How AI answer engines decide who to cite
When someone asks ChatGPT which CRM to buy, the answer usually names four or five products and links to a handful of pages. Those links are the new page one. If you're not in them, you're not in the conversation.
Nobody outside the labs knows exactly how each engine picks its sources, and the details shift with every model update. What you can do is watch closely. This post covers the patterns we look for when we run buyer questions through the assistants, and a method you can run yourself to see who gets cited in your category.
Two problems: getting retrieved, then getting picked
It helps to know roughly what happens when an assistant answers with sources. When ChatGPT search, Perplexity, Gemini or Google's AI Overviews decide a question needs fresh information, they run one or more searches, pull back a set of candidate pages, read them, and write an answer that cites some of what they read.
That gives you two separate problems:
- Getting retrieved. Your page has to show up in the underlying search and be readable by the engine's crawler. This is mostly classic SEO plus crawler access.
- Getting picked. Out of the pages retrieved, the model has to find yours useful enough to quote or link. This is mostly about how the page is written.
When you're missing from an answer, the first job is working out which of the two failed. A page that never gets retrieved won't be fixed by rewriting it. A page that's retrieved and passed over won't be fixed by more links.
Patterns we watch for in cited pages
These are things we check for, not laws. Treat each one as a hypothesis to test in your own category.
A direct answer near the top. Pages that answer the question in the first two sentences are easier to cite than pages that get there after a story or a table of contents. Models pull fragments, and the cleanest fragment is usually the opening.
Something quotable. A specific number with its source, a clear definition, a short ranked list, a price. If a page has no clean fragment to lift, it can be read and still skipped.
Trust that already exists elsewhere. Cited pages often rank well on Google or Bing for a related query, or come from sites the engines already lean on: review platforms, established publications, Wikipedia, active forum threads. Answer engines don't replace search authority. They borrow it.
Third-party pages that name you. For "best X for Y" questions, the cited source is often a comparison article or review roundup, not a vendor's own site. In those categories, being named on the pages that get cited matters as much as your own page.
Content a crawler can read. Strong pages sometimes go missing because the content only appears after JavaScript runs, sits behind a login or cookie wall, or because robots.txt or a CDN rule blocks the AI crawlers.
Current facts where the question is time-sensitive. For pricing, rankings and anything with a year in the prompt, a visible updated date and current details seem to help. For evergreen definitions, less so.
Things we don't expect to matter on their own: word count, keyword density, and how often you publish. Long pages can do well when they cover more questions, not when they repeat one. A daily blog doesn't help if each post answers nothing new.
How to measure it yourself
You don't need a tool to start. You need a fixed set of questions, a schedule and a spreadsheet.
1. Build the prompt set
Write down the questions buyers actually ask before they buy. Pull them from sales call notes, support tickets, demo request forms and the questions your team answers over and over. Keyword tools help with phrasing, but they won't tell you what people ask an assistant.
Start with 30 to 60 prompts, grouped by intent:
| Intent | Example prompt |
|---|---|
| Category | What does payroll software do for a small business? |
| Shortlist | Best payroll software for a 40-person company in the US |
| Comparison | [Vendor A] vs [Vendor B] for a fully remote team |
| Problem | How do I run payroll for contractors in three states? |
| Brand | Is [your company] a good fit for startups? |
Write prompts the way a person types them, not the way a marketer would. Once you start, keep the wording fixed, so changes in the results reflect the engines and not your edits. Add new prompts as a separate batch.
2. Pick engines and fix the settings
Cover the engines your buyers use. For most B2B categories that's ChatGPT with search on, Perplexity, Gemini, Google's AI Overviews (and AI Mode, if your buyers use it) and Microsoft Copilot. Record the settings each time: logged in or out, location, and which model or mode. Answers change with all three, so keep them the same from round to round.
3. Sample more than once
Assistants rarely give the same answer twice. Run each prompt at least three times per engine per round, in fresh sessions, so one lucky or unlucky answer doesn't skew the picture. Weekly rounds are enough for most teams. Monthly works if you're only setting a baseline.
4. Record the same fields every time
| Field | What to write |
|---|---|
| Date, engine, mode | 2026-10-12, Perplexity, default |
| Prompt ID | So you can group results by intent |
| Brands named | Every company mentioned, in order |
| Cited URLs | Every link or source card shown |
| Cited domains | The domain of each URL, for rollups |
| You mentioned? | Yes or no, and position in the list |
| You cited? | Yes or no, and which URL |
| Notes | Wrong facts, old pricing, odd framing |
5. Roll it up
Two numbers matter most. Mention rate is the share of answers that name you. Citation rate is the share that link to a page you control. Track both by engine and by intent group. Then sort the cited domains column by frequency: the sites that keep showing up are where your category's answers come from.
A worked example
Here's how this might play out for a hypothetical company that sells scheduling software to dental clinics.
They write 40 prompts and run three weekly rounds. A pattern shows up quickly. They're named in most "what is" answers, but rarely in "best scheduling software for dental clinics." The cited domains for those shortlist prompts are mostly two review sites and one industry blog's comparison post. Their own site never appears.
So the shortlist gap isn't a writing problem on their site. It's a presence problem on the pages the engines already trust. The plan becomes: complete and update their profiles on both review sites, and send the blog's author accurate, current details. Separately, their notes column shows one engine quoting an old price. Their pricing page loads prices with JavaScript, so they render the prices in the HTML.
Neither fix would have come from staring at Google rankings.
Reading the results
When you're missing from a prompt, sort it into one of four buckets:
- Not retrieved. Your page doesn't rank for the underlying search, or the crawler can't read it. Fix access and SEO first.
- Retrieved, not picked. You rank, but the answer cites someone else. Put your page next to the cited one. Does theirs answer faster, with something more quotable?
- No page exists. Nothing on your site answers the question directly. Write the page.
- The category runs through third parties. Cited sources are mostly reviews, directories and roundups. Work on being in them.
The bucket tells you which team owns the fix: engineering, content or PR. That's the real value of tracking. It turns "we're not showing up in AI" into a list of specific jobs.
What to do this week
- Pull 30 real buyer questions from sales and support, and write them as plain prompts.
- Set up a spreadsheet with the fields above.
- Run every prompt three times on ChatGPT, Perplexity and Gemini, and check Google for AI Overviews.
- List the ten domains cited most often across all the answers.
- Sort your five most important missing prompts into the four buckets.
- Check your robots.txt and CDN settings so the AI crawlers can reach you.
That's most of what we mean by GEO: write the clearest answer on the internet, then make sure the machines can find it.
