Google’s AI Mode generates responses which can be 4x longer than AI Overviews (on common). After we first observed this, the pure assumption was that AI Mode merely expands on the identical info, taking AI Overview’s concise reply and including extra element from the identical sources.
However after analyzing 730,000 response pairs, we discovered one thing surprising: AI Mode and AI Overviews attain very related conclusions (86% semantic similarity) whereas citing completely different sources (solely 13.7% quotation overlap).
This issues as a result of it suggests these aren’t only a “brief model” and “lengthy model” of the identical reply. They’re two distinct techniques that occur to converge on related conclusions via completely different paths.
For entrepreneurs and search engine optimisation professionals, this raises vital questions:
- In case your model is cited in AI Overviews, will it seem in AI Mode too?
- Do you want separate optimization methods for every?
- What does it imply when two AI techniques agree on what to say however not the place they discovered it?
We analyzed quotation patterns, content material similarity, and entity mentions throughout a whole lot of 1000’s of queries to seek out out. Right here’s what we found.
- Quotation overlap: Solely 13.7% of citations overlap between AI Overviews and AI Mode, indicating they cite completely different sources in observe.
- Phrase-level overlap: There’s very low (16%) word-level overlap between AI Overviews and AI Mode. They begin with the identical first sentence solely 2.51% of the time.
- Semantic similarity: 9 out of 10 occasions, AI Mode and AI Overview agreed on what to say; they simply mentioned it otherwise and cited completely different sources.
- Entity overlap: AI Mode responses embrace 2.5x extra individuals and model entities than AI Overviews. However, in case your model will get talked about in AI Overviews, there’s a 61% likelihood it’ll additionally seem in AI Mode’s longer response.
- Model gaps: 32.8% of responses had no particular person or model talked about at all.
- Quotation gaps: For 3% of responses, AI Mode doesn’t cite any sources. AI Overviews don’t cite any sources 11% of the time.
We analyzed September 2025 US knowledge from Ahrefs’ Model Radar, inspecting 540,000 question pairs for quotation and URL evaluation, and 730,000 question pairs for content material similarity evaluation. For every question, we captured each an AI Mode response and an AI Overview response.
We measured quotation overlap by figuring out what number of URLs appeared in each AI Mode and AI Overview responses for a similar question. We additionally tracked area preferences to see which web sites every platform cited most regularly.
For content material similarity, we used two metrics: Jaccard similarity to measure word-level overlap (calculated as distinctive phrases in widespread divided by all distinctive phrases), and cosine similarity to measure semantic overlap on a scale from 0 (fully completely different) to 1 (an identical which means).
We additionally analyzed entity overlap by counting how many individuals, organizations, and types have been talked about in each responses.
Sidenote.
This evaluation compares single generations of AI Mode and AI Overview responses. Our earlier analysis confirmed that 45% of AI Overview citations change between generations. This implies the quotation swimming pools out there to every system could overlap greater than our single-snapshot comparability suggests. Nonetheless, the low overlap we noticed (13.7%) signifies that even when a number of sources might help a solution, AI Mode and AI Overviews typically choose completely different ones in observe.
AI Mode and AI Overviews cited the identical URLs solely 13.7% of the time. After we checked out simply the highest 3 citations from every, the overlap was barely increased at 16.3%. Which means that 87% of the time, these two techniques are pulling from fully completely different sources to reply the identical question.
The area preferences additionally differed between the 2 techniques:
Probably the most notable variations:
- YouTube held the highest place in AI Overviews, being cited greater than encyclopedic sources
- Wikipedia appeared in 10% extra AI Mode citations than for AI Overviews
- Whereas Reddit was cited an identical quantity, Quora appeared 3.5x extra in AI Mode citations
- Well being web sites have been cited virtually twice as a lot in AI Mode as in AI Overviews
- Fb was additionally cited twice as a lot in AI Mode as in AI Overviews
AI Mode seems to lean extra closely on encyclopedic and detailed medical sources because it builds longer responses, whereas AI Overviews present a stronger desire for video content material and community-driven platforms like Reddit.
Nonetheless, each techniques additionally confirmed related preferences for content material varieties:
AI Overviews cited movies and core pages (like homepages or principal class pages) practically twice as typically as AI Mode, whereas each techniques overwhelmingly most popular article-format content material.
Do that your self
You possibly can see what sources AI Overviews and AI Mode cite essentially the most in Ahrefs’ Model Radar. Run a clean search and examine the Cited domains report.
Filter by AI Overview and AI Mode individually to see their most popular sources.
You may as well filter on your goal matters or particular manufacturers if you wish to analyze a extra particular section of the information. Try our information to the most effective filter mixtures to attempt.
AI Mode and AI Overview responses had a Jaccard similarity rating of simply 0.16, which means solely 16% of the distinctive phrases overlapped between the 2 responses for a similar question.
They began with the very same first sentence solely 2.51% of the time, and produced an identical responses in simply 0.51% of circumstances.
This tells us that AI Mode isn’t merely taking AI Overview’s reply and including extra element to it. As an alternative, each techniques seem to analysis the question independently and formulate completely new responses with minimal phrase or quotation overlap.
Even after they agree on the reply (which they do 86% of the time semantically, as we’ll see subsequent), they’re expressing it in basically other ways.
Right here’s the place issues get fascinating: regardless of low phrase overlap (16%) and minimal quotation overlap (13.7%), AI Mode and AI Overview responses achieved a semantic similarity rating of 86% on common.
We measured this utilizing cosine similarity. On a scale the place 1.0 means “an identical which means” and nil means “fully unrelated,” practically 90% of response pairs (89.7%) scored above 0.8, indicating sturdy semantic alignment.
Put merely: 9 out of 10 occasions, AI Mode and AI Overview agreed on what to say. They simply mentioned it otherwise and cited completely different sources.
There are a lot of doable the explanation why this may occur. Nonetheless, Google’s documentation confirms that each AI Mode and AI Overviews use “question fan-out” to assist show a wider and extra various set of useful hyperlinks.
Question fan-out is a course of that runs a number of associated searches to seek out supporting content material whereas responses are being generated. Since AI Mode and AI Overviews use completely different fashions and strategies, they will simply cite completely different sources even when reaching related conclusions.
This explains why they agree on what to say whereas disagreeing on the place they discovered it.
Consider it like two specialists answering the identical query.
They may use fully completely different phrases and reference completely different research, but when they’re each educated in regards to the subject, their solutions will convey the identical core info. That’s what’s taking place right here; the techniques are drawing from a constant understanding of every subject, whilst they specific it otherwise.
AI Mode responses are roughly 4x longer than AI Overviews and point out considerably extra individuals and types (3.3 entities on common in comparison with AI Overviews’ 1.3).
For instance, trying on the key phrase “cloud storage alternate options”, the AI Overview mentions seven manufacturers solely as soon as every, and towards the top of the response:
Whereas the AI Mode response consists of 23 model mentions, repeating some model mentions all through the reply. Additionally, each part of the response comprises a model, together with the very first sentence:
It’s straightforward to assume that since AI Mode solutions are 4x longer, that’s naturally why in addition they embrace extra manufacturers in absolute phrases. However it’s typically the case that AI Overviews is extra selective with whether or not it mentions manufacturers, limiting the quantity of name publicity out there throughout the response itself.
However right here’s what’s maybe most fascinating: 61% of the time, AI Mode consists of each entity that AI Overview talked about, then provides extra on high.
For instance, an AI Overview would possibly point out Mayo Clinic as a well being authority. AI Mode’s response consists of Mayo Clinic but additionally provides Cleveland Clinic and WebMD. The core authority seems in each, however AI Mode expands the professional pool.
Which means that in case your model is talked about in AI Overviews, there’s a great likelihood it’ll additionally seem in AI Mode’s longer response.
However you’ll be sharing house with further rivals or sources that didn’t make the AI Overview reduce.
This sample suggests AI Mode is constructing on an identical basis of associated entities as AI Overviews relatively than ranging from scratch with a completely completely different strategy.
Not each response consists of citations or model mentions. For example, 59.41% of AI Overview and 34.66% of AI Mode responses include no manufacturers or entities.
About one-third (32.8%) of all AI responses in our dataset talked about no manufacturers or individuals in any respect. These are normally informational queries the place no model is looked for or anticipated within the response, similar to “Nov 21 zodiac”, “income cycle”, and “meditation earlier than mattress”.
Some responses additionally confirmed no cited sources. When citations are lacking completely, it’s usually for edge circumstances like:
- Easy calculations (“4 divided by 1/2”)
- NSFW or delicate content material (“sexual abuse,” “my spouse hit me”)
- Redirects to assist facilities
- Unsupported languages (“678是美国哪里的区号”)
Why the distinction in quotation charges?
AI Mode is extra dependable for attribution. Solely 3% of responses lack citations in comparison with 11% for AI Overviews.
A number of doable causes:
- Size and format necessities: 4x longer responses want extra grounding and supporting proof
- Consumer expectations differ: AI Overviews seem in conventional search outcomes, the place customers typically simply want a fast reality. AI Mode is an interactive chat or analysis expertise the place customers count on transparency about sources
- Question filtering: AI Overviews possible floor for extra edge circumstances that don’t warrant citations, whereas AI Mode could filter these out or deal with them otherwise
These distinctions between AI Mode and AI Overviews have direct implications for the way you strategy AI optimization.
- Observe visibility for AI Overviews and AI Mode individually: With solely 13.7% supply overlap, being cited in a single doesn’t assure visibility within the different. Monitor your visibility in each by utilizing Ahrefs’ Model Radar.
- Concentrate on semantic authority, not actual wording: The 86% semantic similarity reveals each techniques look for a similar themes, simply expressed otherwise. Construct topical authority and complete protection (relatively than concentrating on particular phrases) so your model is seen as related to the subject.
- Think about the format variations: AI Mode’s 97% quotation price and entity enlargement (3.3 vs 1.3 entities) favor longer, well-sourced content material, particularly from encyclopedic sources. AI Overviews desire video and neighborhood platforms like Reddit. The identical content material could not carry out equally in each.
- Put together for extra competitors in AI Mode: For those who’re cited in an AI Overview, there’s a 61% likelihood you’ll seem in AI Mode too, however alongside further rivals who didn’t make the shorter reduce.
- Put money into encyclopedic content material: Wikipedia seems in 28.9% of AI Mode citations versus 18.1% in AI Overviews. Think about how your content material can function a complete reference or be cited in present encyclopedic assets.
The underside line: deal with AI Mode and AI Overviews as separate channels with overlapping targets however completely different execution. Optimize for each, however don’t assume success in a single interprets to the different.
