geo

Brand mentions vs. backlinks in generative retrieval

What is documented, what is inferred, and what is guesswork about links and unlinked mentions in AI answers, plus the one practical move that holds up.

Nobody outside the vendors knows how much a backlink or an unlinked brand mention counts toward being named in an AI answer. Google documents that links and references from other sites feed its quality systems, and that its AI features run on the same Search infrastructure. OpenAI, Anthropic, and Perplexity document how their crawlers reach pages, not how sources are chosen. Everything else written on the subject, including the confident version of “mentions are the new links,” is inference.

That doesn’t make the question useless. It makes it a question with three layers, and the useful work is keeping them apart: what a vendor has written down, what follows from how retrieval works, and what is a guess. This post goes through all three and ends with the one practical move that survives under any of them: earn accurate mentions on the pages that already get retrieved for your questions.

None of it promises a citation. A citation is a probabilistic, multi-factor outcome, and the Citability and entity signals guide explains why this site treats it that way. This post is about which inputs you can reason about honestly.

What do the vendors actually say?

Google says the most, and what it says is about Search in general. Its How Search Works page on ranking states that “one of the factors used to determine quality is understanding if other prominent websites link or refer to the content,” and calls that “generally a good sign that the information is trustworthy.” Note the verb pair: link or refer. Google’s own wording covers references as well as links, without saying how each is weighed.

For AI Overviews and AI Mode, Google’s AI features documentation says there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary,” and that both may use “query fan-out,” issuing multiple related searches across subtopics. The reading that follows is plain: whatever Search already uses to judge a page, including links and references, is upstream of what its AI features can surface. Google doesn’t publish a separate weighting for the AI surfaces, and it doesn’t mention unlinked brand mentions anywhere on that page.

OpenAI’s crawler documentation says OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features.” That’s an access statement. It says nothing about how a page is ranked or chosen once it’s reachable. Anthropic’s and Perplexity’s crawler pages are the same kind of document: Perplexity says PerplexityBot exists to “surface and link websites in search results,” and neither page describes source selection. As of the check date, none of the three publishes anything about links or mentions as selection inputs.

ClaimStatusWhere
Links and references from other sites feed Google’s quality assessmentDocumentedGoogle, How Search Works
Google’s AI features use Search’s systems, with no extra requirementsDocumentedGoogle, AI features
Unlinked mentions are weighted the same as linksUndocumentedNowhere
ChatGPT, Claude, or Perplexity weigh backlinks when selecting sourcesUndocumentedCrawler docs cover access only
Brand frequency in training data shapes what a model “knows”InferredFollows from how models are trained; no vendor quantifies it

What follows from how retrieval works?

A retrieval-based answer is built from pages a system fetched for the query, so the pages that get retrieved are the only ones whose text reaches the answer. That’s the mechanical fact, and it leads to two different roles for links and mentions.

Links act upstream. Retrieval systems lean on a search index, and links have always influenced which pages an index finds and how high it ranks them. A link that helps your page rank for a query helps it get retrieved for that query and for the fan-out sub-queries around it. Query fan-out is why this matters more than it used to: one question can pull in several result sets, and each one is a separate chance to be retrieved.

Mentions act in the context. When a retrieved page mentions your brand, your name is now in the text the model is reading while it writes. It doesn’t matter whether that mention is linked. The model reads text, not link graphs. If a widely-retrieved roundup of approaches in your category describes your product accurately, an answer built from that roundup can name you without ever fetching your site.

That second role is the strongest argument for mentions, and it’s an inference, not a vendor statement. It also has a hard limit: a mention only enters the context if the page carrying it is retrieved. A mention on a page nobody’s retrieval system pulls for your topic does nothing through this route.

What is still speculation?

Three popular claims sit here.

“Mentions replaced links.” Nothing documented supports replacement. Google’s own text names both. The defensible version is that links do more of their work upstream (indexing, ranking, retrieval) and mentions do more of theirs downstream (what’s in the context window), and a page benefits from both.

“Co-occurrence trains the model to associate your brand with a topic.” Plausible for what a model knows without retrieval, since training corpora are built from web text. No vendor quantifies it, it can’t be audited from the outside, and it acts on a timescale of model releases. Don’t put it in a plan you have to defend.

“Unlinked mentions are an authority signal in AI ranking.” This is the version vendors would need to confirm, and none has. The research that does exist points elsewhere. The GEO paper from Princeton and collaborators, accepted to KDD 2024, reported visibility gains of up to 40% in generative engine responses. The methods it tested were rewrites of the source page’s own text, not off-site mentions. That’s evidence that page-side content matters, which What makes a page citable covers. It is not evidence about mentions either way.

Where do you earn mentions that get retrieved?

The practical answer is: on pages that already get retrieved for the questions you want to be part of. That turns a vague brand-awareness goal into a list of URLs you can work through.

Step 1: write the question list. Take ten to twenty questions a buyer actually asks in your category. Pull them from Search Console query data, support tickets, and sales calls, the same sources the prioritization sheet uses for its question-demand column.

Step 2: record what gets cited. Run each question by hand in the assistants you care about and record every cited URL in a spreadsheet: question, assistant, date, cited URL, whether your brand is named. Expect the answers to change between runs; What prompt tracking can and can’t tell you explains why one run is an anecdote. Run each question several times on different days, and keep only URLs that recur.

Step 3: classify the recurring URLs. Add a column for page type. Most will fall into a few groups:

Page typeHow a mention gets thereWhat to avoid
Category explainers and guidesOffer data, a definition, or an expert quote the author can useAsking for a link in exchange for anything
Directories and listingsSubmit or correct your entry through their normal processDuplicate or inconsistent entries
Community threads and Q&AAnswer the question well, disclose who you areUndisclosed promotion
Documentation of adjacent toolsIntegration pages, where a real integration existsClaiming integrations you don’t have
Your own pagesAlready yours; check they’re accessible and accurateAssuming they’re retrieved because they rank

Step 4: fix accuracy before volume. On the recurring pages that already mention you, check what they say. An outdated price, a discontinued product, or a wrong category label on a retrieved page is going into answers today. A correction request to a page that gets retrieved is worth more than a new mention on a page that doesn’t.

Step 5: make the mention unambiguous. A mention helps only if the model can tell which entity it refers to. Use one consistent brand name, and make your own entity legible with Organization schema and sameAs, so that “Acme” on a third-party page and “Acme Robotics” on yours resolve to the same thing.

At scale, the spreadsheet is the method. Fifty questions run three times in two assistants is a few hundred rows and an afternoon of work, and a pivot table on cited URL by recurrence gives you the target list without any tooling. If you want to check whether those third-party pages are reachable and readable at all, a crawler GUI pointed at the URL list in list mode will tell you status codes and robots.txt blocks for the whole set.

Yes, through the route Google documents: links and references feed how Search judges quality, and Google’s AI features sit on top of Search. For the other assistants, it’s undocumented. The reasonable working assumption is that whatever helps a page rank in the search index an assistant retrieves from also helps it get retrieved.

What changes is what you ask of a link. A link from a page that is itself retrieved for your questions does two jobs at once: it’s a link, and it puts your name in the context when that page is read. A link from a page nobody retrieves does only the first. When you’re choosing where to spend outreach time, the retrieval spreadsheet is a better filter than domain-level metrics, because it’s measured against the questions you care about.

One boundary applies to both. Google’s spam policies define link spam as “creating links to or from a site primarily for the purpose of manipulating search rankings,” and list buying or selling links for ranking purposes, unless the link is qualified with rel="sponsored". Nothing published exempts AI surfaces from that. A paid placement written to look like an editorial mention carries the same risk it always did, plus the reputational one when a reader works out what it was.

How do you measure whether it worked?

Re-run the question sheet monthly and track two things per question: whether the recurring cited URLs changed, and whether your brand is named in answers built from them. Report it as observed frequency over a stated number of runs, with the date, not as a ranking. Measuring AI visibility without overstating it covers how to phrase that number so it holds up.

Keep the page-side work separate in the report. Whether your own pages can be fetched, parsed, and quoted is deterministic and can be re-checked after each change. The GPTBot vs OAI-SearchBot field guide covers which crawlers have to be allowed for your own pages to be retrievable at all. Mentions are off-page and probabilistic. Mixing the two in one score hides which one moved.

GEOBee covers the page-side half: whether the AI crawlers can reach your pages and whether those pages carry the entity, author, and citability signals listed in the three lenses. It doesn’t track mentions or citations. If you want that half audited on your own machine, download GEOBee and run the 7-day trial on your site.

On this page