Backlinks from contextually relevant pages track AI rankings

What I found

• All 21,600 similarity curves passed the relevance test. The pages linking to or mentioning a recommended brand were more similar to the most closely related quarter of categories than to the least related quarter.
• In all 17 AI channels, the number of backlinks from contextually relevant pages had the strongest correlation with AI score. For GPT-6 Astra, the average correlation within industries was 0.57, and it was positive in all 100 industries.
• Relevance still matters after accounting for the total number of backlinks. The remaining correlation is 0.18 to 0.20 for the GPT-6 models, the Gemini Flash models and ChatGPT web.
• The same relevance pattern appears on pages that mention a brand without linking to it. For GPT-6 Astra, those mention counts correlate with AI score at 0.41, compared with 0.55 for total backlinks.
• The brands AI recommends first have been established for years. Of the top 10 brands with PageRank records, 94% already had PageRank in July 2017.
• The models also recommended domains that do not exist. Of 117,453 distinct domains in the answers, 28,674 (24.4%) were not registered.

How I tested it

For each of 100 industries, I asked for the top 25 recommendations 25 times, using 25 slightly different wordings. That gives 2,500 prompts. None specifies a brand, persona, city or price. I sent those prompts to 13 OpenAI models, three Gemini models and the ChatGPT web interface, then combined each model’s answers into an AI score. Position 1 earns 1 point, position 2 earns 1/2 point, and so on.

Next, I searched two OppAlerts indexes for three groups of pages: pages linking to each brand, pages mentioning it, and pages mentioning it without a link. One index covers news sites, blogs and podcasts; the other covers link and marketing opportunities. The backlink analysis includes 28.4 million page-and-brand pairs. I compared each page with 5,970 categories from four taxonomies: Google Ads Topics, IAB Content, IAB Ad Product and NAICS.

This compares brands as they were. I did not add links to a brand and then measure what changed.

What predicts a higher AI score

For each measure, I calculated its Spearman correlation with AI score within each industry, then averaged those correlations across the 100 industries:

The percentage of a brand’s backlinks that are relevant has almost no relationship with AI rank. Large brands get discussed in more contexts, so their backlink profiles are broader. For example, about 55% of Marriott’s backlinks come from pages strongly about hotels. Kimpton’s share is 93%. Even so, Marriott has about 21 times as many relevant backlinks.

The size of the gap

Here are the median counts for GPT-6 Astra’s recommendations, using the industries represented at each AI rank:

The size of that gap depends on your industry. These are the median relevant backlink counts for brands in the top 10:

Mentions without links

The content around unlinked mentions is almost as relevant as the content around links. On average, the unlinked mention curve is about 90% as steep as the link curve. But links have a stronger relationship with AI score. For a typical brand, one additional backlink corresponds to about as much AI score as 10 additional unlinked mentions. Across six models, that ranges from 6 to 15. For brands named in Google’s AI Overviews, one backlink corresponds to about 1.4 unlinked mentions.

Established brands stay ahead

I grouped GPT-6 Astra’s recommendations by their current AI rank, then looked back at their PageRank history. The groups were already in the same order in July 2017. The top 10 had a median PageRank of 35 then and 37 by September 2026. Ranks 51 to 100 had a median of 13 at both points. The length of time a site has had PageRank, on its own, correlates with AI score at only 0.14 to 0.21.

Hallucinated domains

I checked every domain against ICANN zone files, RDAP and DNS on October 6, 2026. All 28,674 unregistered domains came from the 17 AI channels. Google’s organic results returned no unregistered domains for the same 2,500 searches.

• The further down the answer you go, the more common hallucinated domains become: 1.13% at position 1, rising to 4.73% at position 25.
• After combining scores across the 17 channels, no hallucinated domain made an industry’s top 25.
• Industries with fewer relevant backlinks had more hallucinated domains. In cosmetic dentistry, 21.1% of recommended domains were unregistered. For colleges and universities, that was 0.02%.
• Eleven of the 17 channels recommended the unregistered tour-edge.com across 91 answers. The golf brand’s actual domain is touredge.com.

What the study does not show

• My view is that the relationship is causal. But this study does not prove causation: I compared brands as they were, without adding links to measure the effect.
• This does not give you a formula for how many links will move a brand a certain number of positions.
• I used broad industry prompts. They do not cover every persona, location, price range or user.
• The OppAlerts indexes cover part of the web. A brand with few pages in these indexes may have plenty of coverage elsewhere.
• After accounting for total backlinks, the mix of opportunity types, such as guest posts and events, adds almost nothing. The exception is listicle inclusion, with a correlation of 0.13 for GPT-6.

How to use this in your campaigns

• Measure how many of your backlinks come from contextually relevant pages, along with your total backlink count.
• Find relevant pages that already mention your brand and ask for a link.
• Use the brands AI recommends in your own industry to set your goals. The gap ranges from single digits to more than 17,000 relevant backlinks.
• Plan for ongoing work. The brands leading AI rankings were already established nine years ago.
• Repeat your prompts and measure more than one model.