How does ChatGPT pick who to name?
ChatGPT runs searches of its own, reads a small set of the pages it retrieves, and writes a short answer naming the companies those sources describe clearly. The answer is usually built from a handful of cited sources, so the businesses inside those sources are the ones it names, and everyone else is left out.
Traditional search runs a tournament: pages compete for a keyword, an algorithm scores the field, and the winners are displayed in order. An answer engine is doing something different. It is writing a document, on deadline, and looking for sources it can quote without embarrassment.
That difference changes the scoreboard. There is no eleventh place inside an answer: a source is either cited or it is not, so the return behaves like a threshold rather than a curve. It also decouples the new contest from the old one. Ranking and citation correlate, but they are not the same contest (GEO vs SEO sets out the split). Google says its AI features display "a wider and more diverse set of helpful links" than a classic web search (Google Search Central), so a well-structured page outside the top results can still be cited. Which is why years of decent SEO can coexist with total absence from the answers, and why absence from an answer is not a lower rank. It is a zero. The full argument is in Ranked vs. Cited, a chapter of The Screenshot, the free book (available after email verification).
Which engines and sources build the answer?
Six engines matter for buying questions: ChatGPT, Perplexity, Gemini, Google's AI Overviews, Copilot and Claude. Each finds sources its own way and mixes live web retrieval with what it learned in training, so the same buyer question can name different companies on different engines.
Training data is the accumulated text of the public web as of some cutoff, which is why engines sometimes describe companies in confidently outdated terms. Working across all of them at once, judged by what each one actually says, is the job of an LLM SEO agency.
Two behaviors matter most for who gets named. First, most of what assistants cite is earned media rather than a company's own pages: in Muck Rack's May 2026 study of more than 25 million links cited by ChatGPT, Claude and Gemini, earned media made up about 84% of citations and paid content 0.3% (Muck Rack), so the record other people publish about you often counts more than the record you publish about yourself; repairing that outside record is the AI PR agency lane. Second, Google says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources to develop a response (Google Search Central), so one page or cluster that genuinely covers a territory beats ten shallow pages targeting ten keywords. The per-engine field guide is the book's chapter The Six Engines and Who Each One Trusts.
Can the engines even read your site?
Only if their crawlers can fetch your pages. Each engine runs separate agents for search, for visits a user requests and for model training, and a robots.txt block on the search or user agent can keep your pages out of that engine's answers. Check those rules first, because every other repair depends on access.
Search crawlers include OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot and Bingbot. ChatGPT-User, Perplexity-User and Claude-User handle user-initiated retrieval. GPTBot, ClaudeBot and CCBot are model-development crawlers, while Google-Extended is a separate control that does not govern Google Search, AI Overviews or AI Mode. Each vendor documents its own agents: OpenAI, Anthropic, Perplexity and Google. A block can prevent retrieval through one lane. Landing that fix, and the ones after it, in production is GEO implementation.
This is not rare. On 20 September 2026 Suede AI requested robots.txt and llms.txt from 1,000 top domains on a pinned Tranco list: of the 497 robots.txt responses it could evaluate, 138 (27.8%) disallow at least one of five AI crawler tokens at the root, and 78 llms.txt files passed its format checks, per the AI Crawler Access Index, with the per-domain data and the script published alongside it.
Access is only the gate. Past it comes extraction: AI systems extract passages, not pages. When an engine cites you it lifts a sentence to a short paragraph and builds its answer from that, so the unit of AI visibility is the passage. The test is simple: pulled out alone, with no surrounding context, does a paragraph from your site still say something true, complete, and attributable? Most marketing copy fails instantly: "we take a fundamentally different approach" extracts to nothing. The repair manual for this is the book's chapter Extractable or Invisible.
Why do receipts beat claims?
Because engines are tuned to cite sources they can trust, and evidence is harder to fake than adjectives. The research that named the field, GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024), found that targeted content changes can raise a source's visibility in generative engine responses by up to 40%.
The web is now flooded with plausible machine-written text, and citing something wrong or fake embarrasses an engine in front of its user. So engines are tuned, and continuously re-tuned, to prefer sources that look like they know what they are talking about. Machines evaluate trust the way skeptical buyers do. Adjectives are free and infinitely forgeable; any text generator can produce "industry-leading." Receipts are expensive to fake: a named author who verifiably exists, a specific number with a date and a methodology, a real result someone can open, a document trail third parties corroborate. If your competitors' pages carry receipts and yours carry adjectives, the engine has its reason. The book's chapter on it is Receipts Beat Claims; this site holds itself to the same standard on the receipts page.
How can you see your own gap in an afternoon?
Put ten to twenty real buyer questions to ChatGPT, Perplexity, Gemini and Google's AI on more than one day, screenshot every answer, and record whether you were named, whether you were described correctly, and which sources were cited. That cited-source list is your repair map.
The inputs are the part you control. Repair the pages, records and sources an engine reads, then measure again.
You do not need a vendor to see the gap. Run the question your actual buyer actually asks, "best tool in your category for your exact customer", through ChatGPT and Perplexity and read what comes back. If two competitors are named and you are not, you have your screenshot, and nobody argues with the screenshot. To turn the anecdote into a measurement, do what the audit does (the generative engine optimization method sets out the sampling):
- Write ten to twenty real buyer questions. The free AI search question builder turns a service and a city into a ready set to copy.
- Run them across engines on more than one day.
- Screenshot every answer.
- Record two columns: were you named, and was it correct.
- Record which sources each answer cited. That cited-source list is usually the most direct map of where the repair work should go.
The full afternoon version is the book's chapter The Founder's Visibility Audit. For a first reading on four of those questions, the free AI citation check takes one for you.
How do you get ChatGPT to recommend your business?
Get into the sources ChatGPT reads and trusts: let its crawlers fetch your pages, state what you do in plain quotable text, back claims with evidence, and keep your business records consistent. Then measure again on the same questions.
Run the audit yourself with The Founder's AI Visibility Checklist, the five lanes in severity order, free and printable. The reasoning behind every lane is The Screenshot, an eleven-chapter book, free on this site after email verification. And if you would rather have the measurement and the repairs run for you by a GEO agency, with dated captures before and after, start with the free AI visibility assessment or book a 30-minute call.