Updated on 19 October 2026, by João Leite, Founder of murmur.

Backlinks still matter for appearing in AI, but in the Ahrefs study brand mentions on other sites track more closely with the answers. A backlink is a link from another site to yours. In an Ahrefs study of 75 thousand brands, from December 2025, backlinks had a correlation of 0.19 to 0.24 with the brand appearing. Mention of the brand name on the web was between 0.66 and 0.71, in ChatGPT and in two Google AI models. Correlation measures whether two numbers rise together, from 0 to 1, and does not prove that one causes the other. As of October 2026, we found no published test that adds links to a site and measures what changes in the answers.

  • The number of backlinks tracks weakly with the brand in the answers. In Ahrefs, the correlation was between 0.19 and 0.24, and for mention of the brand name, between 0.66 and 0.71.
  • The link with the brand name in the clickable text tracks more closely. In the same calculation, it was between 0.51 and 0.63.
  • For a site to become a ChatGPT source, the number of sites linking weighed more in the SE Ranking model. Those with more than 32 thousand sites linking had 3.5 times more chance of citation than those with up to 200.
  • Backlinks can reach the answers through search, and nobody has measured how much. In March 2026, 37.1% of the URLs cited in AI Overview were among Google’s top 10. In Claude, it was 79.2% among Brave Search’s top 10, in June 2026.
  • The effect of gaining a backlink on the answers still has no published test. Google’s guide to AI features in search does not mention backlinks.

murmur, a GEO agency in Brazil that puts brands into the answers of ChatGPT, Claude, Google AIO and other AI agents, works in three parts: studying the client’s sector, writing the pages the buyer looks for and monitoring the AI models every week.

What has been measured about backlinks and AI is correlation, not the effect of gaining a link. The studies count two different things, and the difference changes the answer. Ahrefs counted the brand named in the answer text. SE Ranking counted the site used as the source of the ChatGPT answer. In both, the link is one signal among several, and neither added links to a site to see what happened.

Two things about backlinks and AI remain to be measured. The first is the effect of gaining links, and we found no published study, with the data, that compares the answers before and after. The second is the weight of backlinks in Claude and Perplexity, which the link calculations we found left out. Being cited as a source and being recommended are different outcomes, explained in why AI recommends some companies and not others.

Among the backlink measures, the link with the brand name in the clickable text is the one that tracks most closely with the brand in the answers. Ahrefs, which sells SEO and backlink tools, crossed each measure with the brand appearing in three AI models. The three are ChatGPT, AI Overview and AI Mode. AI Overview, or Google AIO, is the text Google’s AI writes at the top of some searches, above the ordinary links. AI Mode is Google’s search mode in which the person talks with the AI.

In the study’s chart, the number of backlinks came in at 0.191 in ChatGPT, 0.244 in AI Mode and 0.218 in AI Overview. The number of different sites that link, the referring domains, was a little higher, from 0.27 to 0.31. The link score of each page, the URL rating, was between 0.17 and 0.20. The Domain Rating (DR), a score from 0 to 100 that Ahrefs gives to the links of an entire site, was between 0.27 and 0.33.

The link with the brand name in the clickable text, the brand anchor, was between 0.51 and 0.63. Ahrefs reads that link as two things at once: someone cited the brand by name and chose to link to it. On the link measures, Ahrefs sums up:

Once again, we found very weak correlations between link metrics (“number of backlinks” and “URL rating”) and brand mentions across all AI systems.

Mention of the brand name on web pages was between 0.66 and 0.71, above any link measure. The role of that mention, alongside the text on your own site, is in text on your site and mentions on other sites do different jobs. The Ahrefs sample only has sites with DR above 40, and sites with weak links were left out. Ahrefs itself warns that correlation is not causation.

The weight of backlinks changes from one AI to another, and we only found published link calculations for ChatGPT and the Google AI models.

AIin favor of backlinks, with the dateagainst, or with a caveat, with the date
ChatGPTNovember 2025, updated in March 2026: the number of sites that link was the highest-weight factor for the site being cited as a source, and more than 32 thousand gave 3.5 times more chance than up to 200 (SE Ranking)December 2025, updated in August 2026: the number of backlinks had 0.191 of correlation with the brand appearing, the weakest of the three AI models studied (Ahrefs)
ClaudeJune 2026: 79.2% of the pages cited were among Brave Search’s top 10; that is position in search, not link (Profound, with MERJ)no study found on backlinks in Claude; Anthropic did not confirm the connection with Brave (MERJ)
Google AIODecember 2025: DR had 0.326 of correlation, the highest of the three AI models (Ahrefs)March 2026: 37.1% of the URLs cited were among the top 10 links in search, against about 76% in July 2025 (Ahrefs)
Perplexitydata from July 2025: 28.6% of the links cited were among Google’s top 10, the highest share among the AI models compared; that is position in search, not link (Ahrefs)no study found on backlinks in Perplexity

For ChatGPT, the two link studies seem to disagree, because they count different things. SE Ranking, which sells an SEO and backlink tool, cross-referenced 129 thousand domains with ChatGPT citations in 100 thousand questions. The study is from November 2025, updated in March 2026. In its model, the number of sites that link was the highest-weight factor for the domain being cited. Those with more than 32 thousand sites linking had 3.5 times more chance of citation than those with up to 200. The calculation is for sites used as a source, and shows association, not the effect of gaining a link.

At Ahrefs, ChatGPT was the AI model in which the classic authority signals tracked the brand the least. The number of backlinks was 0.191, and DR was 0.266. Ahrefs writes that ChatGPT may be the best entry point for a brand with little search, few backlinks and few mentions. The two calculations fit together: in ChatGPT, the link tracks with the site becoming a source and tracks little with the brand being named.

In Google AIO, DR had the highest correlation of the three AI models in Ahrefs, 0.326, and the number of backlinks was 0.218. Ahrefs itself calls DR an intermediate factor, much weaker than brand mentions. In March 2026, Ahrefs found 37.1% of the URLs cited in AI Overview among the top 10 links in search. In the July 2025 study, the share was about 76%, and it fell by half. Ahrefs hypothesizes that, with Gemini 3, searches derived from the original question began to carry more weight in the choice of sources.

In Claude, we found no study that measures the weight of backlinks, and what exists is position in Brave Search. Brave Search is a search engine with its own index. Profound, which sells brand monitoring across AI models, found 79.2% of the pages cited by Claude among Brave’s top 10. Among Google’s top 10, it was 34%, in about 35 thousand citations from June 2026. The study came out with MERJ, which, like Profound, sells AI search services to brands, and which writes that Anthropic did not confirm Claude’s connection with Brave.

In Perplexity, we also found no backlink study. In an Ahrefs study with data from July 2025, 28.6% of the links cited by Perplexity were among Google’s top 10. That was 15 thousand questions, and in the other AI models compared, among them ChatGPT, the share was close to 8%. Ahrefs warns that the numbers may have changed since collection.

A backlink can reach the AI answer through search: the link helps the page rise in search engines, and AI models look for sources there. In MERJ’s reading, Brave builds part of its index from what the bigger search engines already put at the top, and from there it reaches Claude. If that is the path, the link counts more where the AI depends more on the top of search, as in Claude and Perplexity. In Google AIO, the top counts for less than in 2025, and more than 60% of citations come from outside it.

Nobody has measured how much of that path comes from the backlink. Position calculations do not separate the link from the other reasons that put a page at the top, such as the text and the technical side. What position in Google changes in ChatGPT is in being at the top of Google helps you appear in ChatGPT. Claude’s path through Brave is in how to get your company cited by AI models.

Google’s guide to the AI features of search, updated in July 2026, gives no instruction about backlinks. It explains that the answer comes from the systems that rank pages. About the technique that looks for pages in the index to write the answer, Google writes:

Ela se baseia nos nossos principais sistemas de classificação para recuperar páginas da web relevantes e atualizadas do nosso índice da Pesquisa.

Original quote in Portuguese, preserved as published. In English: “It relies on our core ranking systems to retrieve relevant, up-to-date web pages from our Search index.”

Google’s guide adds that the AI features of search also depend on the systems that block spam.

Buying backlinks and exchanging links at scale put the site at risk in Google, and no study we found shows a gain from that in AI models. Google’s spam policy treats as spam a link created with the main purpose of manipulating position in search. That includes buying or selling links for ranking purposes, exchanging links excessively, creating links with an automated program and spreading low-quality links in directories. Under the same policy, a site that breaks the rules can drop in search or disappear from it. When search goes, so does the path that takes some pages to Google AIO.

A paid post with a link follows the rules when the link is marked as sponsored. Where the line is, with what CONAR asks of paid articles, is in manipulating what ChatGPT recommends. The same policy treats as spam the many pages generated to manipulate search. The line is between one page for each genuinely different buyer question, with the complete answer, and one page for each way of writing the same question. What that changes for those who publish content is in GEO and content marketing are not the same thing.

Backlink in an article, list and review: what counts is the text the AI reads

In a backlink inside an article, list or review, the AI reads the text of the page, and the link is only a part of it. What the AI can repeat is what the page writes about the company: the name, what it does and how it compares. The weight of the press in each AI is in press relations and ChatGPT. The weight of reviews is in reviews and Reclame Aqui, and the weight of other sites, in other sites talking about your company.

To check whether a backlink changed the answers, compare what the AI models answer to the same questions before and after the link. Note the day the link was published and run one round before it and another a few weeks after. Open the sources of each answer and check three things: the page that links, your page and the company name in the text. Also check, on the same dates, the position of your page in Google and in Brave Search.

Brand presence is the number that sums up this comparison. Presence is how many answers the brand appears in, divided by the total number of answers. Each question appears in the round more than once, always in a new conversation, since the same AI answers differently on each attempt. If the link came in the same week as a new page or an article, the calculation does not separate the effect of each one. The spreadsheet for this calculation, and how to repeat the questions over time, are in the guide on how to measure brand presence in ChatGPT.

murmur works alongside SEO and checks every week what changed in the AI answers

murmur works on the AI answers, not on position in Google, and its work adds to the SEO the company already has. Each page written by murmur goes to a subdomain that carries the company’s brand. Everything murmur writes belongs to the company, and keeps belonging to it after the contract ends. One specialist follows the company from the diagnostic onward and studies its sector, and the same person stays with the company throughout.

At murmur, the same questions from the client’s market go to ChatGPT, Claude, Perplexity and Google AIO week after week. The client sees on a dashboard, included in the service, its own presence and that of each competitor. With weekly measurement, the change in the answers shows up in the week it happens. Then it can be cross-referenced with what the company did that week: a new backlink, an article, a published page. Anyone who invests in backlinks without redoing the questions ends up not knowing whether the investment reached the answers.

Frequently asked questions

From what was measured, backlinks are weakly associated with the brand being named by ChatGPT, and more strongly with the site becoming a source. In Ahrefs, the number of backlinks had a correlation of 0.191 with the brand appearing in ChatGPT, the weakest of the three AI models. In SE Ranking, the number of sites that link was the factor with the greatest weight for the domain being cited by ChatGPT. Both studies measure association, and neither tested what happens when a site gains links.

Ahrefs did not measure the link without the brand name separately, and the link with the name was more associated with the answers. Total backlinks were between 0.19 and 0.24, and the link with the brand name in the clickable text, between 0.51 and 0.63. For the mention with no link at all, the answer is in does a mention without a link help you appear in AI.

The backlink helps the page rise in Google, but few ChatGPT sources are at the top of Google for the same question. With data from July 2025, Ahrefs found only about 8% of the links cited by ChatGPT among the top 10. ChatGPT’s steps between the question and the sources it cites are in how ChatGPT search works.

Buying backlinks to rise in search is spam for Google, and we found no study showing a gain from that in AI models. A site that breaks the spam policy can drop in search or disappear from it, and loses the path to Google AIO along with it. What Google, Bing, OpenAI and Anthropic write about manipulating the answer is in manipulating what ChatGPT recommends.

Just for AI, link building, which is the work of getting links from other sites, has little in its favor in the published studies. Ahrefs, which sells a backlink tool, writes that gathering links in quantity is not enough. For it, what counts most is the brand being cited across a broad range of sites. The link that is already part of SEO still counts for search. How much of the budget goes to each area is the topic of GEO and SEO: which to invest in first.

Where to start

Before investing in backlinks with AI models in mind, find out which sites they already use as sources for your buyer’s questions. Those are the sites where a link to or a mention of your company enters the text the AI reads. murmur’s free diagnostic gives a first picture of what AI models answer in your market, with no sign-up and in under two minutes. Each sector study by murmur records the AI answers to 25 purchase questions in a niche sector, in a single round. In most of them, the sites cited as sources also appear. The murmur case studies are published with each client’s name.

Read also

References

  • Ahrefs, Louise Linehan, with Xibeijia Guan, “Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)”, 12 December 2025, updated on 12 August 2026. That is 75 thousand brands, from domains with DR above 40 and with at least 800 searches per month for their most searched keyword; the mentions in the answers come from Brand Radar, an Ahrefs tool, and the correlation is Spearman’s. The per-AI numbers for number of backlinks (0.191 in ChatGPT, 0.244 in AI Mode and 0.218 in AI Overviews), URL rating (0.172, 0.204 and 0.18), referring domains (0.271, 0.307 and 0.295) and DR (0.266, 0.285 and 0.326) are in the page’s charts; brand anchors (0.511, 0.628 and 0.527) and web mention (0.66 to 0.71) are also in the text. Ahrefs sells SEO, backlink and AI monitoring tools, and writes that correlation is not causation. https://ahrefs.com/blog/ai-brand-visibility-correlations/ (accessed on 8 October 2026).
  • SE Ranking, Yulia Deda, “How to optimize for ChatGPT: LLMs.txt doesn’t matter but brand mentions on Quora and Reddit do”, 24 November 2025, updated on 27 March 2026. That is 129 thousand domains and 216,524 pages from 20 niches, with ChatGPT citations in 100 thousand questions, analyzed with a regression model (XGBoost) and by reading the weight of each factor (SHAP); the page does not state the country or the dates of the questions. Sites with up to 2,500 referring domains had 1.6 to 1.8 citations on average, and those with more than 350 thousand had 8.4. SE Ranking sells tools for SEO, backlink analysis and ChatGPT monitoring. https://seranking.com/blog/ranking-factors-for-chatgpt/ (accessed on 8 October 2026).
  • Ahrefs, Louise Linehan, “Update: 38% of AI Overview Citations Pull From The Top 10”, 2 March 2026, modified on 31 May 2026. That is 863 thousand result pages and 4 million URLs cited in AI Overview, from Brand Radar. With common links only, 37.10% were among the top 10, 26.20% from 11th to 100th and 36.70% outside the top 100; counting ads and highlights as blocks, 37.9% were in the top 10 blocks. The previous study, from July 2025, gave about 76%. https://ahrefs.com/blog/ai-overview-citations-top-10 (accessed on 8 October 2026).
  • Ahrefs, Louise Linehan, with analysis by Xibeijia Guan, “Only 12% of AI Cited URLs Rank in Google’s Top 10 for the Original Prompt”, 11 August 2025, updated on 31 May 2026. That is 15 thousand long-tail questions, with data collected in early July 2025; Perplexity had 28.6% of cited links among Google’s top 10, and the other AI models, close to 8%. https://ahrefs.com/blog/ai-search-overlap/ (accessed on 8 October 2026).
  • MERJ, Ryan Siddle, with Josh Blyskal, of Profound, “How Brave Search discovers new pages: A deep dive into the Web Discovery Project”, 9 July 2026, modified on 16 August 2026. The 79.2% figure is Profound’s, with about 35 thousand Claude citations across 400 questions, in June 2026, and 34% among Google’s top 10; MERJ writes that Claude’s connection to Brave is well documented as of July 2026, but was not confirmed by Anthropic. MERJ and Profound sell AI search services to brands. https://www.merj.com/blog/how-brave-search-discovers-new-pages (accessed on 8 October 2026).
  • Google, “Guia do Google para otimizar recursos de IA generativa na Pesquisa Google”, Google Search Central, Portuguese version last updated on 14 July 2026. The page does not deal with backlinks. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide?hl=pt-br (accessed on 8 October 2026).
  • Google Search Central, “Políticas de spam para a Pesquisa Google na Web”, last updated on 16 September 2026, link spam section. https://developers.google.com/search/docs/essentials/spam-policies?hl=pt-br#link-spam (accessed on 8 October 2026).
  • murmur, free diagnostic: https://murmur.marketing/en/diagnostic/; sector studies, with 25 purchase questions per niche sector: https://murmur.marketing/estudos/; case studies published with client authorization: https://murmur.marketing/en/cases/ (accessed on 8 October 2026).