AI assistants are rewriting the principles relating to visibility on the internet. The web sites they point out are chosen primarily based on completely different standards than conventional search rankings.
As an alternative of simply taking a look at which web sites are talked about probably the most, I wished to grasp whether or not these mentions really match the recognition of the matters being mentioned.
I in contrast every area’s Point out Share (how typically it exhibits up) to its Impression Share / Potential Attain (how typically you’d anticipate it to point out up primarily based on search quantity for these matters).
This comparability helps to uncover biases and present whether or not a specific system is leaning into sure sources roughly than anticipated primarily based on the recognition of the matters the web site covers.
I regarded on the prime 50 web sites cited in Ahrefs Model Radar for Google AI Overviews, ChatGPT, and Perplexity.
That is throughout ~76.7M AI Overviews, 957k ChatGPT prompts, and 953.5k Perplexity prompts for the month of June 2025.
- Google AI Overviews lean closely on UGC websites like Reddit, Quora, and YouTube. Wikipedia and well being websites like Mayo Clinic and Cleveland Clinic are under-represented. YouTube is owned by Google, so they could want it. Google has a licensing cope with Reddit and has given them a variety of visibility in Google Search already.
- ChatGPT under-represents Wikipedia and Information like Reuters.
- Perplexity is pretty impartial general, however under-represents Wikipedia.
- Wikipedia was barely under-represented in Perplexity, however under-represented fairly a bit in Google and ChatGPT. Regardless of Wikipedia’s recognition in AI assistants, it appears like these methods are pretty biased in opposition to them.
Right here’s the nerdy information.
Google appears to be counting on extra user-generated content material (UGC) websites than they’re reliable websites, these thought-about to have extra EEAT.
Listed here are a pair definitions to maintain in thoughts:
- Point out Share is how incessantly a website is talked about throughout all AI responses.
Point out Share = (Variety of responses that point out the area ÷ Complete variety of AI responses analyzed) × 100
- Impression Share weights mentions by Google search quantity to estimate how a lot potential visibility a website will get throughout high- vs. low-demand matters.
Impression Share = (Sum of search quantity for queries the place the area is talked about ÷ Complete search quantity of all AI-analyzed queries) × 100
The distinction between Point out Share and Impression Share tells you whether or not a web site is being cited extra in high-visibility queries or low-visibility ones. It reveals systemic biases in how AI assistants present completely different web sites.

Over-relying on:
- Reddit (7.4% vs 4.0%): talked about 3.4% greater than its anticipated visibility
- Quora (3.6% vs 1.4%): talked about 2.2% greater than its anticipated visibility
- YouTube (9.8% vs 7.8%): talked about 2.0% greater than its anticipated visibility
Beneath-utilizing:
- Wikipedia (8.4% vs 11.6%): talked about 3.2% lower than its anticipated visibility
- Mayo Clinic (2.9% vs 4.5%): talked about 1.6% lower than its anticipated visibility
- Cleveland Clinic (2.9% vs 4.5%): talked about 1.6% lower than its anticipated visibility
You need to use Ahrefs Model Radar to see when these biases might have been launched.
For instance, listed below are the Mentions for the highest 5 websites in AI Overviews over time, however weighted to the market. It appears like Google is shifting away from Wikipedia, shifting in the direction of Reddit closely in December and March, in the direction of Google Translate in March, in the direction of Quora in December and March, however away from Quora in April, and should have moved away from YouTube again in October, however introduced it again in April.
That is the way you see algorithm updates and tendencies within the new period.


ChatGPT was the system that I believed would have probably the most radical variations. Total, they appear to under-represent Wikipedia and information websites.
Another definition:
- Potential Attain Share (%) estimates how a lot publicity a website may be getting in an AI assistant responses. There is no such thing as a dependable search quantity for AI assistants, so this weights in search quantity from Google searches and acts as a proxy primarily based on historic internet recognition.
Potential Attain Share = (Sum of search quantity for prompts the place the area is talked about ÷ Complete search quantity for all prompts analyzed) × 100


Over-relying on:
- Google.com (2.3% vs 1.7%): 0.6% over
- Apple.com (4.0% vs 3.8%): 0.2% over
Beneath-utilizing:
- Wikipedia (16.3% vs 19.3%): 3.0% below
- Reuters (4.3% vs 6.4%): 2.1% below
I used to be anticipating Perplexity to bias in opposition to Wikipedia extra. The CEO has made some feedback about Wikipedia’s bias and even provided to help anybody who wished to construct an alternate. I couldn’t have been extra mistaken.
Perplexity seems to be probably the most balanced, displaying point out patterns that roughly align with matter recognition on the normal internet.


Over-relying on:
- YouTube (16.1% vs 15.7%): 0.4% over
- Apple.com (3.5% vs 2.6%): 0.9% over
Beneath-utilizing:
- Wikipedia (12.5% vs 13.4%): 0.9% below
- Tuasaude.com (1.6% vs 2.5%): 0.9% below
Closing ideas
Search is fragmenting. We’re so used to only optimizing for Google. Now we would want to have a look at optimizing for various methods with completely different biases.