Off-site mentions and citation sources
Your own site is the smaller half of this. The strongest published signal in the field points somewhere uncomfortable: being written about on pages you do not control.
Short answer
Across roughly 75,000 brands, branded web mentions correlated with AI Overview visibility at 0.664 against 0.218 for backlinks. The authors of that study state plainly that correlation is not causation, and so do we. It is still the strongest signal anyone has published, and it is where we spend the most effort: getting you described accurately on sites you do not own.
What the evidence actually says
Ahrefs measured branded web mentions against AI Overview visibility across roughly 75,000 brands. Mentions correlated at 0.664. Backlinks correlated at 0.218. Mentions tracked roughly three times more strongly than links.
Separately, Muck Rack's analysis of more than 25 million citations across ChatGPT, Claude and Gemini found earned media accounted for about 84 percent of AI citations. Paid and advertorial placements accounted for 0.3 percent.
Both figures are correlational and neither proves causation. We say that on every call. But the direction is consistent across two independent datasets, nothing points the other way, and the practical conclusion is the same either way: being included in a credible roundup is worth more than publishing your own.
For hospitality this matters more than average
An assistant answering “best restaurants in Lisbon” or “boutique hotel near Shibuya” is reading roundup articles, city guides, review platforms and forum threads. Your own site barely participates in that answer. It participates in the follow-up, when someone asks about you by name, which is a different and much easier query to win.
So the work splits. Your site has to be readable and accurate for the branded query. The wider web has to describe you correctly for the category query, which is the one that brings people who have never heard of you.
Which platforms, and one claim worth distrusting
The specific sources vary by category and city, which is why we sample before we act rather than working from a generic directory list. Some of what circulates about this is wrong. The widely repeated claim that Foursquare powers ChatGPT's local results did not survive testing: a run of 2,880 prompts across twelve metros recorded Foursquare citations at 0.00 percent, with Yelp grounding appearing in the overwhelming majority of runs, following OpenAI's non-exclusive Yelp licensing deal of 23 July 2026.
Treat that as one test rather than a settled fact. Treat anyone selling you Foursquare optimisation as a ChatGPT lever with rather more suspicion.
What we do
- Sample first. We run your category and city prompts and record which domains the assistants actually cite. That list, not a generic directory checklist, is the target.
- Fix consistency. Name, address, phone, category and attributes identical across every listing and review surface. Contradictions between sources are what make a model hedge or route around you.
- Earn the placements. Getting into the guides and roundups that came back in the sample. This is the slowest part of any engagement and the part with the most leverage.
- Rewrite what you control elsewhere. Your own description fields on third-party platforms, matched to the language your site uses, so the sources agree.
What we will not do
No paid placement sold as editorial. No link schemes. No mass directory submission. Paid and advertorial accounted for 0.3 percent of citations in the Muck Rack dataset, which tells you what it is worth, and the reputational cost of the other two is not worth the discussion.
Common questions
It overlaps heavily and we will say so. The difference is what gets targeted and what counts as success. Traditional PR chases reach and links. Here the target list comes from sampling which domains assistants actually cite for your category, and an unlinked mention that describes you accurately counts as a win, because the evidence values the mention over the link.
Longer than the technical work, and we cannot put a date on it. Editorial timelines belong to editors. What we can do is show you the target list, report what has landed, and sample the prompts on a fixed cadence so you can see whether the answers are changing.
No. Nobody can, and an agency that offers it is either buying placement and calling it earned, or misrepresenting what it controls. We commit to the target list, the outreach and the reporting.
What else an engagement covers
Want to know whether this is your problem?
Book a 30 minute call. We will fetch your site the way the crawlers do and tell you which of the prerequisites is failing. If none of them are, we will say so.