Across engines, not one of them
How to show up in AI search results
There is no single AI search to show up in. We put the same 16 buying questions to three engines and recorded the sources behind every answer. They did not reach for the same things, and one of them cannot be audited from the outside at all.
Same questions. Different sources behind every answer.
16 questions, four industries, one city. Every engine answered all 16 and attached sources to all 16. What they attached was not the same.
ChatGPT returned 84 citations drawn from 63 distinct domains, a median of 5 per answer. Perplexity returned 313 citations drawn from 181 distinct domains, a median of 20 per answer. The two overlap, but neither is a subset of the other.
Source. 16 questions across four industries, Cincinnati · sampled 24 and 25 August 2026 · ChatGPT gpt-5.4, Perplexity sonar, Gemini 3.6-flash, each through its API · counts derived from the stored answers and their cited URLs.
Gemini's links point to Google, not to the publisher
All 142 of Gemini's citations in that sample came back as a single Google redirect host rather than the publisher's own address. Each citation did carry the publisher's domain as its title, so the site is named, but the link itself does not go there.
This matters when buying measurement. A report that lists Gemini's source domains is reading those titles or following the redirects, and it should say which.
What follows for the work
Because the source sets differ, work aimed at one engine does not automatically land in another. A page that is quotable may be reached by one engine and ignored by another that preferred a directory, a government page or a forum thread on the same question.
The order that makes sense is to find out which engines answer your category at all, then which sources they reach for in it, and only then decide what to change. That order is the reason to measure before writing anything.
The parts that are ordinary
Pages have to be reachable, readable without running scripts, and specific enough to answer a question on their own. Facts about a business have to agree wherever they appear, because a machine reading three different opening hours has no way to choose. Google's own guidance for its AI features points at ordinary search fundamentals and does not ask for special AI markup. Google's documentation.
Most of the useful work is unglamorous and already understood. The measurement is what tells you which part of it is missing for one business.
What sampling cannot establish
These counts describe the questions we asked under recorded conditions. They are not market share, not impressions and not buyers. Sampling records answers and does not establish why any business was included. API results are observations under recorded conditions and are not the consumer apps. We do not guarantee citations, rankings, traffic or revenue.
How we measure · How to rank in ChatGPT · Being recommended, not just mentioned · What answer engine optimization is · Google AI Overview not showing up?
Questions we get asked about this.
Is there one AI search index to get into?
No. In our sample the three engines answered the same 16 questions with different source sets. ChatGPT drew on 63 distinct domains and Perplexity on 181, and neither set contained the other.
Does showing up in one engine mean showing up in the others?
Not reliably. The engines reached for different kinds of sources on the same question, so presence in one is not evidence of presence in another. Each engine has to be checked separately.
Can you tell which websites Gemini used?
Yes, from the title on each citation. Every Gemini citation in our sample was returned as a Google redirect address rather than the publisher's own domain, but each one carried the publisher's domain as its title. Any report that lists Gemini's source domains should say whether it read the titles or followed the links.
Do I need special markup for AI to read my site?
Google's published guidance for its AI features points to ordinary search fundamentals and does not require AI-specific markup. Structured data helps a machine read a page reliably, which is worth doing on its own merits, but we do not sell it as a route to being recommended.
How long does any of this take to show up?
We do not know, and neither does anyone else with evidence we would accept. Answers change between runs and between model versions, which is why we measure repeatedly and report the conditions rather than a date.
We publish what we find.
We're measuring AI answers in Cincinnati and will publish what we find, including our own results, including when they're unflattering. No pitch attached.
One email a month at most, and you can leave with one click. We do not sell your address and we will not add you to anything else.