If you only read one number about AI search this year, make it this one: across 23,387 sources cited by five different engines, 77% of what the models quoted about a brand came from somewhere other than that brand’s own website.
That figure comes from Omniscient Digital’s January 2026 analysis of 240 branded prompts run across ChatGPT, Perplexity, Gemini, AI Mode and Google’s AI Overviews. It is the most useful piece of published work we have found on where AI answers actually come from, and it is worth sitting with, because it says something uncomfortable and then something genuinely useful.
Where the citations came from
Omniscient sorted every source into three buckets. The split:
- 48%Earned media — editorial 16%, forums and social 11%, review sites 11%, directories 10%
- 30%Commercial brand content — the paid and partner surface around you
- 23%Owned brand content — your actual website
Read that badly and you conclude your own site barely matters. Read it properly and you notice something else: the study also broke results down by what the person was asking, and the picture changes completely depending on the question.
The exception is the whole opportunity
Two intent categories moved hard in opposite directions.
On customer review queries — is this any good, what do people say, is it worth it — earned media took 82% of citations. You are not going to write your way into that one. Those answers are built from other people talking about you, and the honest advice is that a page on your site will not displace them.
But on functionality and integrations queries — does it fit, does it work with, what are the specs, is it compatible — owned content took 50%, the highest share of any bucket in any intent group. When someone asks a factual question about how your product works, the models go looking for the manufacturer’s own answer first.
That is not a small carve-out. In ecommerce, that is most of the questions that precede a purchase. Will these fit my car. Do they run true to size. Will it work with the model I already own. How long does it last. Those are functionality questions wearing everyday clothes, and they are the ones where your own page can win the citation outright.
Most answers do not recommend anyone
The second study worth knowing is Omniscient’s December 2025 look at 5,323 outputs across five models, using 180 prompts split evenly across three stages of buyer awareness. It asked a simpler question: does the answer name any brand at all?
- 19%Problem unaware — “my feet hurt after long runs”
- 28%Problem aware — “what causes toe numbness on trails”
- 79%Solution aware — “best wide-toe-box trail shoe”
The gradient is steep and it is the practical part. Four out of five answers name a brand once the shopper knows what kind of thing they want. Fewer than one in five do at the top of the funnel. If you are trying to appear in answers, the solution-aware questions are where the naming actually happens.
Category matters too. Electronics prompts returned recommendations 65% of the time; consumer products overall, 43%. So the surface exists in physical goods — it is just narrower than the B2B software numbers that dominate most GEO writing.
What we take from it
Three things, and we will say plainly which are ours and which are theirs.
Theirs: off-site sources carry roughly three quarters of brand citations; owned content wins on functionality and integration questions; brand naming climbs sharply with buyer awareness.
Ours: the practical move for an ecommerce brand is to stop trying to out-argue the review sites and start owning the answerable questions. Fitment. Sizing. Compatibility. Durability. What the product is not for. These are the questions where your catalog is the authoritative source, and where most storefronts currently publish a spec table instead of a sentence.
We have written more about how to do that in the paragraph your spec table never had.
One caveat we would want if we were reading this. Both studies weight toward B2B SaaS, and neither is ecommerce-specific. The direction is well evidenced; the exact percentages for a bike parts store or a running shop are not known, by us or by anyone. Measuring that properly for physical goods is what the AI Shopping Index exists to do.