TL;DR
- Keep AI Mode and AI Overviews as distinct observations, even when the buyer question is the same.
- A high semantic similarity score does not mean the same sources, brands or facts appear in both answers.
- Research found different source patterns under its recorded conditions. Those historical averages are not your brand’s current chance of being cited.
- Compare the saved answers on meaning, brand presence and exact linked pages. Publish shared facts clearly; do not create duplicate pages just to chase each feature.
Name the Google Feature Behind the Result
A dashboard labeled “Google AI visibility” can hide two questions: which feature did it observe, and what counted as presence? For a B2B team, that can change the next action. A product might be named in an answer while a rival’s guide gets the source link. A report needs to preserve both facts.
Similar Meaning Does Not Imply Shared Sources
Ahrefs’ December 2025 comparison used September 2025 US data. Its citation analysis covered 540,000 query pairs; content similarity used 730,000. The reported citation overlap was 13.7%. Mean semantic similarity was 0.86 on its cosine measure, while unique-word overlap was 0.16 on a Jaccard measure.
Those are different measurements. A mean score of 0.86 is not a count showing that 86% of answers agree. Citation overlap is also not the chance that your brand transfers from one feature to the other. Ahrefs describes single generations per feature; a new run could select other links.
For a buyer, shared meaning can be useful: two answers might both advise checking access controls. For a supplier, the source pages still matter. One answer could cite its setup guide, while the other cites a trade publication. Neither semantic score tells us that the supplier was named, recommended or linked in both.
A Separate Study Found Different Source Breadth
Semrush’s July 2025 comparison used 5,000 keywords, balanced across four intent groups and two search-volume tiers, in desktop results. It reported an average of three unique linked domains per query in AI Overviews and seven in AI Mode. This is a count of distinct domains, not a share of buyers or a number of clicks.
More domains can give a reader more sources to inspect. It does not show that any one vendor receives more attention. Nor does it establish that a B2B supplier should lengthen all its pages. The study did not test the effect of that edit.
The studies complement each other rather than validate one common percentage. Ahrefs compares source overlap and answer content. Semrush describes source breadth in a different sample. Both give reasons to retain the feature label; neither supplies a conversion factor between two visibility scores.
Semrush report, July 21, 2025; 5,000 keywords with balanced intent and volume groups. Collection dates are not clearly specified. Counts are historical study means, not visitor reach or current brand odds. Semrush study and method.
What the Evidence Can and Cannot Carry Into a Report
| Direct source | Actual evidence | Reporting boundary |
|---|---|---|
| Ahrefs, December 2025 | 13.7% citation overlap; mean cosine score 0.86; word Jaccard 0.16 | Different units; one-generation comparison, not brand transfer odds |
| Semrush, July 2025 | Mean unique linked domains: three in AI Overviews, seven in AI Mode | Different sample; domain breadth does not measure clicks or supplier reach |
| Google’s current guidance | The features may use different models and techniques, so answers and links vary | Official product explanation, not an independent effect study |
Figure 2. Evidence comparison. The table presents actual research results alongside their measurement limits. Ahrefs and Semrush both sell search tools. Their query samples do not establish the mix of real B2B buying questions, and their results should not be pooled into one platform ranking.
Google’s site-owner guidance, checked November 9, 2026, distinguishes the two experiences and says their models and techniques may differ. It also says an AI Overview often does not trigger. A missing Overview therefore needs its own state; it is not an answer in which your brand was considered and rejected.
This explanation fits the research without proving the cause of each missing citation. A public product description cannot tell an analyst why one page was selected in a saved answer. Keep that distinction when moving from a chart to an action plan.
A Paired Review Can Change the Recommended Action
Consider an illustrative record for a fictional access-management vendor. The analyst saves both outputs for the same question about audit logs, in the same market, language and device setup. The entries below are invented to show the review, not measured SemShift results.
| Review field | AI Overview record | AI Mode record |
|---|---|---|
| Observed state | Overview present; brief answer saved | First-turn answer saved in a fresh conversation |
| Meaning | Advises checking log export and retention | Advises the same checks and asks about storage needs |
| Vendor presence | Vendor not named | Vendor named as one option |
| Exact source link | Trade publisher’s audit-log guide | Vendor’s export guide |
| Claim accuracy | No vendor-specific claim to verify | Says export is included in every plan; approved plan record says otherwise |
| Action | Record absence under this feature; inspect competing source coverage | Investigate the plan-scope error and whether source wording is clear |
A single “visible on Google AI” label would hide the useful finding. The team has a scope error in one answer and no brand mention in the other. The immediate content task is to check the export guide against approved plan facts. It is not to declare success or write a second page with the same content.
If the source guide is vague, make the plan limit clear there and keep related pages consistent. If it already states the limit, record the generated-answer error. A website change cannot force either feature to repeat the corrected fact. Retain the original answer so a later check can distinguish a correction from a new sample.
For an audit, use the same core buyer tasks across features where possible, but keep the entry point and follow-up context in the record. A first-turn AI Mode answer and a later conversational answer are different observations. If an Overview does not appear, record “not triggered” separately from a generated answer with no brand mention.
Repeated observations are still needed to assess stability. Our snapshot and trend guide covers that separate problem. The task here is to avoid substituting one Google feature for another in the first place.
AI Mode and AI Overviews: Common Questions
Should we build a separate page for each feature?
Not just because the features differ. Start with the buyer’s question and the facts you can support. Create distinct pages when the reader tasks differ. These studies did not test duplicate feature-specific pages as a way to win citations.
Can the two features share a content plan?
Yes. They can need many of the same product facts and evidence. Keep observed results distinct so a shared plan does not become a claim that both features showed the same brand or page.
Does more source breadth make AI Mode better for our brand?
The reported count does not answer that. Inspect whether your brand appears, how it is described and which exact page is linked. More domains can also mean more competing sources.
Can we combine the results into one summary?
You can give a clearly labeled aggregate with its counting rule and component results. It should not stand in for a feature-level finding. Keep missing generations and non-triggered Overviews visible, and avoid treating the sample as real audience reach.
Bring the Feature-Level Evidence
If your report combines AI Mode and AI Overviews without saved outputs, bring the question set, feature labels, answers and source URLs to SemShift’s AI visibility audit. We can discuss which findings the record supports and which buyer tasks or product facts need a closer review.