclusion: do not use success on one surface as proof of success on the other.

How should you build a useful monitoring process?

Choose a documented question set covering branded, category, operational and comparison intent. Capture both experiences where available, using comparable conditions. Record the answer, date, cited URLs and brand description.

Repeat the exercise. A single answer can reveal an error, but it cannot establish a stable visibility rate.

Then compare patterns. Are important pages absent everywhere? Is a particular topic missing? Are competitors cited because they provide a specific explanation you do not? Distinguish those gaps from random changes in source selection.

What should you change on the website?

Improve what you can substantiate: accessible content, accurate service descriptions, clear explanations, original evidence and useful internal links. Google’s published requirements do not call for separate AI Mode markup.

Stratozphere’s recommendation is to separate the observation, not automatically duplicate the production effort.