The Advice Loop Nobody's Testing
If you've spent any time on a GEO or AEO Slack channel this year, you've heard the same advice on loop. Be on Reddit. Be on YouTube. Meet buyers where the community already trusts.
That advice is not wrong. It's just incomplete, and in some cases quietly misleading.
I've watched teams reallocate six-figure content budgets toward Reddit engagement squads and YouTube shorts because a smart-sounding person with a newsletter said that's where AI citations live. The anecdotal evidence is convincing. A ChatGPT answer about project management tools does tend to mention a few Reddit threads. So you go look. You screenshot. You write the LinkedIn post. You budget the headcount.
Then someone runs an actual study, and the picture gets more complicated than the Slack-loop version.
What the Citation Data Actually Shows
Search Engine Journal covered a Ten Speed analysis of 1,712 commercial evaluation prompts run across four AI engines: ChatGPT, Perplexity, Gemini, and Google AI Mode. The setup was simple. Ten Speed picked 428 prompts that looked like things a buyer would actually type when close to a decision, ran each across the four engines, then hand-classified every citation URL into buckets like product pages, comparison content, blog posts, review sites, and community forums.
The headline: brand-controllable content accounts for roughly 88% of AI citations at the evaluation stage. That's not a rounding error against the community-channel narrative. It's the opposite of it.
Within that 88%, product pages alone pulled 24% of all citations. Homepages contributed another meaningful slice. Together, those two formats land close to a third of citations at the exact moment a buyer is narrowing a shortlist.
Compare that to the community channels everyone is panicking about. Ten Speed's number for Reddit plus YouTube combined: about 4%.
Four.
Why Product Pages Outperform the Formats Marketers Love
Here's the part I keep repeating to clients. AI engines at the evaluation stage are doing a specific job. Someone asks "what does this tool do, who is it for, what does it integrate with" and the model needs a plain-language answer it can paraphrase.
Your product page is the only place on the internet where you write that answer in your own words. Not a blogger guessing. Not a Redditor half-remembering a demo. You. With pricing tables, integration lists, positioning statements, and feature descriptions written by someone who has actually touched the roadmap.
The models lean on that page because it is dense, structured, and current. Clever brand copy with lots of metaphor and personality actually performs worse here. The model wants facts, not voice.
That's a weird result for a lot of content marketers. We were trained to make product pages sing. AI citation data suggests we should make them answer instead. If you want a broader framework for how brands earn recommendations inside AI answers, we broke one down in our guide to winning AI search recommendations.
Comparison Pages Punch Above Their Weight
Ten Speed found that comparison content (the classic "X vs. Y" format) showed up at 1.33x the rate you'd expect from its share of the corpus. If your comparison pages exist only because sales begged for a battlecard against your single most annoying competitor, you're underinvesting in the format that's already outperforming its share of buyer questions.
Build comparisons against more than the obvious rival. Build them against the adjacent category, too. The models use these to populate shortlists in prompts that never name a vendor directly.
The Reddit Counterpoint, And Why It Still Matters
Now for the uncomfortable part. The Foundation lab ran a separate study on Reddit citations in B2B SaaS with AirOps and landed somewhere very different: Reddit accounted for 21% of third-party citations in their B2B SaaS dataset.
Both studies can be true. They measured different things. Ten Speed counted all citations including owned domains. Foundation split the denominator to look only at third-party surfaces. Change the denominator, get a different headline.
That's the trap. Every AI citation stat floating around Slack right now is smuggling a definition of "citation" and "relevant" that nobody is reading past. Foundation's own conclusion is worth quoting in full because it's the most honest sentence written about GEO this year: no single source wins, but stacked presence does. Same insight repackaged across your blog, a LinkedIn post, a YouTube breakdown, a Reddit thread, a support doc, and a trade publication pitch multiplies the number of surfaces a model can pull from. That stacking effect is exactly why brand mentions have become the new off-page SEO for AI search: even where a link never appears, repeated third-party presence gives models more surfaces to associate your name with a category.
Where YouTube Fits (And Where It Doesn't)
YouTube does not win at the evaluation stage. The Ten Speed data is clear on that. Where YouTube earns its keep is earlier in the journey and later in the journey. The explainer a buyer watches six months before they start comparing vendors. The teardown a frustrated customer watches after they've decided to switch. YouTube shows up when the prompt is broad and conceptual, not when the prompt is "best alternative to X with SOC 2 compliance."
If you're a B2B SaaS company and you're cutting YouTube because it doesn't earn evaluation-stage citations, you've made a good call for one funnel stage and a bad call for two others.
A Practical Reorder For Your Next Content Budget
The Ten Speed report itself closes with three moves, and I think they're right.
Rewrite product pages and homepages like a model is reading them cold. Plain language. Integration lists. "Who is this for" stated explicitly. Cleverness costs citations.
Build comparison content earlier, against more than your two obvious rivals. That 1.33x return on comparison format is too big to leave to sales-battlecard archaeology.
Manage G2 and Capterra profiles as content, not review inboxes. Review sites hit about 7.2% of citations in Ten Speed's data. Stale profile copy misinforms the model right when it's building someone's shortlist.
I'd add a fourth: before you paste any AI citation percentage into a deck, ask what the denominator is. Most of the numbers circulating this year will not survive that question, and mine included.
The Real Takeaway Isn't Anti-Reddit
Nobody is arguing that Reddit is dead. It isn't. It's a meaningful slice of third-party citations and an outsized slice of the trust you can't buy. Community platforms still carry real weight in the right contexts, which is why we've separately covered how multi-location brands win AI visibility via Reddit. But the GEO conversation has collapsed "where citations come from" into "where the loudest people on Slack say citations come from," and the data is telling a flatter, less exciting story.
Buyers on ChatGPT are mostly asking models to read your own website. Get the website right first. Everything else is stacking.
