The shortlist has moved into the answer.
A marketing leader who needs a new partner used to search, open five tabs and compare. Many now ask an assistant a single question, such as “which firms help B2B software companies with this?”, and start from the three or four names it returns.
That shortlist is assembled before anyone visits your site. If your brand is absent, your homepage, your case studies and your pricing never get the chance to persuade. The work therefore starts outside your own pages.
This guide explains how the engines decide whom to name, what the current evidence says moves that decision, and the order in which to act.
Engines read what the web says about you.
When a buyer asks for recommendations, assistants with web search look for pages that already compare providers: directories, “best of” lists, review platforms and forum threads. In a study of 31 “best agencies for…” questions, ChatGPT with search drew 57% of its citations from the firms’ own websites, while Google’s AI Mode and Gemini drew most of theirs from “best of” lists (Onely, October 2026).
The practical reading is simple. Your own site has to state clearly what you do and for whom, because one engine quotes it directly. And you need to be present on the independent pages the other engines trust, because that is where they look first.
- 01The buyer asks a comparison question
- 02The engine fans out into sub-questions
- 03It retrieves lists, reviews and your pages
- 04It names the brands that recur
A simplified model of observed behaviour, not a documented algorithm.
Each engine sees a different web.
The engines barely agree with one another. Across 3.7 million URL citations in ChatGPT, Perplexity and Google AI Overviews, 91% appeared in only one engine, and only about 2.4% appeared in all three (Growth Memo, May 2026).
Two consequences follow. A single test in one assistant tells you very little, so any visibility check has to cover several engines and be repeated. And winning one engine does not carry over to the next; each needs its own evidence of where you stand.
This is why a credible baseline records the exact questions asked, the engines used, the dates of each run and the brands named each time, rather than a single screenshot.
Being cited is not the same as being named.
An engine can link to your page and still never say your name. In a study of 3,981 appearances in AI answers, 61.7% were citations without a mention, while only 13.2% were both cited and named (Semrush and Growth Memo, June 2026).
The same study found that comparison questions produced brand mentions 2.4 times as often as purely informational ones. If you want to be named, publish pages that compare approaches and providers honestly, and say clearly who you are within them.
Where the answer comes from on a page.
Position on the page matters. Published analyses of ChatGPT citations find that a disproportionate share of cited passages come from the opening third of a page, and that definitions buried deep in a page are retrieved far less often.
Freshness matters as well. Of 7,683 pages cited by ChatGPT, Gemini and Perplexity between March and June 2026, three in four had been updated within the previous year, and most had not been newly published (Seer Interactive, July 2026).
| Habit | What it looks like |
|---|---|
| Answer first | The direct answer in the opening paragraphs, detail afterwards. |
| One idea per section | A clear heading, then two to four short paragraphs that stand alone. |
| Dated evidence | Every number with its source and period, so it can be quoted safely. |
| Visible updates | A real “updated” date, changed only when the content changes. |
What does not move the needle.
Some popular fixes have weak evidence behind them. An llms.txt file is the clearest example: across 137,210 domains, 97% of published files received no requests at all in the month studied (Ahrefs, June 2026). It costs little to publish, but it is not a visibility strategy.
Google states that there are no additional requirements to appear in its AI features beyond being indexed and eligible for a snippet, and no special markup or AI text files are needed (Google Search Central). Structured data still helps engines understand who you are, but it is not a shortcut to being recommended.
Be cautious of anyone who promises a position in AI answers. The engines change, they disagree, and no one controls them.
A practical order of work.
Start with a baseline: a sample of the questions your buyers actually ask, run through the main AI engines, with the date and every brand named. Without it, you cannot tell whether anything you change makes a difference.
Then fix the independent sources. Make sure you are listed, accurately described and reviewed on the directories and comparison pages the engines read for your category. Next, rewrite your key service pages so the answer comes first and every claim is dated and sourced.
Finally, publish comparison content that helps a buyer choose, and refresh it on a schedule. Re-run the same questions monthly and record what changed. For the measurement side, our guide on how to check whether your business appears in AI answers sets out the method step by step, and our AI search visibility service covers the work end to end. If you are weighing where this sits in your budget, see AEO, GEO and SEO: a leader’s guide to the investment decision.
References and definitions
Cited means the engine links to your page as a source. Named means it says your brand in the answer text. Query fan-out is the engine’s practice of splitting one question into several sub-questions before retrieving sources.
Sources: Google Search Central, AI features and your website; Semrush, the ghost citations study; Growth Memo, the consensus gap; Onely, how ChatGPT decides which brands to recommend; Seer Interactive, content recency and AI visibility; Ahrefs, llms.txt study. Figures are as published by each source and were checked on 6 November 2026; several are vendor studies.