Do backlinks matter for AI visibility? Links, brand mentions and what the evidence says
Yes, but mostly indirectly. Links move the rankings that AI assistants retrieve from, and a branded link is also a mention. Among established brands, mentions predict AI visibility better than link counts do, and nearly all of the evidence is correlation from tool vendors.
In short
- An assistant answers from two places, what the model absorbed in training and what it retrieves from a search index at answer time, and links act almost entirely on the second.
- In Ahrefs' December 2025 study of 75,000 brands, branded web mentions correlated 0.66 to 0.71 with AI visibility and Domain Rating 0.27 to 0.33, in a sample limited to sites above Domain Rating 40.
- Semrush's October 2025 study of 1,000 randomly chosen domains found its Authority Score correlated 0.65 with AI mentions, which fits links working as an entry threshold more than as a dial.
- Google's guide to generative AI features, read on 4 October 2026, says those features are rooted in its core Search ranking and quality systems, so whatever links do for rankings carries over.
- Every study on this question is a correlation, nearly all come from companies that sell tools, and no public experiment has shown what adding links or mentions does to one brand.
Backlinks still matter for visibility in AI assistants, but mostly at one remove. An assistant draws on two things when it answers: what the model absorbed during training, and what it retrieves from a search index while you wait. Links do most of their work on the second. They help a page get discovered, indexed and ranked, and ranked pages are what retrieval returns. On the first, the model’s memory, links as such do very little, because a model learns from the words on a page and has no use for the hyperlink. What it picks up is how often a brand is named and what is said next to the name.
That is why the published studies look contradictory and are less so than they seem. Counts of links and authority scores are a poor guide to which established brands get named, while mentions across the web are a much better one. Even so, a link with the brand as its anchor is itself a mention, and the answer changes from one assistant to the next because each searches a different index. The limits should be stated just as plainly. Every study on this subject is a correlation, nearly all were published by companies that sell tools in the field, and nobody has run a public experiment that adds links or mentions to one brand and measures what happens. What follows is the best reading of imperfect evidence, with the imperfections marked.
Why the argument exists
Three camps are arguing, and each has a real observation behind it. The first says links are still required. Its strongest form goes like this: assistants that search the web retrieve from search indexes, search indexes rank pages partly on links, and so a site with a weak link profile will seldom have its pages retrieved, whatever else it does. This camp can point to Google’s own documentation, to Microsoft’s, and to studies in which sites with more referring domains are cited more often. It tends to be argued by link builders, outreach software companies and SEO agencies, who would lose a product line if it were false. That interest does not make the argument wrong, though it does explain why the argument is usually stated without its caveats.
The second camp says links have stopped mattering for AI. Its strongest form rests on overlap data. When Ahrefs compared the pages that assistants cite with the pages that rank in Google’s top ten, the overlap for ChatGPT was around one in ten, and when Seer Interactive tested 10,000 questions in January 2025 its authors reported that they had expected backlinks to play a big role and found their impact “weak or even neutral”. If the cited pages are mostly ones that do not rank, and link counts do not predict who is named, the reasoning runs, then the whole link economy is aimed at the wrong target. This view is most often heard from people selling a new discipline under a new acronym, and from content and PR practitioners who never liked link building in the first place.
The third camp says only mentions matter. Rand Fishkin of SparkToro put the core of it in October 2024, writing that “the currency of large language models is not links” but mentions, by which he meant words that appear frequently near other words across the training data. A model is trained on text, the argument goes, so it learns which brand names appear near which category words, and a hyperlink adds nothing to that. The camp’s best evidence is Ahrefs’ study of 75,000 brands, in which mentions tracked AI visibility far more closely than any link measure. Digital PR agencies, brand monitoring vendors and PR software companies argue it most loudly, for reasons as obvious as the link builders’. Each camp has seized one part of how an assistant works and treated it as the whole. The disagreement mostly dissolves once the two parts are laid side by side.
How an AI assistant produces an answer
Think of a well-read consultant who is asked for a recommendation. Often they answer from what they already know. Sometimes they say “give me a minute”, look a few things up, and answer from what they have just read. An assistant does both, and the two routes are fed by different things. The first route is the model’s training: a very large body of text, collected up to a cut-off date, from which the model learned patterns of language, including which names tend to appear in which contexts. The second route is live retrieval. The assistant runs searches against an index, reads the pages or passages that come back, and writes an answer with sources attached. The analogy stops being useful at that point, because a consultant remembers where they read something and a model generally does not.
The retrieval route has more steps than most people assume, and each step is a place where a page can be included or dropped. The assistant first decides whether to search at all, and many answers come from memory alone with no citations. If it searches, it rewrites the prompt into several searches of its own, a step known as query fan-out. Research by Seer Interactive and Nectiv, cited by Ahrefs in March 2026, found that 59% of prompts trigger 5 to 11 of these hidden searches, and the same Ahrefs article says over 95% of them have no recurring human search volume. The results are then merged, passages are picked and the answer is written. Google’s guide to generative AI features, read on 4 October 2026, describes its own version plainly: retrieval works “by relying on our core Search ranking systems” to fetch current pages from the Search index.
Two different outcomes come out of this, and the camps often argue past each other because they are measuring different ones. A mention is the brand named in the text of the answer. It can come from memory, or from what the retrieved pages say, and those pages are usually somebody else’s. A citation is a URL shown as a source, which can only come from retrieval. Semrush’s 2026 AI Visibility Index found that on Gemini the overlap between mentioned brands and cited domains was as low as 30%, which is vendor data, but it shows how far the two can drift apart. How AI search chooses its sources walks through the pipeline in more detail.
Now set the three camps against that picture. Links act on retrieval, since retrieval runs against an index that ranks pages, and they act there in the ordinary way: by helping pages get crawled, trusted and ranked. Mentions act on both routes. They are what the model saw in training, and they are what the retrieved third-party pages contain. The “links are finished” camp is looking at the fan-out step, where the hidden searches are so many and so unlike the head keyword that the page ranking first for the obvious query is often not the page cited. So the link camp is describing the index, the mentions camp is describing the text, and the sceptics are describing the gap between the query a person types and the queries a machine runs.
Which search index feeds each assistant
The retrieval route depends on an index, and the assistants do not share one. This matters more than any other single fact for the links question, because whether links can influence retrieval on a given assistant depends on whether the index behind it uses links to rank. The table sets out what each company documents and what is only reported. Which index each assistant uses has the full account, including the crawler names.
| Assistant | Where it retrieves from | How sure | What that means for links |
|---|---|---|---|
| Google AI Overviews and AI Mode | Google’s index, through its core ranking systems | High. Google documentation | Links count as far as they count in Google rankings |
| Gemini | Google Search, through a grounding feature | High for the API, unverified for the consumer app | Probably the same as above |
| ChatGPT search | Third-party search providers plus OpenAI’s own crawler. Bing was the launch partner | High for the documented part. A report that Google results were also used is unconfirmed | Indirect, through whichever engine supplies the results |
| Microsoft Copilot | Bing’s index | Medium to high. Microsoft reports Copilot citations inside Bing Webmaster Tools | Links count as far as they count in Bing |
| Perplexity | Its own index, built by PerplexityBot | High. Perplexity documentation | Unknown. Its ranking method is not published |
| Claude | Brave Search, reportedly | Medium. Neither company has confirmed it | Unknown. Brave runs its own index |
Google’s case is the clearest. Its documentation on AI features says: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” A page has to be indexed and eligible to show with a snippet, and the newer guide adds that the AI features are built on the same core ranking and quality systems as ordinary Search. An exhibit from the US antitrust case, reported in May 2025, describes anchors feeding Google’s relevance score and PageRank feeding its quality score, as described in the lesson on the court documents and the leak. Put those two together and the conclusion is hard to avoid: on Google’s AI surfaces, links matter to the extent that they matter in Google Search. That extent is smaller than it was a decade ago and well above nothing.
Microsoft says much the same in fewer words. Krishna Madhavan of the Bing team wrote in October 2025: “Crawlability, metadata, internal linking, and backlinks remain essential for ensuring your content is discoverable.” Note the word chosen. Microsoft frames backlinks as a matter of being found, which is a more modest claim than saying they decide which source an answer quotes. How Bing weighs links is covered elsewhere on this site, and since Bing was the launch partner for ChatGPT search, Bing’s treatment of links reaches further than Copilot alone.
The other three are murkier. OpenAI’s crawler documentation says OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features”, and says nothing at all about how results are ordered. Perplexity documents its crawler and not its ranking. For Claude, TechCrunch reported in March 2025 that Brave Search had appeared on Anthropic’s list of data subprocessors and that the programmer Simon Willison had found a search parameter named after Brave, while neither company confirmed the arrangement. So for half the assistants in the table, nobody outside the company can say how much links weigh in the index that feeds them. Anyone who states it confidently for ChatGPT, Perplexity or Claude is guessing, whichever camp they belong to.
The evidence that links matter less than people assume
The case against links rests on two kinds of study. The first kind correlates link measures with how often a brand is named. Ahrefs’ May 2025 study of 75,000 brands, which covered Google’s AI Overviews only, found that referring domains correlated with AI mentions at 0.295 and the raw number of backlinks at 0.218, against 0.664 for branded web mentions. Its December 2025 update added ChatGPT and AI Mode and put Domain Rating at 0.266 to 0.326 across the three. Seer Interactive’s January 2025 test reached the same place by a different road, with backlinks weak or neutral as a predictor of being named. A correlation near 0.2 or 0.3 means the measure moves with the outcome only loosely, so a brand’s link count tells you little about its chances.
The second kind compares what assistants cite with what Google ranks. In Ahrefs’ August 2025 study of 15,000 long-tail prompts, the share of cited URLs found in Google’s top ten was 8% for ChatGPT, 8% for Gemini and 8.6% for Copilot. Semrush reported in July 2025 that about 90% of the time ChatGPT cited pages ranking at position 21 or lower for related queries. Even on Google’s own surface the tie is loosening: Ahrefs found in March 2026 that 37.9% of URLs cited in AI Overviews ranked in the top ten for the same query, down from about 76% in July 2025. If link building exists to push one page up the results for one keyword, these figures say that prize is worth less inside an AI answer than it is on a results page.
Both findings are real, and both are routinely stretched past what they show. Start with the correlations. The slogan that mentions beat links “three to one” divides one correlation coefficient by another, which measures nothing about effect. The Ahrefs sample was also limited to sites above Domain Rating 40, and that filter matters more than it looks. When a study only includes sites that already have substantial authority, authority cannot explain much of the difference between them, because they all have it. Height tells you little about who scores most among professional basketball players for the same reason, and nobody concludes from that that height is irrelevant to basketball. The study tells you that among established brands, more links do not separate the visible from the invisible. It cannot tell you what links do for a site that has few.
The overlap studies need the same care. A cited page that does not rank in the top ten for the prompt has not escaped ranking. It has usually ranked for one of the hidden searches the assistant wrote, which is a different query from the one the researcher checked. Ahrefs’ own explanation for the falling AI Overviews figure is heavier fan-out, and its September 2025 study found that while only 10% of ChatGPT’s cited URLs were in Google’s top ten, overlap at the level of the domain was 31.8%. The same sites recur through different pages. There is also a point about authority scores that both sides tend to forget. In Ahrefs’ January 2025 study of one million keywords, Domain Rating correlated with ordinary Google rankings at only 0.131, which is lower than its correlation with AI visibility. Nobody sensible reads that as proof that links are irrelevant to Google. A third-party score is a rough summary of a link profile, and a weak correlation for the summary says little about the thing summarised, as Domain Authority versus Domain Rating explains.
The evidence that links still matter
The strongest evidence on this side is documentary, which puts it in a different class from any correlation. Google says its AI features retrieve through its core ranking systems, and court exhibits show links inside those systems in two places. That is a description of how the machine is built. For Google’s AI Overviews, AI Mode and, by the API documentation at least, Gemini, the question “do links matter for AI visibility” reduces to the older question of whether backlinks still matter for rankings, and the answer there is yes, less than before, with a threshold that varies by market. Google’s Gary Illyes said in April 2024 that Google needs “very few links” to rank pages. Very few is still more than none.
The correlational support comes from the ranking side. Seer’s study, the one that found backlinks weak, also found that ranking on Google’s first page correlated about 0.65 with being named by the model, and Bing rankings 0.5 to 0.6. Seer removed forums, aggregators and social sites from its sample, which flatters those figures, but the direction is clear: the pages that rank are disproportionately the pages a model draws on, and how search engines evaluate links is part of why they rank. On Perplexity the tie is tighter still. Ahrefs’ September 2025 study of short keywords found Perplexity matched Google’s top ten for 65% of its citations. So a team that says links do nothing for AI visibility has to explain why the systems that reward links keep supplying the sources.
Then there are the studies that find authority predicting AI visibility directly, and they deserve a careful reading because they appear to contradict Ahrefs. Semrush published a study in October 2025 of 1,000 randomly selected domains across ChatGPT, Gemini, AI Overviews and Perplexity. Its Authority Score correlated with the number of AI mentions at 0.65 on one measure and 0.57 on another. SE Ranking’s November 2025 analysis of 129,000 domains, as reported by Search Engine Journal, found that sites with up to 2,500 referring domains averaged 1.6 to 1.8 ChatGPT citations while those above 350,000 averaged 8.4, with a jump from 2.9 to 5.6 at around 32,000. The likeliest reason these disagree with Ahrefs is the sample. A random draw of domains includes the small and the obscure, so authority has room to explain differences, whereas a sample that starts at Domain Rating 40 has removed most of that room.
Read together, the three studies describe a threshold. Below some level of authority a site is rarely retrieved or named at all. Above it, more of the same buys little, and other things decide who is chosen. That is the same shape Illyes’ remark implies for ordinary rankings. Semrush’s data has a second detail that cuts against a simple reading in favour of link building: when the outcome was share of voice against competitors, the correlations with Authority Score fell to about 0.23 and 0.36. Being a big site gets you mentioned somewhere. It does much less to decide whether you are picked ahead of a rival for a given question.
One study on this side does not show what its headline claims. Pitchbox, which sells outreach software, reported a correlation of 0.62 between referring domains and position in assistants’ recommendations. The sample was a little over 20 B2B software brands in two categories, tested on prompts such as “best SEO software”. The best-known companies in a category have the most links and are also the ones a model names first, so a correlation was guaranteed, and two dozen data points from two markets cannot carry a general claim. Pitchbox does state that correlation does not imply causation. The finding is compatible with links mattering and equally compatible with fame causing both.
What happens to a site with no links
No public study has followed sites with no links into AI answers, so this part is reasoning and should be read as such. A page has to be discovered before any index can hold it, and links are how crawlers find most pages, which is the sense in which Microsoft calls backlinks essential for discoverability. A linkless site can be submitted through a sitemap and indexed, but being in the index is only the entry condition. To be cited it has to rank for at least one of the hidden searches, in competition with pages that have links. To be named from memory it has to have been written about. A site with no links and no mentions has neither route open. A site with no links and plenty of mentions can be named, with other people’s pages cited as the source, which is a real outcome and a weaker one than being cited yourself.
The evidence for mentions
The mentions camp has the largest sample and the most consistent result. In Ahrefs’ December 2025 data, branded web mentions correlated with AI visibility at 0.664 for ChatGPT, 0.709 for AI Mode and 0.656 for AI Overviews, and mentions on YouTube were the strongest single factor at 0.737 for ChatGPT. The earlier edition gave a sense of scale: the top quarter of brands by web mentions averaged 169 mentions in AI Overviews, against 14 for the next quarter down, and 26% of the brands in the study had none at all. That last figure is worth holding on to. Every brand in the sample had a Domain Rating above 40, so a quarter of sites with respectable link profiles were invisible in AI Overviews. Authority on its own did not get them named.
The mechanism is plausible on both routes. For memory, Ahrefs notes that both Google and OpenAI have trained models on YouTube transcripts, and a brand discussed in thousands of documents is present in what the model learned before any search runs. For retrieval, the pages an assistant fetches are mostly third-party pages, and whichever brands those pages name are the brands available to the answer. Peec AI’s 2026 analysis found ChatGPT added the word “best” to 24.3% of advice-style questions when the user had not typed it, which sends retrieval straight to comparison lists and reviews. Brand mentions and citations covers the signal in general, and the related idea that models learn which brands sit near which category words is set out in co-citation and co-occurrence. That idea is well reasoned and still a hypothesis, since no public controlled experiment at scale has tested it.
Where the mentions sit matters, because assistants read some places far more than others. Peec AI’s analysis of 30 million US citations in March 2026 put Reddit, YouTube, LinkedIn, Wikipedia, Forbes, G2, Yelp, Facebook, Medium and TechRadar in the overall top ten, with Google’s surfaces leaning towards social platforms and ChatGPT towards editorial sites. The shares are smaller than the charts suggest. In Profound’s analysis of 680 million citations, Wikipedia was 7.8% of ChatGPT citations and Reddit 6.6% of Perplexity’s. They are also unstable: Semrush recorded Reddit’s share of ChatGPT citations falling from 3.8% to 0.5% within August 2026. The separate lessons on Reddit threads, YouTube mentions, third-party “best X” lists and review platforms take each surface in turn. SE Ranking’s November 2025 analysis adds one directional figure on review sites: domains present on several review platforms averaged 4.6 to 6.3 ChatGPT citations against 1.8 for those absent, though big companies are on more platforms, so the usual caution applies.
Linked and unlinked mentions
The camps treat a link and a mention as rivals, and the data keeps refusing the distinction. In the Ahrefs table the second-strongest web factor after mentions is branded anchors, meaning links whose clickable text is the brand name, at 0.511 to 0.628. A branded anchor is a mention that happens to be clickable. Semrush’s October 2025 study found something that points the same way: nofollow links correlated with AI mentions almost exactly as well as followed links, 0.340 against 0.334 on one measure and 0.509 against 0.504 on the other. If links helped AI visibility mainly by passing ranking credit, nofollow links should do visibly worse. That they do not suggests that much of what a link contributes here is the appearance of the brand on a page that gets read, which is what a mention contributes too.
That gives a useful order of value for a single placement. A link in a relevant article that names the brand and says what it does serves rankings and serves both routes into the answer. An unlinked mention in the same article serves the answer and not the rankings. A keyword-anchored link on a page that says nothing about the brand serves rankings at most, and often not even that. The weakness in the mentions camp’s position is the same as in the others: it is correlational, and large brands have more of everything. Google has also declined to endorse it. At Search Central Live in Sydney on 15 May 2026, Gary Illyes and Cherry Prommawin of Google did not confirm that authentic mentions give any direct benefit in AI search, according to Kenichi Suzuki’s account as reported by Search Engine Roundtable.
The main studies side by side
The table gathers the studies quoted on this page. The last column is the one to read first, because each study’s weakness tells you how far its finding can be pushed.
| Study | Sample | What was measured | Finding | Main weakness |
|---|---|---|---|---|
| Seer Interactive, January 2025 | 10,000 questions, GPT-4o through the API, finance and software | What tracks being named by the model | Google page one about 0.65, Bing 0.5 to 0.6, backlinks weak or neutral | Forums, aggregators and social sites removed. API answers, two industries |
| Ahrefs, May and December 2025 | 75,000 brands above Domain Rating 40 | Correlation of brand factors with mentions in ChatGPT, AI Mode and AI Overviews | Web mentions 0.66 to 0.71, branded anchors 0.51 to 0.63, Domain Rating 0.27 to 0.33 | Established brands only. Size drives every factor. Vendor |
| Ahrefs, August and September 2025 | 15,000 long-tail prompts, 3,311 short keywords | Overlap of cited URLs with Google’s top ten | ChatGPT 8% to 10%, Perplexity 28.6% to 65%, domain overlap 31.8% | Compares with the typed query, not the hidden searches |
| Semrush, October 2025 | 1,000 random domains, five AI surfaces | Authority Score and link types against AI mentions | Authority Score 0.65 and 0.57. Nofollow about equal to follow | Vendor’s own score. No test of cause. Share of voice much weaker |
| SE Ranking, November 2025 | 129,000 domains, ChatGPT | Average citations by referring domains and other traits | 1.6 to 1.8 citations at the low end, 8.4 at the top | Averages by bucket. Read through a trade report |
| Peec AI, March 2026 | 30 million US citations | Which domains are cited, by engine | Reddit and YouTube lead. Google leans social, ChatGPT editorial | Describes sources, not what gets a brand named. Vendor |
| Pitchbox, 2026 | About 20 software brands, four assistants | Link counts against position in recommendations | Referring domains 0.62 | Tiny sample, two categories, fame confound. Vendor |
Three patterns run through the table. Every study that samples widely finds authority tracking visibility, and every study that samples only established brands finds it tracking weakly. Every study that measures mentions finds them tracking more closely than links. And not one of them changes anything and watches what follows, so the whole table describes which things travel together.
Where each camp is right and where it overreaches
With the mechanism and the studies in view, the popular claims can be judged one at a time. The verdicts below are the site’s editorial reading of the evidence above, and each is only as firm as the weakest study it leans on.
| Claim | Verdict | Why |
|---|---|---|
| You need backlinks to appear in AI answers | Partly right | True for having your own pages cited on assistants fed by Google or Bing. Untrue for being named, which needs no link at all |
| More links or a higher authority score means more AI visibility | Mostly wrong above a threshold | Weak correlations among established brands. Stronger only when small sites are in the sample |
| Links no longer matter for AI | Wrong | Google and Microsoft both document retrieval through their ranking systems, and those use links |
| Rankings are irrelevant because cited pages do not rank | Overreach | Cited pages rank for hidden searches. Domains recur even when URLs differ |
| Only mentions matter | Overreach | Mentions track visibility best, but the evidence is correlational and retrieval still needs indexed, ranking pages |
| An unlinked mention is worth as much as a link | Right for the answer, wrong for rankings | The text is what a model reads. The link is what a ranking system counts |
| Mentions can be bought the way links were | Works briefly, then breaks rules | Disclosed placements are rarely cited, and undisclosed ones are deception |
The link camp is right about plumbing and wrong about volume. It is correct that retrieval runs on indexes, that indexes rank, and that a site nobody links to will struggle to get its own pages in front of an assistant. Where it overreaches is in selling the old unit of work, a number of links per month to a target page, as if AI visibility scaled with the count. The evidence says it does not. Past the point where a site is findable and trusted, the correlations flatten, and what distinguishes the brands that get named is how widely they are discussed.
The sceptics are right about the target and wrong about the cause. A page ranking first for a head keyword is no longer a reliable ticket into the answer, and anyone still reporting AI visibility as a list of keyword positions is measuring the wrong thing. But the low overlap figures come from fan-out, which multiplies the searches. It does not abolish ranking. Each hidden search is still a search against an index, and the old question of which pages deserve links is still a live one for the pages that could answer those searches.
The mentions camp is right about the direction and wrong about the exclusivity. Mentions are the best single correlate anyone has measured, and the mechanism holds up on both routes. The overreach is in the word “only”, and in the leap from correlation to promise. A brand cannot be retrieved from an index its pages are not in, and it cannot control what third parties say unless it has a site worth citing for the facts. There is a quieter problem as well: the camp’s favourite evidence was gathered on brands that already had authority, so it shows what matters after the entry condition is met and says nothing about meeting it.
How to split the effort in practice
The practical consequence is that links and mentions should be planned as one programme with two measures of success, with the balance set by where the site stands. The detailed plan, from choosing prompts to getting onto cited pages, is in link building for AI visibility, so what follows is about emphasis. One principle applies in every case: prefer the placement that does both jobs. Coverage that names the brand, describes it accurately and links with the brand as the anchor serves rankings, retrieval and memory at once, which is why digital PR sits at the centre of most sensible plans for this.
A new site
A new site’s first problem is the entry condition, and that is a links problem before it is anything else. Picture a small online shop selling refurbished film cameras, six months old, with a handful of links and no press. Assistants cannot cite pages their indexes barely hold, and the model has never read about the shop, so both routes are closed. The early effort belongs to the unglamorous base: being crawlable by every assistant’s search bot, indexed in Bing as well as Google, listed where shops of its kind are listed, and linked from suppliers and partners. Your first 100 links puts those tactics in order of cost and risk, and notes that digital PR is a weak fit for a brand-new site with no credibility. For the shop, mentions at this stage come from places that take small sellers seriously: enthusiast forums, a few video reviewers, review platforms. The measure of success is modest. It is whether the brand appears at all for narrow, specific prompts.
An established site
An established site has usually cleared the threshold, which changes the arithmetic. Consider a mid-sized maker of walking boots with ten years of links, steady rankings and patchy AI visibility. More links of the kind it already has will add little, on the evidence above. The gap is far more likely to be in what third-party pages say: the comparison lists that omit it, the review sites with stale details, the video roundups that feature three rivals. Here the weight shifts towards mentions in the places assistants read for that category, towards original data and expert comment that journalists will quote, and towards turning unlinked mentions into links where a publication allows it. Link building does not stop. It narrows to the pages that answer the recurring hidden searches, such as pricing, comparisons and alternatives, and it favours branded anchors over keyword ones, for the reason given earlier: a branded anchor is a mention as well as a link.
Entity work belongs in this tier too, as hygiene. One consistent description of what the company does, repeated on its own site, its profiles and its press material, makes it more likely that scattered mentions are credited to the same brand. Evidence that this changes AI answers directly is thin, and entity building says so, but the cost is low and the failure it prevents, a model confusing the brand with a namesake, is real.
A brand in a regulated or high-risk market
Regulated markets change the supply side of both links and mentions. Take a licensed online sportsbook operating in one country. Between 55% and 85% of publishers refuse restricted niches, according to ESBO’s September 2026 dataset of 16,625 publisher quotes, and 41.5% accept gambling links. Many of the community surfaces assistants favour are closed as well, since most subreddits ban promotion of gambling. So the mentions that reach an assistant in this market come from a narrower set of sources: licensed affiliates and comparison sites, sports media, regulator and licensing pages, and news coverage. The iGaming hub and the wider guide to high-risk markets cover how link supply works there.
Two consequences follow. Because so much of the available inventory is paid, the temptation to buy mentions the way the sector has long bought links is strong, and that is exactly the behaviour the next section describes as a policy risk. And because few independent sources exist, each one carries more weight in the answer, so accuracy on those few pages matters a great deal: the licence details, the terms, the facts a comparison site copies. The honest emphasis is on earned sports and data coverage, on properly disclosed affiliate relationships, and on keeping the ordinary link profile strong enough that the brand’s own pages are retrieved for the questions buyers ask.
What can get you into trouble
The argument about links and mentions has produced a market in manufactured mentions, and the rules have caught up with it faster than they did with paid links. Google’s spam policies page, read on 4 October 2026 and last updated on 28 August 2026, defines spam as techniques used to deceive users or manipulate Search systems, and gives as an example “attempting to manipulate generative AI responses in Google Search”. Trade coverage dates that wording to 15 May 2026. The page lists no separate policy called “inauthentic mentions”, although some coverage refers to one. The phrase comes from Google’s guide to generative AI features, which says that seeking inauthentic mentions across the web is not as helpful as it might seem and that its AI features depend on the systems that block spam. What counts as link spam covers the rest of the policy page.
Google staff have been direct about the parallel. At the Sydney event in May 2026, Illyes and Prommawin strongly cautioned against buying or manipulating mentions and compared the practice to paid links, which Google’s systems detect and ignore, again by Suzuki’s account as reported by Search Engine Roundtable. Bing has moved too. Search Engine Journal reported in February 2026 that Bing’s webmaster guidelines now cover content engineered to trigger citations and treat prompt injection as abuse. The practical reading is that a mention bought to plant a brand in AI answers is in the same position as a link bought to pass ranking credit: it may work for a while, it is against the rules, and the most likely penalty is that it quietly stops counting.
The specific practices are easy to recognise once named. Paid places in “best of” lists are one. Disclosed, they are advertising, and paid or advertorial content made up 0.3% of AI citations in Muck Rack’s May 2026 report, a vendor figure taken from a press release. Undisclosed, they deceive the reader, and paid link disclosure law applies. The share of digital PR practitioners who say they never pay for placements fell from 87.3% to 80% in BuzzStream’s State of Digital PR 2026, so the practice is growing. Before paying for any placement, it is worth running the page through the publisher page checker to see how many sites it links out to and whether the site sells placements. Self-published lists that rank the publisher first are another. Peec AI found they made up about 11% of AI citations in software between December 2025 and February 2026, and Lily Ray reported in February 2026 that several sites using them at scale lost 29% to 49% of their Google organic visibility, while noting it was one factor among several.
Seeded community posts are the third. Reddit said its detection caught about 25,000 spam posts and comments a day in the first quarter of 2026, and the brand damage when moderators expose a seeded thread tends to outlast the campaign. Hidden instructions aimed at assistants are the fourth: Microsoft Security reported in February 2026 that it found more than 50 attempts by 31 companies to plant recommendations in assistant memory. And placing commercial pages on a trusted host to borrow its standing is parasite SEO with a new target. GEO spam and AI manipulation documents each of these. A working test covers them all. If a mention exists only because money changed hands and the reader is not told, assume it is on the wrong side of the policy.
How to test the question on your own brand
The published studies describe averages across thousands of brands, and none of them can say whether links or mentions are the constraint for one particular site. That can be tested directly, with patience and a fixed method. The method matters because AI answers vary from run to run: SparkToro and Gumshoe’s January 2026 research found the same list of brands came back less than 1% of the time across 2,961 runs, so a single answer proves nothing and only rates across many runs mean anything. How to measure AI visibility sets out the metrics. The sequence below is a working method, not a validated procedure.
- Confirm access first. Use the AI crawler access checker to see whether robots.txt lets each assistant’s search bot in, and check that key pages are indexed in Google and Bing.
- Write 30 to 50 prompts a buyer would ask before they know your name, and keep brand-name prompts in a separate set.
- Run each prompt many times on each assistant that matters to you, by hand or with a tracker from the AI visibility tools comparison, and record two things per answer: which brands are named and which URLs are cited.
- Sort the cited URLs into your own pages, competitors’ pages and third-party pages, and note which third-party pages mention you.
- For the competitor pages that are cited, compare their referring domains with those of your equivalent page, as in a link gap analysis.
- Pick one change, make it, and run the same prompts monthly for a quarter.
The fourth and fifth steps are the diagnosis. If the answers are built mostly from third-party pages that name your rivals and not you, the constraint is mentions, and no quantity of links to your own site will fix it. If competitors’ own pages are being cited where yours could be, and those pages have clearly stronger link profiles, the constraint is retrieval, and links are a reasonable place to spend. If you are named often but your pages are never the source, you have mentions without citations, which usually points to thin or missing pages on the themes the assistant keeps searching. And if nothing of yours appears on one assistant while it does on others, go back to step one, because the cause is more often access than authority. Bing Webmaster Tools helps here at no cost, since its AI Performance report lists the grounding queries behind Copilot citations of your pages.
The sixth step is the nearest thing to an experiment that a single brand can run, and its limits should be written into any report. One site, one change and a few months cannot separate the effect of the change from everything else that moved in the period, including the assistants themselves. What it can do is stop a team arguing from other people’s averages. When a campaign earns coverage, confirm what was actually published: the backlink status checker shows whether a link is present and followed, and brand mention monitoring catches the coverage that named you without one. Log both, with dates, next to the monthly mention and citation rates. AI visibility tracking tools explains what the paid trackers add and what they cannot see.
What nobody knows yet
The honest list of unknowns is longer than any camp admits. Nobody outside the companies knows how ChatGPT’s providers, Perplexity’s index or Brave’s order their results, so the weight of links on three of the main assistants is undocumented. Nobody can say, for a given answer, how much came from memory and how much from retrieval, which means the split that explains the whole argument cannot be measured from outside. No public study has shown that adding mentions causes a rise in AI visibility, or that adding links does, or that the number of reviews changes anything. Google’s patent on implied links shows it has thought about counting unlinked references, and its staff, asked in 2026, would not confirm that authentic mentions help. The sources assistants favour have changed sharply several times in a year, so even a correct finding has a short shelf life.
It is also unclear how much any of this is worth in traffic. SE Ranking’s study of 101,574 sites put AI referrals at 0.32% of all traffic in 2026, as how much traffic AI search sends explains, so the value of being named is mostly influence on buyers who never click. That makes branded search and brand demand one of the few places the effect might show up in a site’s own data. And the ground is moving again. Assistants that act for a user, comparing options and returning a shortlist, would make third-party reputation a filter for who is considered at all, a possibility that for now is reasoned and unmeasured.
Another study claiming to settle the question will arrive shortly, and a handful of checks will tell you how much to believe it. Look first at who was in the sample, because a study of established brands and a study of random domains will reach opposite conclusions about authority from the same underlying reality. Check what the outcome was: a count of mentions, a share against competitors and a citation of a URL are three different things. Ask whether the comparison is with the typed prompt or with the searches the assistant ran. Note who published it and what they sell, without dismissing it on that ground alone, since vendors hold the only large datasets. Check the date against the last known change in the assistant’s behaviour. Above all, see whether anything was changed and measured afterwards. Until someone does that at scale, every finding in this field is a description of which things are found together, and should be acted on only where the action makes sense for other reasons as well.
Where to go next
The working plan that follows from this page is the one linked in the section on splitting effort, and it is the natural next read. If the vocabulary of the debate is the obstacle, what GEO, AEO and LLMO mean sorts the acronyms from the work underneath them, and the site glossary defines mention, citation, fan-out and the other terms used here. For the older version of this argument, about links and ordinary rankings, link relevance versus authority shows how the same habit of grading evidence applies there.
Common questions
Do backlinks help you get cited by ChatGPT?
Indirectly, on current evidence. ChatGPT search retrieves pages through search providers and its own crawler, so links help where they help a page get found and ranked. Counts of links are a weak guide to who gets named among established brands.
Are brand mentions more important than backlinks for AI search?
Mentions track AI visibility more closely. Ahrefs' December 2025 study found 0.66 to 0.71 for branded web mentions against 0.27 to 0.33 for Domain Rating. That is a comparison of correlations among large brands and does not show that mentions cause visibility.
Does Google use links to choose sources for AI Overviews and AI Mode?
Google's documentation says its AI features retrieve pages using its core Search ranking systems and that a page must be indexed and eligible for a snippet. Those ranking systems use links, so links matter there to the extent they matter for ordinary rankings.
Do nofollow links and unlinked mentions count for AI visibility?
Probably, because an assistant reads the text of a page and has no need of the link. Semrush's October 2025 study found nofollow links correlated with AI mentions almost exactly as well as followed links. Neither finding proves cause.
Can a site with no backlinks appear in AI answers?
It can be named if other sites talk about it, since mentions need no link. Having its own pages cited is harder, because a page has to be discovered, indexed and ranked for the hidden searches an assistant runs, and links help with each of those.
Is buying brand mentions against Google's rules?
Google's spam policies define spam to include attempting to manipulate generative AI responses in Google Search. Google staff at Search Central Live in May 2026 were reported to compare bought mentions to paid links, which Google's systems detect and ignore.
Should I stop link building and do digital PR instead?
No. Keep earning relevant links, because rankings feed retrieval, and prefer placements that also name and describe the brand. Digital PR does both jobs in one placement, which is why it suits this goal well.
How do I know whether links or mentions are my problem?
Run a fixed set of buyer prompts many times and record who is named and which pages are cited. If cited pages never mention you, the gap is mentions. If competitors' own pages are cited and yours are not, the gap is retrieval, which is where rankings and links come in.

