What we know about how AI assistants choose their sources
Taken from answers we collected every day from 15 to 29 August 2026, about our own comparison website. Other people’s research is labelled as theirs.
The topic mattered more than the pages
On telehealth questions, AI assistants used our site in 31% of answers. On business software, 5%. Same site, same standards. On software, Reddit, Salesforce and Capterra already held the answers. Before you spend on pages, check whether your customers' answers already belong to someone else.
Evidence: 125 of 401 answers to 22 telehealth questions; 11 of 223 answers to 16 business-software questions.
Head-to-head comparisons get used by more than one assistant
Two of our three comparison pages were used by more than one assistant. The third was used by Perplexity alone. It's three pages, so read it as a strong pattern.
Evidence: Moshy vs Juniper, used by all three assistants in 41 answers; Mosh vs Pilot by all three in 31; Moshy vs Pilot by Perplexity only, in 19.
Being used as a source isn't the same as being named
One of our pages was used in 12 of 15 answers to one question. Across all 41 answers that used it, our business was named once.
Our reading, which we haven’t measured. Assistants seem to credit claims that need someone behind them, like a verdict, and quietly use ones that don't, like a price.
Evidence: The question “juniper vs moshy”, 12 of 15 answers; named in 1 of 41.
Where a fact first appears decides whether it's trusted
Our discount code is worth more than the merchant's own, and it still loses, because assistants say ours “comes from an external coupon site”. We're now testing a code published first on our own page.
Evidence: Our code is $120 off, the merchant's $100. Theirs appeared in 37 sentences, ours in 24; ChatGPT quoted theirs 9 times and ours once.
The assistants read completely different things
Claude used Reddit in 0 of 88 completed answers that carried a source list. Perplexity used around fifteen websites per answer, ChatGPT around three. Advice for one assistant can be wrong for another.
Other people’s research
Otterly (June 2026, software questions) found Reddit cited zero times in 379,321 Claude citations.
OpenAI says ChatGPT may search with Bing, and a site must let its OAI-SearchBot crawler read it to be included. (OpenAI Help Center, read 11 September 2026)
Evidence: Median distinct websites per answer: Perplexity 15, Claude 6.5, ChatGPT 3. Claude's figures use its completed searches only.
For a local business, most websites that matter won’t be its own
A prediction we haven’t tested yet
Across our measurements an answer draws on a median of eight websites, from about thirty per question. We expect local businesses to follow that pattern, but we haven’t measured one with web search on yet.