99.99% of AI citations point somewhere that isn't your website
That is the number to start from. Of the 49,391 citations analysed in Victorious's Q2 2026 quarterly report — covered on 29 July 2026 by Search Engine Journal — 99.99% pointed at third-party sites. Four brands out of 150 earned a citation to their own domain. Four.
Which means the SEO work you can control is not the work that decides whether an assistant names you. Your site is the material it describes you with, once it has decided to name you. The naming happens on other people's domains.
Your site is the source of the description. The shortlist gets written somewhere else.
Recognition is not recommendation
The same report tested 175 brands across five verticals — legal, healthcare, SaaS, financial services and ecommerce — on eight platforms, from ChatGPT to Meta AI. Put the brand name in the prompt and 96% of the brands with evaluable answers are described accurately. Ask about the category instead and, across the 150-brand cohort the mention rate was measured on, only 11% appear. The other 89% never do.
Two different tests, treated as one by almost everybody. The first asks whether the model has you. The second asks whether it lists you. We covered the underlying split in GEO versus SEO.
The 2,000-mention threshold
Brands with fewer than 2,000 third-party mentions appeared in AI answers just 3% of the time. The strongest measured correlation is not with what you publish but with the count of domains that mention you: 0.49. Third-party web mentions: 0.45.
And the number that should worry a service business most: on prompts where someone describes a problem rather than a category — "my site loads slowly and I'm losing orders, what do I do?" — a brand was named in 0.10% of answers. The assistant explains the fix and names nobody who could deliver it.
Two thousand sounds unreachable for a small firm. It helps to know what counts: not bought links, but appearances. Supplier directories, trade press, "best agencies for X" lists, complete profiles, reviews, podcasts, signed technical comments.
What we measured on our own site
Our on-page work is finished: JSON-LD for organisation, article and FAQ, llms.txt, a question-and-answer structure in every piece, hreflang across three languages, an aggregate rating built from the 21 real Google reviews we have.
Search Console, 2–29 July 2026: 1,890 impressions, 13 clicks, average position 24.4, 125 pages with impressions. Our article on getting cited by AI search: 112 impressions, average position 38.7, zero clicks.
We finished what could be done on the site and it is not enough. The report explains why. We publish this because it is cheaper to hear it from us than to buy an on-page audit for an off-site problem.
89% of the demand has no owner yet
Semrush data for January–June 2026: 1,094 US product and service categories, five prompts each, more than 50,000 brands, over 600,000 citations tracked. Only 15.2% of categories have a clear owner — one brand in at least four of five prompts, five percentage points clear of the runner-up. 53.7% have several contenders and no winner. 89.3% of estimated demand sits in categories nobody owns.
Leaders are not comfortable either: 90.4% month-over-month retention, and a median lead of 2.9 percentage points for those who hold. Where the leader changes, the gap is 1.3 points. Easy to lose, which is the same reason it is winnable.
What to do, in this order
- Narrow the category until it means something. "Marketing agency" is a sector, not a category. "Google Ads agency for small hotels in Italy" has five realistic prompts and three competitors.
- Get mentioned where assistants already read. The 0.49 correlation is with the domains that mention you — trade directories, industry press, supplier lists.
- Publish a number that is yours. A guide gets paraphrased without attribution. A measurement you took has to be attributed.
- Judge results in impressions, not clicks. Inside an AI answer the mention is the delivery — the mechanism is in zero-click search.
- Keep on-page as the entry requirement. Without structure and schema you risk being described badly in the 96% of cases where you are described at all.
Three questions we get
Is this just link building with a new name?
No, and the difference decides your budget. Link building optimises for a link that passes authority. This optimises for a mention on a page an assistant is likely to read when answering about your category — often with no link at all. A directory listing with no follow link and a paragraph of context beats a bought link on an unrelated blog.
How do I check whether I appear in AI answers?
Not by asking for your company name — that is the 96% test and everyone passes it. Ask ChatGPT, Perplexity and Google AI Mode the category question your customer would ask, and write down the names that come back. Repeat monthly: with 90.4% month-over-month retention, the list moves enough to be worth checking.
How long before I show up?
Nobody claiming to know has the data. What can be said with numbers: the threshold observed sits at 2,000 third-party mentions, and below it the appearance rate is 3%. Not a one-month project, and not one that ends.
We do SEO and AI search for companies in Italy and Romania — in your position we would start with the ten questions your customer would ask, run across three assistants, and the names that come back today.
It describes you well. It just won't suggest you.
We run your customer's questions through ChatGPT, Perplexity and Google AI Mode and show you who appears instead of you. No pitch, just the list.
