The new backlinks: why third-party mentions drive AI recommendations

Ask a language model "what's the best help desk for small e-commerce teams?" and watch what it does. It doesn't rank pages. It doesn't tally backlinks. It produces the two or three names that its training data and retrieved sources most consistently associate with that description — and it repeats the characterizations it has seen most often. "Popular with small teams." "Known for fast setup." "Often mentioned alongside Shopify."

None of those characterizations came from a single authoritative page. They're aggregates — statistical residue of hundreds of independent texts saying roughly the same thing. Which points to the core strategic fact of GEO: you cannot corroborate yourself. Your website supplies the claims; only third parties can turn them into facts.

Consensus is the mechanism, not a metaphor

This isn't hand-waving about "authority." It follows from how models are built and how answers are assembled.

During pretraining, a model ingests enormous swaths of the public web — filtered web crawl, curated corpora, Wikipedia — as documented in the original GPT-3 paper.¹ The model has no ledger of which domain said what; it learns associations in proportion to how often and how consistently they appear. One page saying "Acme is the accuracy leader in humidity sensing" is a whisper. Forty independent pages saying it — in reviews, forum threads, distributor catalogs, conference writeups — is a pattern the model will reproduce without hesitation, and often without attribution.

At answer time, retrieval-augmented engines add a second consensus check: they fetch several live sources and synthesize. Claims that appear in multiple retrieved documents survive synthesis; claims that appear in one — especially a self-interested one — get hedged ("the vendor states…") or dropped. The GEO research bears out the direction: content carrying citations, quotations and statistics — external evidence signals — measurably gained visibility in generative answers, with lifts up to ~40% on some metrics, while shallow keyword tactics did nothing.²

And the platforms are institutionalizing it. OpenAI's partnership with Reddit explicitly brings community discussion into its products,³ which tells you what the labs think grounds a trustworthy answer: not brand copy — other people talking about you.

This is not link building with a new name

The reflex is to treat mentions as backlinks and reach for the old playbook: outreach at scale, guest posts, anchor-text spreadsheets. The reflex is wrong, because the scoring system changed underneath.

PageRank needed a hyperlink — a machine-verifiable edge in a graph — and cared enormously about where the link came from. A language model doesn't need the link at all. What it registers is co-occurrence: your brand name appearing near category terms ("field service management"), near attribute terms ("offline-first," "SOC 2"), and near or instead of competitor names, across many independent documents. An unlinked mention in a well-read forum thread can shape the model's associations more than a followed link in a footer nobody reads.

Notice the last failure mode. In link building, the risk was doing it too aggressively. In mention building, the bigger everyday risk is incoherence: ten placements that describe you ten different ways teach the model nothing, because no single association crosses the consensus threshold.

Five corroboration plays that actually work

Before the list, a filter for evaluating any placement opportunity: ask what sentence about your brand will exist afterward, and on whose domain. A logo on a sponsor wall produces no sentence and teaches the model nothing. A quote from your CTO in a trade-press piece produces exactly the kind of text — your name, your category, your claim, someone else's domain — that compounds. Judge every PR hour by the sentences it leaves behind.

With that filter in hand, here are the plays, ranked roughly by leverage per hour of effort for a B2B or considered-purchase brand:

1. Analyst and industry-body coverage. A single inclusion in a category landscape, market guide or association resource page seeds the exact sentence structure models love: "vendors in this space include A, B, C." That's co-occurrence with the category term, from a neutral source, in list form. Brief the analysts with your positioning phrase; they often reuse the language you hand them.
2. Niche publications over big mastheads. A trade publication your category actually reads is denser in category vocabulary than a national outlet — and its archive is exactly the kind of specialist text that survives corpus filtering. One substantive feature in a respected niche pub typically beats a passing mention in general press.
3. Comparison and review sites. "X vs Y" and "best X for Y" pages are the most-retrieved documents for shortlist queries. Make sure your listings exist, are current, and describe you the way you describe yourself. Fresh reviews matter twice: once as retrieval fodder, once as training-data reinforcement.
4. Podcasts — but only with transcripts. Audio is invisible to a text model; a published transcript is a few thousand words of your founder using your positioning phrase in natural conversation, hosted on someone else's domain. Always ask whether the show publishes transcripts, and offer to supply one if not.
5. Original data others cite. The compounding play. Publish a benchmark, survey or teardown that's genuinely useful, and every writeup of it repeats your name next to your category — corroboration you didn't have to ask for. This is also the tactic most aligned with what generative engines reward directly: statistics and citable claims.²

One reality check: almost nobody clicks the citations inside AI summaries — Pew measured about 1% of visits.⁴ So don't evaluate these placements as referral-traffic channels. Their job is to change what the answer says, not to be clicked.

The consistency discipline: one phrase, everywhere

If consensus is the mechanism, consistency is the strategy. Pick one positioning phrase — twelve words or fewer, containing your category and your sharpest differentiator — and use it verbatim in every context you influence: site, press boilerplate, review-site profiles, analyst briefings, podcast intros, speaker bios, partner pages.

"Verbatim" is the part teams resist, because writers are trained to vary phrasing. Resist the resistance. Paraphrase is expensive for the model: "collaborative whiteboard for remote workshops," "visual workspace for distributed teams" and "online canvas for team brainstorms" are three weak associations instead of one strong one. You are not writing for readers who see every instance; you are training a model that does.

Measuring it: share of voice, not placement counts

Link building had crisp accounting: referring domains went up, rankings followed, everyone got a dashboard. Mention building needs equally honest accounting, and placement counts aren't it — ten mentions that never move an answer are worth less than one that does.

The metric that captures the outcome is share of voice in sampled answers: across a fixed suite of category prompts, run repeatedly on each engine, what fraction of all brand mentions are yours? Corroboration work done well shows up as that number climbing over weeks and months — first on retrieval-heavy engines like Perplexity, which react to new sources within days, later in the slower parametric layer as models are updated.

• Fix a prompt suite of 20–50 real buyer questions and don't change it mid-experiment.
• Run each prompt multiple times per engine — answers are stochastic, single runs are noise.
• Track share of voice and mention rate with confidence intervals; judge movement against them.
• Annotate the timeline with placement dates (feature published, review batch landed, report cited) so you can connect cause to effect.
• Watch the wording of answers, not just the counts — the goal is your positioning phrase coming back to you in the model's voice.

That last one is the tell. The day an assistant describes you in words you wrote — on a page you don't own — the loop has closed: claim, corroboration, consensus, answer.

Expect the timeline to be lumpy rather than linear. Corroboration accumulates silently below the consensus threshold for weeks, then a retrieval index picks up two new sources in the same week and your share of voice steps up all at once. Teams that report weekly tend to panic during the flat stretch and overclaim during the step; monthly reporting against quarterly targets fits the physics of the channel better.

Frequently asked questions