Do Backlinks Help AI Citations? 2,880 Queries Compared Against 2.8 Billion Links

This article was first published in Korean by Intrix Lab on August 16, 2026.

The results were opposite depending on the model. Across 6 industries, we asked two models the same 144 questions ten times each, for 2,880 calls in total. We then checked backlinks for the 2,672 domains that appeared against 2.8 billion public web links (measured on 2026-08-16). In GPT-5.6, the top 25% by link count received 5.66 times as many citations as the bottom group. In Gemini-3.7, more links were associated with fewer citations. The same link metric had a correlation of +0.272 in one model and -0.144 in the other.

Why did we measure this?

Until now, our recommendations have centered on "get indexed." In an earlier experiment, Google indexing was associated with a 9.53-fold difference in AI citations, while the effect of published page count disappeared when indexing was included. Publication volume was effectively a proxy for indexing.

But that leaves the next question unanswered. What drives indexing? The industry's standard answer is backlinks. We therefore checked directly against a public dataset of 2.8 billion web links.

How did we measure it?

We covered six industries. The YMYL industries were dermatology, plastic surgery, and legal services. The non-YMYL industries were website development, marketing agencies, and video production. We matched the question composition so that differences in questions would not obscure differences between industries.

We asked each question repeatedly because a single measurement is insufficient. Even in this dataset, 48% of the domains GPT cited appeared exactly once in ten runs. Half the domains on a list from a single run disappear the next time.

Does AI cite sites with more backlinks more often?

The answer depends on the model. We divided domains into quartiles by link count and examined their actual citation counts.

GPT-5.6 cited the bottom 25% by link count an average of 1.12 times and the top 25% an average of 5.66 times. That is a fivefold difference. The rank correlation in the healthcare-related group was +0.272, the largest value in this experiment.

Gemini-3.7 showed the opposite pattern. The group with 0 links averaged 3.95 citations, the middle group 2.53, and the top group 4.40. The correlation was negative at -0.144. The direction was the same in all four combinations of the two industry groups and the link count and quality metrics.

The citation targets suggest why. The 1,878 individual businesses Gemini cited most often had a median of just 4 inbound links. By contrast, the 207 professional societies and public institutions GPT cited frequently had a median of 93. Gemini primarily cites individual businesses with few links, while GPT primarily cites institutions with many links.

Does the link effect disappear after controlling for indexing?

It did not. That was the central hypothesis we preregistered before starting this experiment. Just as publication volume had been a proxy for indexing, we expected links to work only through indexing.

The results showed otherwise. Comparing before and after controlling for indexed page count, GPT's correlation in the healthcare-related group barely declined, from +0.272 to +0.256. Gemini's moved from -0.144 to -0.131. In the non-healthcare group, Gemini's correlation actually strengthened, from -0.131 to -0.153.

This means our hypothesis was wrong, and that our recommendations were missing a factor. Links do not operate only through indexing; they have a separate pathway.

What about Google rankings?

The relationship was weaker than expected. The correlation between link count and high Google rankings ranged from +0.115 to +0.152. That is too small to explain even 2% of ranking variation.

More strikingly, in the healthcare-related group, links explained GPT citations (+0.272) more than twice as well as Google rankings (+0.115). Our preregistered expectation was the opposite.

What does this mean in practice?

Allocating budget to link building makes sense when the goal is GPT citations, specifically in regulated fields such as healthcare or law. If Gemini is the target, we found no supporting evidence in this dataset. If Google rankings are the goal, it would be worth reassessing the cost relative to the effect.

There is one more point. The bottom 25% by link count had a median of 0 inbound links. More than half the businesses in that group appeared in results for queries with top-ranking results without any links. Links are not a prerequisite for entry.

Limitations of this experiment

We will start with the most important limitation. This is an observational study, so we cannot say that "adding links increases citations." The direction could be reversed. Sites may have attracted links because they were cited and ranked highly. In particular, professional societies and courts, which GPT cited frequently, are institutions that naturally attract many links. Establishing causation requires a randomized experiment, and we have preregistered a separate design for that.

Our control for indexing was also coarse. We counted indexed pages only up to a fixed ceiling, so large sites could not be distinguished from one another. Refuting the central hypothesis means that controlling at this level of precision did not eliminate the link effect. It does not guarantee the same result with a more precise metric.

The sample consisted of domains that appeared in response to our queries. Businesses with neither links nor rankings were excluded from the outset. These figures therefore describe a relationship within the top-ranking group, rather than across the full range. We also covered only six industries and do not generalize to industries overall.

These results are also based on APIs. They may differ from what users see in an app. In fact, when we asked the same questions through a different API interface from the same company, there was no overlap at all in cited domains (a preliminary observation from 3 queries). We plan to test this in a separate experiment as well.

Finally, none of these figures guarantees a particular outcome. Correlation coefficients describe average tendencies. There is no guarantee that an individual site will follow those tendencies.

Frequently asked questions (FAQ)

Q. Should we invest in backlinks or not?
Define the goal first. If you want GPT citations and work in healthcare or law, that is where we found the strongest relationship in this dataset. If you want Gemini citations, we found no supporting evidence. However, this is an observational study and does not guarantee an effect.

Q. Why do the models produce different results?
It appears to be because they cite different types of sources. Gemini primarily cited individual businesses with a median of 4 inbound links. GPT primarily cited professional societies and public institutions with a median of 93. The same link metric means different things for the two models.

Q. Can we apply these figures directly to our site?
That is difficult. These are average tendencies among domains that appeared near the top across six industries. New or small sites that did not enter the sample may behave differently. Our own three domains were also absent from the link dataset.

Q. Did you choose the analysis after seeing the results?
No. We set the hypotheses and analysis methods before starting data collection. Choosing an analysis after seeing the results makes it possible to produce the picture you want.