I Looked at Every URL Perplexity Cited for 6 Buying Questions. 14 of 23 Belonged to a Vendor.

Our own database has been storing the URLs that AI engines cite since June. Every visibility check writes them to a sources column, and until recently nothing read that column. When I finally went looking, I found 236 stored URLs, and then spent an afternoon working out that most of them were worthless.

Here is the part I'd rather not publish. Of those 236 URLs, 213 came from checks we'd labelled chatgpt, which ran through a model with no web grounding. We pull URLs out of the answer text with a regex. For a grounded engine that gives you real citations. For an ungrounded model it gives you URLs the model remembered from training, which is why those rows are full of bare homepages: later.com, hootsuite.com, ads.reddit.com. They aren't citations. They're recall.

That leaves 23 URLs from grounded Perplexity checks, spread across 6 questions and 2 products, collected between 17 June and 9 July 2026. It's a tiny sample. I decided to write it up anyway, because the shape of it contradicted what I expected and I'd rather publish 23 real citations than 200 fake ones.

The counts

I ran the 23 URLs through our own classifier, the same code that labels sources in the product:

Then I went through them by hand, because the classifier answers "what shape is this page" and the question I actually cared about was "who owns it". Fourteen of the 23 were on the domain of a company selling a tool in the category the question asked about. Not a journalist writing about the category. The vendors themselves, describing themselves, quoted back to the buyer as an answer.

Four were written by someone who wasn't selling anything: a Reddit thread, an Indie Hackers thread, a Medium post by an indie developer, and a LinkedIn article. The rest were blogs belonging to companies adjacent to the question, like Ahrefs and Semrush writing about brand monitoring, or Mercury writing about marketing tools for startups.

The full list

For "How do I monitor when AI assistants cite my competitors but not me?":

• ahrefs.com/blog/brand-mentions/
• radarkit.ai/blog/monitor-competitor-mentions-in-ai-search/
• turboaudit.ai/ai-monitoring
• reddit.com/r/ProductMarketing/comments/1k10tlt/top_5_tools_to_monitor_your_brands_presence_in_ai/
• semrush.com/blog/ai-visibility-toolkit/

For "How do I monitor Reddit for high-intent buyer conversations in my niche?":

• catchintent.com/blog/how-to-find-buyer-intent-signals-reddit/
• catchintent.com/social-listening/
• catchintent.com/tools/reddit-intent-analyzer/
• linkeddit.com/blog/reddit-buying-intent-signals-guide
• relato.com/blog/reddit-monitor-agent/

For "What's the best tool for finding potential customers on Reddit who are actively looking to buy?":

• linkeddit.com
• linkeddit.com/blog/how-to-find-customers-on-reddit
• ogtool.com
• replyagent.ai

For "Best tools for indie hackers to find paying customers through Reddit marketing?":

• medium.com/@faustoct1/side-tools-how-indie-hackers-promote-their-products-for-free-indefinitely-ebd64f0cf3b8
• redditgrow.ai/blog/best-reddit-tool-indie-hackers-2026
• indiehackers.com/post/as-a-solo-indie-maker-what-are-your-go-to-tools-for-building-and-growing-your-product-380ee3dffb
• reddit-radar-marketing.com/use-cases/indie-hackers

For "digital marketing for startups":

• bolsterbiz.com/blog/10-best-digital-marketing-strategies-for-startups/
• mercury.com/blog/best-marketing-automation-tools-startups
• startup.google.com/products/
• linkedin.com/pulse/best-free-marketing-tools-startups-diginext-info-xybsc
• usekaya.com/blog/budget-friendly-marketing-tools-for-startups

Notice the third question. Three of its four citations are a competitor's homepage or a page one click from it. linkeddit.com and ogtool.com and replyagent.ai are small tools with the kind of domain authority that wouldn't get them onto page one of Google for that query. The engine cited them anyway.

What I think this means, with the caveat that it's 23 URLs

None of this says organic search stopped mattering, which is a separate argument I have made at length. The advice I keep reading is that AI visibility comes from being on the big third-party properties: get on G2, get into the listicles, get mentioned on Reddit. A year later I ran the same exercise at ten times the size and got a different answer, which is the 533-URL corpus study. In this sample the big properties were absent. No G2. No Capterra. No Wikipedia. No product comparison pages. One Reddit thread out of 23.

What the engine reached for instead was whichever page most directly answered the question, and a lot of the time that page belonged to a small vendor who'd bothered to write it. catchintent.com got cited three times for one question, on three different pages of their own site. That is a company that wrote about buyer intent signals on Reddit and got quoted for it.

If that holds up at larger sample sizes, the first move for a small product isn't outreach. It's writing the page. Which is convenient for me to conclude, so treat it with the suspicion it deserves.

The Reddit finding is the one I keep chewing on. We spent a year building around the idea that Reddit threads are what AI engines quote when buyers ask for tools, and I had spent a lot of that year reading what a buying question actually looks like on there. One citation out of 23 is not that. It's not zero either, and the one that appeared was a thread titled "Top 5 tools to monitor your brand's presence in AI", which is a listicle that happens to live on Reddit.

How the numbers were produced

Since the whole point of this post is that most published claims in this space have nothing behind them, here is enough detail to argue with mine.

The questions came from our own product database rather than from a keyword tool. They are questions attached to real tracked products, phrased the way a buyer would type them, which is why they read as long and awkward. Six of them, across two products.

Each check ran through Perplexity's grounded API, one question per call, no system prompt beyond an instruction to answer the question. We stored the answer text and the returned source list. The counts in this post come from the source list, not from URLs appearing in the answer text, which matters because those two things are not the same and mixing them is exactly the mistake described below.

"Unique URLs" means deduplicated by full path, not by host. catchintent.com contributing three different pages counts as three URLs and one host, which is why the vendor-domain share (14 of 23) is higher than the number of distinct vendors involved.

The date range is 17 June to 9 July 2026, which is a period in which we were not doing anything to change our own visibility, so nothing here reflects an intervention. It is a passive read.

The classifier that produced the type table is the same code that labels sources in the product. It works on URL shape and page structure, and as noted below it got at least one of these wrong.

What I would do differently

Four things, if you are planning to run this yourself.

More questions. Six is not enough for the counts to be stable. On a set this small, one question that happens to be about a well-covered topic moves a whole row. Thirty to fifty is where I would start.

More than one engine. Perplexity cites openly, which is precisely why it is the engine everybody measures, including us. It is also much smaller by usage than ChatGPT. A finding drawn only from Perplexity describes Perplexity.

Repeat runs. I ran each question once. Retrieval moves between runs, sometimes substantially, so some fraction of what I have counted as a stable pattern may be one day's noise. Two or three passes spread across a couple of weeks would separate those.

Categories I do not sell into. Every question here comes from a market I have a product in, which means the questions are shaped by how I think about the category and the results are shaped by pages written by people I compete with. That is close to the worst possible sampling frame for a general claim.

Running this on your own category

The manual version takes an afternoon and does not need our product.

Write fifteen to thirty buyer questions as full sentences, with constraints in them, pulled from sales calls and support tickets rather than keyword tools. Run each through a grounded engine, twice, a few days apart. For every answer, record the question, the run date, each cited URL, its host, and what kind of page it is. Then pivot on host and count.

What you are reading is the shape of the distribution rather than any single row. A handful of hosts carrying half the citations means you have gatekeepers and a finite list of places to get mentioned. Citations scattered across forty hosts with no repeats means there is nothing to pitch and writing your own answer pages is the better use of the quarter.

One rule that decides whether any of it is real: only count citations from an engine that visibly searched and showed you a source panel. If there is no link panel, there is no citation, whatever the text claims.

Two things this exercise broke

The classifier put the Indie Hackers thread in the listicle bucket. Its URL contains tools-for, which matches a listicle hint, and indiehackers.com isn't in our list of community hosts. So a community thread got labelled as a ranked article. Both are now fixed in our backlog.

The bigger one is that we were storing ungrounded model output in the same table as real citations, with no column saying which was which. Anyone reading that table, including me, would have drawn conclusions from 213 hallucinated URLs. The sources column now only means something when the engine was grounded, and we should have written that down when we built it.

If you want the same read on your own category, our visibility checker runs the questions and shows you what got cited. The methodology page covers how we ask, what we extract, and what the method can't see, including this exact limitation.

I'd like to redo this with 50 questions across categories I'm not selling into. When I do, I'll publish the numbers whether or not they agree with this post.