Community as a Ranking Signal: Brand Mentions Beat Backlinks 0.66 to 0.22

Ahrefs measured 75,000 brands and found branded web mentions correlate with AI Overview visibility at 0.664, against 0.218 for backlink count. Community work produces mentions at volume, which is why community budgets have become a search argument. The honest version of that argument is narrower than the one most decks make, and it survives a finance review.

Short answer. Community investment shows up in AI search as brand mentions, and mentions are the factor most closely associated with visibility in every large correlation study published so far. The relationship is measured, not proven causal. Budget cases that promise mention growth are defensible. Cases that promise revenue attribution are not.

What a community actually contributes to a retrieval system

Community work does not produce backlinks at any useful rate. It produces mentions: your product name typed by somebody who does not work for you, on a page a crawler can read.

For twenty years that was a rounding error. Search ranked pages, pages were ranked by links, and a name in a forum thread with no anchor tag around it counted for nothing you could measure.

That has changed, and the change is measurable rather than rhetorical.

The unit is a mention, not a link

A backlink is an endorsement with an address attached. A mention is your name inside somebody else's sentence. Classical search had no way to price the second one, so it priced the first.

Ahrefs put both against the same outcome across 75,000 brands, using Spearman correlation. Branded web mentions landed at 0.664. Backlink count landed at 0.218, below Domain Rating at 0.326 and below referring domains at 0.295.

Read that as a hierarchy rather than a verdict. It says the things people say about you now track AI visibility more closely than the things people link to.

Retrieval happens before ranking

An answer engine does two separate jobs. It retrieves candidate passages, then it writes an answer out of them. Your page can rank respectably and never enter the second step.

This is why community sits differently from ordinary content marketing. A forum thread is a candidate passage on a domain the retriever already trusts, written in the question-and-answer shape these systems were trained to summarise.

Where citation and classical rank pull apart is examined in the analysis of citation and rank overlap. The compressed version is that they are two races, and you can lead one while losing the other.

The correlation evidence, stated at its real strength

Two Ahrefs studies carry most of the weight here, and they are worth reading with the caveats attached rather than the headline alone.

0.664 against 0.218

The first study, published 26 May 2025, sampled 75,000 brands and correlated eleven factors against brand visibility in Google AI Overviews. Branded web mentions came first at 0.664. Branded anchors followed at 0.527, then branded search volume at 0.392.

A December 2025 follow-up extended the same method to ChatGPT and Google AI Mode. Branded web mentions held at 0.664 on ChatGPT and 0.709 on AI Mode. YouTube mentions scored higher still, at 0.737 on ChatGPT.

Ahrefs states the limit plainly in both posts: correlation is not causation, and every coefficient measured sits in the moderate to weak band. I would go further. Nobody has published an experiment where mentions were added to a holdout set and visibility moved. Until somebody does, this is a strong association and nothing more.

The quartile gap is the number for a budget meeting

A coefficient does nothing in a finance conversation. A quartile table does.

Ahrefs split the sample by web mention volume. Brands in the top quarter averaged 169 AI Overview mentions. The quarter directly below averaged 14. The bottom half registered close to nothing.

That is a step function, not a gradient. It tells a community lead something specific: incremental mention growth inside the bottom half buys very little, and the return appears only once volume crosses into the upper band. Budget requests should be sized against that shape, not against a linear promise.

Why forum content dominates the citation lists

Peec AI analysed 30 million cited sources across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews. Reddit came first, YouTube second, LinkedIn third, with Wikipedia and Forbes in the top five, as reported by Search Engine Land in March 2026.

The mechanism is not mysterious. These systems answer questions, and forums are where questions were already being answered by people with no commercial stake in the reply.

Tinuiti, working with Profound, tracked seven assistants over four months on commercial-intent prompts. Social and community sources rose from 6% of all citations in October 2025 to 9% in January 2026. Reddit roughly doubled its own share over the same window, from about 2% to about 5%.

99% of the citations point at threads, not brand pages

The most useful line in that study is not a share figure. Tinuiti found that roughly 99% of Reddit citations pointed at individual threads with substantive discussion, rather than at subreddits, profiles or brand-authored content.

That single sentence should reshape most community strategies. The asset being cited is a conversation you did not write and cannot edit. Your brand page on the platform is close to worthless as a citation target.

It also explains why the usual content playbook transfers badly here. Publishing more of your own material is a different lever, discussed in the piece on original research as a content moat. Community is the lever that produces text on somebody else's domain.

The measurements disagree, and that changes what you can promise

Two credible studies put Reddit's importance at wildly different levels. Semrush reports Reddit appearing in tens of percent of ChatGPT responses. Tinuiti reports Reddit at about 5% of all citations.

Both are right. They count different objects. Share of responses asks how often a domain shows up anywhere in an answer. Share of citations asks what fraction of all citation slots a domain fills. A domain cited once in most answers scores high on the first and low on the second.

If you quote either number in a budget deck without saying which one it is, somebody numerate will find the other one and your credibility goes with it.

Three of these four are vendor research published by companies selling visibility tracking. That does not make them wrong. It does mean none has been independently replicated, and the sampling frames differ enough that the numbers cannot be compared directly.

One model update cut Reddit's ChatGPT share by five sixths

Semrush analysed 230,000 prompts producing more than 100 million citations between 14 July and 12 October 2025. Reddit appeared in roughly 60% of ChatGPT responses in early August. By late September it appeared in roughly 10%.

Wikipedia fell in the same window, from about 55% to under 20%. Crucially, neither drop showed up on Google AI Mode or Perplexity, which stayed relatively consistent throughout.

That is the single most important fact in this post for anyone signing a cheque. The value of a community mention is set by a source-weighting decision inside a model you do not control, and it can be revised without an announcement.

Crawlability decides whether the work produces any signal

Here is the part most community plans get wrong, and it is the cheapest thing on this page to fix.

A mention only counts if a crawler can read it. That makes the platform choice, not the engagement rate, the variable that determines whether community work has any search return at all.

A private Discord is invisible to a retrieval system

Discord messages sit behind a login. They are not indexed, so a thriving Discord with thousands of daily messages produces exactly zero retrievable text and zero third-party mentions.

Teams that want both outcomes mirror public channels to indexable pages, using open tools such as Answer Overflow or Linen. That is a real fix rather than a workaround, and it converts an invisible asset into a crawlable one.

A public forum on your own domain does the same job with fewer moving parts. Google supports DiscussionForumPosting markup, which makes a forum eligible for the Discussions and Forums treatment. Note the word eligible. Markup makes a page legible to a parser. It does not make anybody cite it.

What a community budget is actually buying

Three distinct assets, with three different levels of evidence behind them. Separating them is what makes the pitch honest.

The first is aggregate licensing value, and it has a public price. Reddit signed a content licensing deal with Google reported at $60 million a year in February 2024, and a comparable arrangement with OpenAI estimated at around $70 million a year, per the Columbia Journalism Review in October 2025. That price is paid to the platform, not to you. It tells you what the buyers think community text is worth.

The second is retrievable passages: threads that answer the questions your buyers ask, sitting on domains the retrievers already favour. This is the asset the citation studies actually measure.

The third is mention volume, which is the only one you can count weekly. It is also the one that maps onto the Ahrefs quartiles, which makes it the natural reporting metric.

What a community budget is not buying is trackable traffic. Assistant answers frequently resolve a query without a click, and the measurement problem that creates is set out in the piece on measuring dark discovery. Where clicks do arrive, they behave differently from search clicks, which is covered in the analysis of AI referral conversion rates.

How to build the case without overclaiming

The failure mode in community budget requests is not weak evidence. It is claiming a causal chain the evidence does not support, then losing the whole request when one link is challenged.

Split the claim into what is measured, what is yours to count, and what nobody can currently prove. Then state the third category out loud, before your CFO does.

The version of the pitch that survives a CFO

Bring three numbers and one baseline. Count unlinked brand mentions on crawlable pages. Count how many assistant answers name your product across a fixed prompt set. Record both before the spend starts, then re-measure monthly with the same tool and the same prompts.

Report those as counts, not as revenue. A count with a stated method is defensible. A modelled revenue figure built on a correlation of 0.664 is not, and the same standard applied to vendor claims is set out in the piece on verifying case study outcomes.

Where this argument is weakest

Three objections, and the third is the one I would raise if I were reviewing the budget.

Reverse causality is the obvious problem. Large brands get mentioned because they are large, and they appear in AI answers because they are large. The correlation may be measuring size twice. Nothing in the Ahrefs method separates those, and Ahrefs does not claim it does.

Most of the citation research is also vendor research. Semrush, Ahrefs, Profound and Peec AI all sell products that benefit if you believe AI visibility is measurable and improvable. The work looks methodologically careful and the sample sizes are large, but none of it has been independently replicated, and the field has no shared definition of a citation. Content saturation adds a further wrinkle, discussed in the piece on saturation and visibility.

The platform can collapse under you

Stack Overflow is the case study nobody in community marketing wants to sit with. It received more than 200,000 questions a month at its 2014 peak. In December 2025 it received 3,862, a fall of about 78% year on year, per devclass. By July 2026 the figure had reached four digits at the low end.

A brand that had built its developer presence on Stack Overflow in 2020 held an asset that shrank by four fifths in two years. The cause was partly AI assistants absorbing the questions and partly a community culture that punished new participation.

So the durable position is not that community is a ranking signal. It is that mentions are a signal, community is currently the cheapest way to produce them, and the venue will change. Build for mention volume across several crawlable surfaces, and hold any single platform loosely.

Frequently asked questions

Where to start this week

Start with an inventory, because most teams do not know which of their community surfaces a crawler can even see.

List every place your community currently talks. Mark each one crawlable or not crawlable. Anything in the second column is producing retention value and no search value, and that is a fact worth knowing before the next planning cycle, not after it.

Then set the baseline. Pick 20 prompts a buyer would actually type, run them across two assistants, and write down how many name your product. That number, dated and repeatable, is the only thing that will make next year's community budget argument on evidence rather than on faith.

Related reading