Getting cited and recommended · · By Gerald De Oleo
What Sources Does AI Cite About You? How to Find Them
The short answer: it depends on the engine and the question, which is why borrowed percentages are useless here. What matters is which sources shape your answers — and you can find that out in about an hour.
Here’s the method, and what to do with what you find.
Find your own source list
1. Write 20–30 questions your buyer asks before they know you exist.
Real constraints attached: their CRM, team size, industry, sales cycle, compliance requirements. Pull the phrasing from discovery calls and support tickets, not from your homepage.
Never include your brand name. Asking “is [company] good?” introduces the brand yourself and teaches you nothing.
2. Run them across ChatGPT, Perplexity, Gemini and Google’s AI results.
Separately. They don’t share a source pool and they frequently disagree.
3. Log the sources, not just the answer.
For every response, record every cited URL and what type it is: your site, a competitor’s, a comparison list, a review platform, a community thread, editorial, documentation.
4. Sort by frequency.
The domains at the top are writing your category description. Most of them won’t be yours.
5. Read the top ten properly.
Check how you’re described in each, whether you’re mentioned at all, and which competitors are. The finding is usually not absence — it’s an accurate-sounding description of the company you were two years ago.
That sorted list is your work plan. Everything below is what to do with it.
What tends to show up
Patterns I see across B2B SaaS audits, with the caveat that your category will differ.
Community threads carry more weight than most teams expect. Reddit in particular. Not because anyone optimized for it, but because it’s where people answer each other honestly.
Review directories carry less than most teams expect. They matter — as an eligibility check rather than a cited source, which I wrote about in Do G2 and Capterra Reviews Get You Recommended by AI.
Your competitors wrote the buyer’s guides. In most categories, the “best [category] software” pages the models read were published by vendors selling in that category. When I looked at marketing attribution, most of the ranking guides were written by attribution companies. That’s covered here.
Your own pages matter, but only the specific ones. Integration pages, comparisons, pricing with real numbers, customer results. Generic feature pages get cited by nobody.
Getting into each type
Ordered by how fast you’ll see anything.
Your own site — fastest, fully in your control. The pages that earn citations are specific: one page per major integration, honest comparisons, an alternatives page, pricing with context, outcomes with numbers. Also confirm AI crawlers can actually read your pages — most don’t run JavaScript, which is the first thing I check in any audit. Details in How to Appear in ChatGPT Answers.
Review platforms — moderate effort, mostly eligibility value. Be present, be in the right category, keep the description current.
Third-party lists you didn’t write — slow, high value. Independent roundups, category pages, industry publications. Many accept vendor submissions or corrections. Being named in a list you don’t control is what makes a model believe your description, because agreement across independent sources is the signal these systems lean on.
Community — slowest, highest ceiling, easiest to get wrong. Participating honestly where your category is discussed. Seeding threads is detectable and both moderators and models handle it badly. There’s no shortcut here worth the risk.
Original data — the most durable. A benchmark or study only you can publish, because it comes from your product. It’s the one source type where other people’s articles need your number. More on that in Your Product Has Data Nobody Else Can Publish.
What doesn’t work
llms.txt. Heavily marketed, and Google has said plainly that no AI system uses it. Genuinely useful if you sell a developer product and want coding assistants to pull your docs correctly. Not a lever for marketing pages.
Hidden instructions for models. Text telling an engine to recommend you, or content served to crawlers but hidden from users. Ignored at best, a credibility risk at worst.
Keyword stuffing. Worse than doing nothing, by the research that exists on this.
Dozens of thin comparison pages. One honest comparison per real competitor beats fifty templated ones.
Publishing before the entity is fixed. If sources disagree about what your company is, more content amplifies the disagreement. Fix the description first.
The part people miss
Your own pages being cited isn’t the finish line. A page can be cited while the model pulls the wrong claim from it — the integration you mention in passing instead of the product you sell.
So when you read your results, check the exact sentence the engine used about you. Not whether your domain appeared. That gap is the whole premise of the first post on this blog.
Quick answers
How do I find which sources influence my AI answers? Run 20–30 unbranded buyer questions across each engine, log every cited URL, and sort by frequency. The most-cited domains are shaping your category description.
Does Reddit matter for AI citations? In most B2B categories, more than the review directories do. Its weight varies by category, and seeding isn’t a viable strategy.
Why do published studies give such different numbers? They measure different engines, different question types and different definitions of a citation. Use them for direction, not for targets.
Is being cited the same as being recommended? No. Most companies named in an AI answer don’t get the citation pointing at their own site.
If you’d rather see the source map for your category than build it by hand, that’s the first thing the audit produces: every question, every engine, every cited page, and how you’re described in each. Before we talk, I run a free Shortlist Check — three or four of your buyers’ questions in one engine — so you can see whether the full audit is worth it. See how the audit works, or book a call.