A rival's site showed up in 57% of the answers, the brand's own in 6%
Ask whether writing on your own site gets you into AI answers and you get two confident replies. One says publish depth on your own domain and the engines will find it. The other says your own domain barely matters. Neither side counts, so we did.
Between 30 September and 3 October 2026 we pulled every citation our sampler stored for unbranded questions, the kind a buyer types before they know your name. Six brands in our own accounts produced citations, spread across mattress retail, construction estimating software, a women's health app, interactive fiction, media training and AI visibility. Four engines return a source list we can read: our ChatGPT engine on OpenAI's API, Google AI Overviews, Perplexity and Claude. 1,205 answers carried at least one citation, 13,386 citations across 2,174 domains. Read every number here as ours.
- 681 of the 1,205 answers, 56.5%, linked at least one of the eight rivals the brand tracks. Interval 53.7% to 59.3%.
- 74 of the 1,205 answers, 6.1%, linked the brand's own website. Interval 4.9% to 7.6%.
- By citation: 2,162 went to those eight rivals' sites, 16.2%. 116 went to the brand's own site, 0.87%, interval 0.72% to 1.04%. Eighteen rival citations for every one of the brand's own.
- Three of the six brands got exactly zero citations to their own website, across 1,805, 1,714 and 393 citations respectively.
Claude cited an own site 0.16% of the time, AI Overviews 2.49%
The blended number hides a fifteen fold spread, and the spread decides where you spend.
- Google AI Overviews: 56 of 2,248 citations were the brand's own, 2.49%, interval 1.92% to 3.22%. 26 of its 292 answers linked the brand, 8.9%.
- Our ChatGPT engine: 32 of 1,927, 1.66%, interval 1.18% to 2.33%. 26 of 286 answers, 9.1%.
- Perplexity: 21 of 4,875, 0.43%, interval 0.28% to 0.66%. 15 of 325 answers, 4.6%.
- Claude: 7 of 4,336, 0.16%, interval 0.08% to 0.33%. 7 of 302 answers, 2.3%.
- The two engines that cited own sites least cite most overall. Perplexity averaged 15.0 sources per answer and Claude 14.4, against 7.7 for AI Overviews and 6.7 for ChatGPT. The longer source list made no room for the brand, but it had room for the rivals, who held 19.7% of Perplexity's citations and 11.3% of Claude's.
18 of the 25 most cited domains in one category were companies' own sites
A share tells you whether to care. Reading the domains tells you what is going on. We opened all 25 of the construction estimating brand's most cited domains. 18 were a company's own commercial website, 15 of them selling construction or estimating software. The other 7 were YouTube, Reddit, Google, two software directories and two statistics sites.
- Those 18 domains took 574 of that brand's 1,714 citations, 33.5%, interval 31.3% to 35.8%. The seven non-vendor domains took 168 between them.
- 508 of the 574 were an inner page rather than a homepage: product pages, comparison pages, and in two cases a vendor's support site.
- The brand's own website appeared in none of the 148 answers that carried a citation.
A study of 149,912 citations outside ours lands in the same place
Six brands over four days is thin, so here is a larger sample that is not ours. Ranqo published a study on arXiv in June 2026 covering more than 100 brands across ChatGPT, Claude, Perplexity, Gemini and Grok, sampled between March and May 2026. We read the paper itself.
- Of 149,912 citations, 4,392 pointed at the tracked brand's own domain. 2.9%.
- 112,763, or 75.2%, pointed at what the paper calls corporate or third-party brand pages: commercial sites owned by businesses other than the tracked brand, meaning competitors, peers and vendors in the same category, homepages and product pages included.
- Own and third-party corporate pages together came to roughly 78% of the citations. Software review sites came to 1.1%.
- Their 2.9% sits above our 0.87%, and sampling explains the direction: their prompts had already mentioned the brand, ours are unbranded whether it came up or not.
What we'd actually do
Cheapest first.
- Split your own domain out of the third-party bucket before you brief anybody. A blended third-party number cannot tell you whether the engine prefers other people's pages or just your rivals' pages, and those call for different work.
- Rule out the eligibility gate before rewriting a word. Google documents that to appear as a supporting link in AI Overviews or AI Mode a page must be indexed and eligible to show with a snippet, and states there are no additional technical requirements.
- Pick the engine you are fighting on. Own-site citations were about fifteen times more likely on AI Overviews than on Claude here. If your buyers are on Perplexity, your own pages are a weak citation surface and your rivals' pages are not.
- Read the rival pages that got cited, not their homepages. 508 of 574 vendor citations in the construction sample were inner pages answering something specific.
- Stop reading zero own-site citations as a verdict on your writing. Three of our six brands sat at zero while fifteen or more rival vendor sites turned up in the same answers. That is an absence from a set, and the set gets built from pages that already get cited.
Where this could be wrong
Six brands over four days is thin and one of them supplied 49% of the citations. The categories are a deliberate spread rather than a sample of anything, so a different six would move these numbers.
Our rival figure counts eight tracked competitor domains per brand, so 16.2% is a floor and not the total vendor share. The real figure is higher, which the construction read shows: 18 hand-checked domains took a third of that brand's citations. We did not classify all 2,174 domains, and 1,057 of them were cited exactly once.
Our earlier co-citation post put 72.7% of cited domains on a software sample outside the vendors'. That counted domains in a smaller sample; this counts citations and splits yours from theirs. We would not stack the two.
These counts come only from the citation annotations each engine emits. An engine can read a page without listing it, so retrieval is undercounted and a page could inform an answer without appearing among these 13,386. Perplexity and Claude only began returning citation rows for us on 30 September, hence the short window.
Nothing here says your own site does not matter. It says your own site was rarely the cited link while your rivals' sites frequently were. Being crawlable and indexed is still the price of entry.