There is a tidy story about AI search that a lot of agencies are currently telling. It goes: the AI engines read the web, the web is ranked by Google, so rank well and the citations follow. Keep paying for SEO and generative visibility arrives as a by-product.
It is a comfortable story because it requires nobody to change anything. It is also, as far as the available data goes, wrong.
What the measurement actually shows
Orbit Media published a citation study at the start of September 2026 covering 13,184 citations across 1,765 answers — 72 prompts across three B2B brands, run weekly against ChatGPT, Claude, Perplexity and Gemini, with data through 23 August. The author describes it as a directional sample rather than statistical proof, and that caveat belongs in every sentence that follows.
The study compared what the assistants cited against what ranked in Google for the same query.
At domain level, between 10% and 30% of cited domains also appeared in Google’s top ten, depending on the model and the question set.
At URL level — the level that actually matters, because a citation points at a page and not at a brand — the overlap collapsed. Roughly 1% to 5% for Gemini and ChatGPT. Perplexity was the outlier at 13% to 23%, which tracks with how heavily it leans on live retrieval rather than parametric knowledge.
The author’s conclusion: page-one ranking is neither necessary nor sufficient for AI citation.
Why this is not an argument against SEO
It would be easy, and wrong, to read that as “SEO is dead” — a claim we have already argued against at length.
Technical SEO is what makes a page retrievable at all. If your site cannot be crawled, if your sitemap is stale, if your CDN is challenging the crawlers, then nothing downstream works. Authority earned conventionally does make citation more likely at the margins. None of that changes.
What changes is the inference. You cannot look at a ranking report and conclude anything useful about your AI visibility. The correlation at URL level is somewhere between one in twenty and one in a hundred depending on the platform. That is not a proxy. That is a different measurement that happens to live in the same neighbourhood.
If your monthly report shows rankings and then asserts AI visibility, it is asserting, not measuring.
The second gap: cited is not the same as mentioned
There is a related confusion that flatters a lot of AI visibility dashboards.
In the same dataset: Claude mentioned the tracked brands in 70% of answers and cited them in 40%. Gemini, 72% and 41%. Perplexity, 67% and 44%. ChatGPT, 56% and 47%.
A mention is the model saying your name from what it already absorbed in training. A citation is the model handing a reader a link to your page. They are different commercial events. One builds familiarity; the other delivers a visitor who can convert.
The gap widens with brand size, which is the part most relevant to a South African SME. Semrush showed a 50-point spread between mentions and citations in the same run. Two much smaller sites showed spreads of four and five points. Big brands get talked about without being linked. Smaller brands, when they show up at all, tend to show up because a specific page earned the reference.
That is actually the encouraging reading. The citations a smaller firm earns are traceable to pages you can identify, measure and improve.
A supporting data point, with its label on
A second study released on 7 October 2026 by Sourcelift, a generative engine optimisation firm, carried the headline that traditional search rankings do not guarantee AI answer citations.
We want to be careful here, because the headline is stronger than the study. Sourcelift tested 40 US holiday gift questions against a single unnamed AI assistant on 5 October. It compared no rankings to citations at all; the rank claim in the headline rests on one anecdote about a software company that ranked page one and was named in two of ten answers.
What it did find is worth having: of the 36 answers that returned sources, 26 — 72% — cited no brand’s own website. Roughly a third of the 66 cited sites had been registered in the previous twelve months.
That is a real and interesting number about where AI assistants go looking. It is not evidence about ranking. We are separating the two because the industry habit of laundering a vendor headline into a law is a large part of why this field is so hard to navigate.
What to do instead of inferring
Measure the citation outcome directly. Fix a set of twenty to forty buying-intent questions a real prospect would type, and stop editing them. The value is in repetition.
Run it against each assistant separately. A blended AI visibility score hides the finding that matters most — that the platforms disagree with each other almost completely. The same study found all four cited the same domain for the same question in just 1.7% of combinations.
Report mentions and citations as two numbers. If they are collapsed into one figure you cannot tell whether you are becoming known or becoming useful.
Keep the variance. Running the same questions two days apart, Claude retained 30% of its cited URLs, Gemini 38% of its domains, Perplexity 65%. If your result moves that much in 48 hours, a one-off audit is describing a day, not a position.
Build the primary material. The consistent signal across the source preferences is that clear, specific, well-structured material on your own domain is what the more selective models reach for. Documentation. Pricing. Methodology pages. The unglamorous ones.
The honest limits
Three brands. Seventy-two prompts. One author who says plainly that it is directional. We are not presenting this as settled.
What we will say is that it is the most substantial independent measurement currently available, that a separate vendor study released in October points the same direction, and that it matches what we observe in client data. Ranking and citation are related jobs with a much weaker link between them than the standard pitch assumes.
If someone quotes you a single AI visibility number, ask which assistant, on what date, and what it was the day before.