Who does Perplexity recommend for cold outreach? We asked 50 buyer questions.

We are about to launch a cold outreach tool. Before sending a single email for it, we wanted an answer to an uncomfortable question: when a solo founder asks an AI assistant what to use, who gets named, and what would it take to be in that answer?

So we measured one engine properly instead of vibing about all of them. Fifty questions a real buyer would ask, sent word for word to Perplexity's retrieval-grounded model, no steering, no system prompt, everything recorded. One engine, one day. Other engines will weight sources differently, which is exactly why the method matters more than any single number here.

How we asked, exactly

The 50 prompts cover ten intents, because buyers do not ask one kind of question: category discovery, problems, how-tos, comparisons, alternatives, cost, validation, trust, use cases, and five branded lookups about our own product. A tool counts once per answer if its name appears in the response, however many times it is repeated. Sources are the URLs cited by the answer, deduplicated within each answer; we report domains both as total citation instances and as answers containing at least one. The full prompt set is at the end of this page, rendered from the same frozen file the benchmark script runs, and the raw answers are available on request.

Eight names, over and over

The concentration is the story. Fifty questions, shaped ten different ways, and the same eight names keep coming back. Ask about cost: Reply.io. Ask about finding addresses: Hunter, Apollo. Ask what a bootstrapper should do instead of hiring an SDR: Perplexity lists founder-led outbound, AI tools, and fractional SDRs, then names the same names.

If you build in this category, this is your real competitive set. Not the tools you compare yourself to in your head. The ones the machine reaches for when nobody is watching.

Reddit runs the answers

Reddit was cited in 23 of 50 answers, 37 citations in all: the most-cited domain in our run by either measure, ahead of every vendor. When the engine wants evidence beyond vendor claims, it keeps reaching for people arguing about the category in public.

And when vendor pages do get cited, they are not homepages. We classified every cited vendor URL: all 12 of Saleshandy's citations point at its blog's comparison and how-to guides, zero at its homepage; 9 of Instantly's 10 are the same shape, including pieces comparing themselves with named competitors. The engine wants liftable facts; an honest comparison page supplies them even when a vendor wrote it, and a homepage supplies none.

What Perplexity told our test founder

These are not our benchmarks. They are the advice the engine surfaced during the run, from its own sources, and we report them as a snapshot of what every founder now hears before reaching any vendor site:

That cold email still works, citing an average reply rate of about 3.4%. That a new domain should send 10 to 20 emails a day per inbox after warmup; for what it is worth, that is the range our own defaults enforce, a coincidence we enjoyed. That you should never send cold email from your main domain, because mistakes land on the domain that sends them. And that a full DIY stack, sender plus data plus verification plus warmup, runs $100 to $200 a month, which is the strongest argument we have seen for collapsing the stack into one tool, said knowing exactly how biased we are.

We got beaten by Google Tag Manager

Now the embarrassing part, published on purpose. Our product is called GTM for One (gtmforone.com). Asked "What is GTM for One?", Perplexity answered about Google Tag Manager. Asked if we were legit, it reviewed ONEGTMLAB instead, a go-to-market agency, a different company in a different business: they build and run your GTM as a service, we sell software you run yourself.

Zero of our five branded questions came back right. The lesson for anyone naming a product near an acronym the internet already owns: you will spend your early life disambiguating, and the fix is discipline, not cleverness. Full name and domain together, in the same breath, everywhere the product is mentioned. You are reading that discipline right now.

What we can't tell you

One engine, one run, fifty prompts, September 2026. These numbers describe Perplexity's answer set on that day, not "AI" in general; an engine with different retrieval will weight sources differently. Mention detection is string matching, so a tool described but never named would slip past. And we build a product in this category, so we are not neutral: the prompt set below, the raw answers and the scoring script are versioned, and the run costs about thirty cents to reproduce.

We will rerun the same fifty prompts every month and publish the deltas. Today the answer set belongs to eight tools and a great many Reddit threads. The interesting question is how long that stays true, and the point of freezing the prompts is that the numbers can argue with each other over time.

The 50 prompts

Rendered from the same frozen file the benchmark runs, so what you read is what was asked: