Something Inc.LoginSchedule a free consultation
LINK BUILDING

You Cannot Buy 32,000 Referring Domains

The brand mentions vs backlinks slide quotes one study and skips the bigger one that says the opposite. Both are honest. The fight is about statistics, not about links.

LINK BUILDINGAI SEARCHNOV 2026

There is a slide going around. You have seen it in a pitch deck this quarter, or in a webinar, or screenshotted into a thread with a one word caption. Two bars. Brand mentions at 0.664. Backlinks at 0.218. The takeaway is that links are finished and mentions are the new currency, and the person showing it to you is not lying.

They are reading one chart, from one study, about one surface, and presenting it as a law of physics. There is a second study. It is larger, its method is more careful, and it reaches the opposite conclusion. Almost nobody shows that one. Not because it is buried or paywalled. It sits on a vendor blog under a boring headline, and it does not fit the slide.

So here is where the brand mentions vs backlinks argument gets interesting, and it is not where either camp wants to have it. The two sides are not disagreeing about links. They are quoting two different statistics, computed on two different units of analysis, against two different target variables. Both are honest work. Neither one answers the question you actually walked in with, which is what goes on next quarter's invoice.

TL;DR · 60 SECONDSThe evidence. Ahrefs published two studies of 75,000 brands by Louise Linehan, the first on May 26, 2025 against AI Overviews and the second on December 12, 2025 across ChatGPT, AI Mode and AI Overviews, both using Spearman rank correlation on data from its own Brand Radar and Site Explorer. The first ranked branded web mentions at 0.664 and number of backlinks at 0.218, with referring domains at 0.295 and Domain Rating at 0.326 in between. The second put branded web mentions at 0.709 on AI Mode, 0.664 on ChatGPT and 0.656 on AI Overviews, with YouTube mentions around 0.737, and reported backlinks only as very weak without publishing a figure. SE Ranking published a different kind of study on November 24, 2025, by Yulia Deda: 129,000 domains and 216,524 pages across 20 niches, target variable built from 100,000 parsed prompts, modelled with XGBoost regression and interpreted with SHAP. It found referring domains to be the strongest single predictor of ChatGPT citations. Our verdict: the studies do not contradict each other, because rank correlation cannot see the step function that the gradient boosted model found, and the step appears at 32,000 and 350,000 referring domains, which is a company size rather than a budget. Fund links against rankings, fund mentions against citations, and stop letting a correlation coefficient do a causal job.
0.664
Spearman correlation between branded web mentions and AI Overview brand visibility in Ahrefs' study of 75,000 brands, published May 26, 2025
0.218
the same study's correlation for number of backlinks, the figure that became the industry's whole position on link building
129,000
domains in SE Ranking's November 24, 2025 model, across 216,524 pages and 20 niches, which found referring domains the top predictor of ChatGPT citations
~0.65
correlation Seer Interactive measured between ranking on page one of Google and being named in an LLM answer, across 10,000 GPT-4o answers

The brand mentions vs backlinks comparison comes from a single Ahrefs study of 75,000 brands, published May 26, 2025, which ranked Spearman correlations against AI Overview brand visibility and put branded web mentions at 0.664 and number of backlinks at 0.218.

Louise Linehan wrote it, Xibeijia Guan contributed the data, and the full table is published, which is more than most vendor research manages. Worth reading the whole column rather than the two bars that got screenshotted. Branded web mentions 0.664. Branded anchors 0.527. Branded search volume 0.392. Domain Rating 0.326. Referring domains 0.295. Branded traffic 0.274. Number of backlinks 0.218. Ad traffic 0.216. Ad cost 0.215. URL Rating 0.18. Number of site pages 0.17.

Linehan's own framing describes these as moderate to very weak, and the study says in plain text that correlation is not causation. The research is careful. The slide built from it is not, and the gap between those two things is the entire problem.

“Correlation isn't causation.”

That line is in the study. It survives roughly none of the retellings. The December 12, 2025 follow up from the same author split the analysis across three surfaces and found branded web mentions at 0.709 on AI Mode, 0.664 on ChatGPT and 0.656 on AI Overviews, branded search volume at 0.466, 0.352 and 0.392, and Domain Rating at 0.285, 0.266 and 0.326. YouTube mentions came in highest of anything tested at roughly 0.737. Backlinks were described as very weak across all three systems, with no coefficient published at all.

Then there is the sampling frame, which almost nobody quotes and which changes what the study is allowed to claim. Both Ahrefs studies selected brands from domains with Domain Rating above 40 that already had a top keyword pulling at least 800 monthly searches. Every brand in the sample was already established. The research therefore cannot tell you what links do for a company below DR 40, because it did not look at one. If you are a Series A company with 300 referring domains, the most quoted chart in AI search was computed on a population you are not in.

STUDYPUBLISHEDSAMPLESTATISTICAL METHODTOP PREDICTOR FOUNDWHERE LINKS LANDED
Ahrefs, AI Overview brand visibilityMay 26, 202575,000 brands, DR above 40Spearman rank correlationBranded web mentions, 0.664Backlinks 0.218, referring domains 0.295
Ahrefs, cross surface brand visibilityDec 12, 202575,000 brands, three AI surfacesSpearman rank correlationYouTube mentions, about 0.737Described as very weak, no figure published
SE Ranking, ChatGPT citationsNov 24, 2025129,000 domains, 216,524 pages, 20 nichesXGBoost regression with SHAP attributionReferring domains, strongest single featureTop of the model, with a threshold shaped curve
Seer Interactive, LLM brand mentionsJan 7, 202510,000 GPT-4o answers from 300,000 plus keywordsCorrelation against SERP positionGoogle page one ranking, about 0.65Weak or neutral as a direct effect

Four studies. Four methods. Four different questions. The row that gets quoted is row one, and the row that would change your budget is row three.

The Study That Says Referring Domains Win

SE Ranking's November 24, 2025 analysis of 129,000 domains and 216,524 pages found referring domains to be the strongest single predictor of ChatGPT citations, which is the direct opposite of what the mentions slide asserts.

Yulia Deda built it across 20 niches with a target variable derived from 100,000 parsed prompts, modelled with XGBoost regression and interpreted through SHAP feature attribution. That is not a rank correlation. It is a gradient boosted model asked how much each feature contributes to predicted citation count, which is a genuinely different question from whether two columns move together.

And the shape of what it found is the part worth sitting with. Sites with up to 2,500 referring domains averaged 1.6 to 1.8 citations. Citations nearly double at the 32,000 mark, moving from 2.9 to 5.6. Sites above 350,000 referring domains averaged 8.4. SE Ranking also reports that domains above 32,000 referring domains are roughly 3.5 times more likely to be cited than those with up to 200.

Up to 2,500 referring domains: 1.6 to 1.8 citations21%
Just below the 32,000 mark: 2.9 citations35%
At and above 32,000 referring domains: 5.6 citations67%
Above 350,000 referring domains: 8.4 citations100%

Average ChatGPT citations by referring domain band, as published in SE Ranking's November 24, 2025 study of 129,000 domains. Bar length is indexed to the 8.4 reading at the top band, so it is not itself a percentage.

Notice that this is not a line. It is a staircase, and it is flat for a long way before the first step. Hold that thought, because it is the thing that reconciles the two camps and then quietly disappoints both of them.

The same study is not even anti mention, which is the other reason it deserves better than the silence it gets. It found Quora and Reddit presence mattering on a similar curve: up to 33 mentions produced about 1.7 citations, while 6.6 million mentions produced 7.0, and domains with mentions in the millions had roughly four times the citation chance of those with minimal activity. Domain Trust showed the same escalation, with high trust domains above 90 earning almost four times the citations of those below 43. It also found llms.txt negligible, to the point that including the file reduced model accuracy and removing it improved predictions. A study that finds links, communities and trust all mattering is not a rebuttal of the mentions thesis. It is a wider frame that happens to rank links first.

Spearman rank correlation on brand level aggregates and gradient boosted feature importance on page level citation counts answer different questions, so both studies can be correct about links without either one being wrong about anything.

Ahrefs asked whether, across 75,000 brands, the rank order of mention counts tracks the rank order of AI visibility. SE Ranking asked how much each feature contributes inside a model predicting citation counts. The first is monotonic and linear in ranks. It is structurally incapable of seeing a staircase. A variable that does almost nothing until 32,000 and then doubles your citations will score low on Spearman and high on SHAP, every time, in any dataset. That is not a contradiction between two findings. That is two instruments working exactly as designed on the same underlying shape.

The second problem is collinearity, and Ahrefs' own table is the cleanest demonstration of it anyone has published. Branded search volume 0.392. Domain Rating 0.326. Referring domains 0.295. Branded traffic 0.274. Backlinks 0.218. Those are not five independent findings about five levers. That is one finding, reported five times at descending strength, and the thing all five are measuring is how large and well known the company already is. The ordering tells you which proxy for company size is cleanest. It does not tell you which lever is strongest, and rank correlations on collinear proxies cannot be made to tell you that no matter how well presented the chart is.

Nobody running these studies claims otherwise. Both Ahrefs pieces say correlation is not causation, in the text, where you can read it. The slide just omits the sentence and lets you draw the conclusion yourself, which is a more effective way to make a claim than making it. We have argued before that links behave like a gate rather than a dial, and the staircase in the SE Ranking data is what a gate looks like when somebody finally measures it with a tool that can see steps.

The Threshold Is A Company Size, Not A Budget Line

SE Ranking's citation gains appear at 32,000 and 350,000 referring domains, link counts that no retainer in the market reaches, which means the study describes what large domains already have rather than a purchase a mid market brand can make.

Do the arithmetic once and the conversation changes. Moving from the top of the flat band at 2,500 referring domains to the first step at 32,000 means acquiring roughly 29,500 new referring domains. At any honest blended cost for earned referring domains, that is not a line item. That is an acquisition, or a decade, or both. No agency on earth is selling it, including us.

Which means that for the overwhelming majority of companies reading that study, the correct reading is not the headline. It is that this variable is effectively pinned. You are operating in the flat part of the curve where SE Ranking measured 1.6 to 1.8 citations, and within that flat part, link volume is not the thing that moves you.

So we land close to the mentions camp, and we get there by a different road, and the road matters more than the destination. Not because links fail to work. Because for you, specifically, links are already near their practical ceiling and mentions are not. That is an argument from your position on a curve, not an argument about mechanism. The distinction is not academic: a mechanism argument tells you to stop buying links, while a position argument tells you to stop buying links as a citation lever and keep buying them for the thing they demonstrably still do.

THE UNCOMFORTABLE VERSIONIf the honest finding is that citation gains need 32,000 referring domains, then for most of the market the entire brand mentions vs backlinks debate is being conducted about a variable nobody in it can move. The useful question is not which signal is stronger in a study of enterprises. It is which signal is still elastic at your size, and mentions are, which is why the digital PR process that earns them is where the marginal dollar goes.

Seer Interactive's study of 10,000 GPT-4o answers found that ranking on page one of Google correlates around 0.65 with being named in an LLM answer, which is the indirect path every brand level correlation study is built to miss.

Christina Blake and Nick Haigler published it on January 7, 2025. They gathered more than 300,000 finance and software keywords, ran them through Google and Bing, extracted nearly 600,000 People Also Ask questions, narrowed those to 10,000, and put them through the GPT-4o API. Google page one ranking came back around 0.65 against brand mentions in the answers. Bing was weaker at roughly 0.5 to 0.6. Backlinks, measured directly against mentions, were weak or even neutral. The correlation with rankings got stronger once they filtered forums, social sites and aggregators out of the result set.

Now put those two Seer findings side by side, because together they say something neither says alone. Backlinks are weak or neutral against LLM mentions directly. Page one ranking is about 0.65 against LLM mentions. Links remain one of the better established inputs to page one ranking. The chain is links, then ranking, then mention.

KEY TAKEAWAYA study that regresses AI mentions on backlinks while ignoring ranking position is measuring the leftover direct effect after the main path has already been absorbed by a variable the model never included. The result comes back near zero. It comes back near zero even in a world where links are precisely why you got mentioned.

That is why we keep funding links and keep refusing to sell them as a citation product. The outcome you are buying is ranking, with AI citation as a downstream consequence you should expect and should not promise. Our link building services are scoped against rankings and qualified referral traffic for exactly this reason, and a vendor who quotes you a citation lift per link is quoting you a number that does not exist in any of the four studies above.

There is a timing problem on top of the measurement problem, and it is live this week. The November 2026 core update began rolling out on November 10 and may run until around November 24, and Google's own guidance is to wait for a rollout to finish before reading Search Console. Anyone running a links against mentions test right now is reading a moving baseline through a holiday traffic period. Whatever that test shows you in early December is an artifact. Run it in January.

1Find your position on the staircase before you argue about the signalPull your referring domain count today. If it is under about 5,000 you are in the flat band where SE Ranking measured 1.6 to 1.8 citations, and link volume is not your citation lever regardless of which study is right. That single number settles the budget argument faster than any correlation chart will.
2Split link reporting from citation reporting, permanentlyReport links against ranking and qualified organic sessions. Report mentions and citations against share of answer on the surfaces you care about. Keeping them on one dashboard invites the causal claim that none of the published research supports, and it is the reporting choice that makes vendors oversell.
3Read the sampling frame of every study before you quote its numberBoth Ahrefs studies sampled only domains above DR 40 with a keyword above 800 monthly searches. If your brand sits outside that frame, the 0.664 and the 0.218 are both describing someone else's company, and treating them as your coefficients is a category error with a budget attached.
4Prefer the elastic signal at your sizeAt 300 referring domains and 40 brand mentions a quarter, mentions are the variable you can plausibly triple this year and links are not. Fund the one that moves. Revisit the ratio annually rather than quarterly, because both signals compound on a timescale longer than a reporting cycle.
5Stop paying a premium for llms.txt while you are hereSE Ranking found the file negligible to the point that including it as a feature reduced model accuracy and removing it improved predictions. Ship it if it costs an afternoon. Do not let it appear on a proposal as a line item with a price next to it.
So do backlinks matter for AI search or not?Yes, mostly indirectly. Measured straight against AI mentions they test weak. Measured through page one ranking, which Seer put at about 0.65 against mentions, they are one of the inputs. Fund them against rankings, not against citations.
Where does the 0.664 versus 0.218 figure come from?Ahrefs' May 26, 2025 study of 75,000 brands by Louise Linehan, using Spearman correlation against AI Overview brand visibility. Branded web mentions scored 0.664 and number of backlinks 0.218. The study itself states that correlation is not causation.
Which study says referring domains win?SE Ranking's November 24, 2025 analysis by Yulia Deda, covering 129,000 domains and 216,524 pages across 20 niches, modelled with XGBoost and SHAP. It ranked referring domains the strongest single predictor of ChatGPT citations.
Why do the two studies disagree?They used different statistics on different units. Rank correlation cannot detect a step function, and SE Ranking's citation curve is a step function with jumps at 32,000 and 350,000 referring domains. Both results are correct about the same shape.
How many referring domains do I need to get cited more?On SE Ranking's curve, meaningfully more citations start around 32,000 referring domains and keep climbing to 8.4 at above 350,000. That is not purchasable. Below roughly 2,500 the curve is flat at 1.6 to 1.8 citations.
Should we move link budget into brand mentions?Move the portion you were funding against AI citations. Keep the portion funding rankings. Mentions are the elastic signal at most company sizes, which is a reason to fund them without being a reason to defund links.
Does llms.txt help with citations?No evidence that it does. SE Ranking found its impact negligible, and reported that including the file actually reduced the accuracy of its citation model. Publish it cheaply if you want. Do not pay for it.
Can we test this ourselves right now?Not cleanly this month. The November 2026 core update is still rolling out and the holiday period distorts the baseline, so start a controlled test in January and hold both the link and mention variables for at least two quarters.
DO THIS NEXTToday, pull your referring domain count and your quarterly branded mention count and write them on the same line, because that pair tells you which signal is still elastic for you and no published study can tell you that. This week, split your reporting so links are measured against ranking and sessions while mentions are measured against share of answer. Then push any links against citations test into January, when the core update has finished and the baseline stops moving. If you want the per surface version of this for your own category, start with how citations get earned inside ChatGPT and the B2B SaaS playbook, which is where the mention curve is steepest.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

TT
Tyler TruffiMANAGING PARTNER, SOMETHING INC.

Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.

Keep reading

All insights →
LINK BUILDINGYahoo Backs AI Licensing. Its Own CEO Just Said It Won't Work.Yahoo CEO Jim Lanzone told HumanX in Amsterdam that AI micropayments "will never be enough, even for someone like Reddit," and argued search ads should pay for the open web instead. Yahoo is also a founding backer of the RSL Collective, the licensing standard built specifically to sell AI companies on paying per use. Both things are true at once, and the gap between them is the clearest read yet on whether AI content licensing deals are a real revenue line or a negotiating chip. Here is what each side of Yahoo's own position is actually selling, and what to do if you are not one of the 91 publishers currently inside a deal.JOSH BERNSTEIN · 11 MINLINK BUILDINGThe AI Citation Process Most Digital PR Teams Don't HaveA 2026 survey of 150+ digital PR professionals found 83% say AI has made their work more important and only about 40% have a repeatable process for earning AI citations. The five-step process that closes that gap, and how to tell it is actually working.TYLER TRUFFI · 11 MINLINK BUILDINGEarned Media Drives 84% Of AI CitationsMuck Rack and Generative Pulse released the fourth edition of their AI citation study on October 28, 2026, covering more than 40 million AI responses and 207 million citations collected across 2026. More than 84% of the citations came from earned and third-party sources. Paid and advertorial content accounted for 0.3%, unchanged since May. Journalism fell from 27% to 22%, the largest category shift the series has recorded, while owned media rose from 13.7% to 15.2% and third-party corporate content rose from 24% to 25.1%. Only 212 domains were cited by all four major providers. The finding that should change how teams work is quieter: when the same prompt was run twice in one day, ChatGPT repeated 41% of its cited domains and Gemini 45%, against 94% for Perplexity. Earned coverage is the asset, access is the gate, and a single-run visibility dashboard is measuring weather, not climate.JOSH BERNSTEIN · 12 MIN
Free consultation

Let us be the last SEO agency you ever work with

A 30 minute call and a free audit of your SEO and GEO position. You keep the findings either way.