Visibility in ChatGPT, Perplexity and Copilot

How Assistants Source Answers

There are two mechanisms, and they behave very differently.

Retrieval — the assistant searches the live web, reads results, and synthesises with citations. Perplexity works this way by default; ChatGPT and Copilot do so when browsing is invoked. Here, conventional search visibility is the dominant input, because the assistant is reading search results.

Parametric recall — the model answers from what it absorbed during training, with no live lookup. Here, visibility depends on how often and how consistently your brand appeared in the training corpus, which you cannot influence quickly and cannot influence at all retroactively.

What Drives Retrieval Visibility

Because retrieval reads search results, standard SEO does most of the work. On top of that:

• Content structured so that individual passages answer questions completely
• Presence on the pages assistants tend to consult for your category — comparison articles, roundups, reputable review sites, established community threads
• Being described consistently across the web, so retrieval across several sources reinforces rather than confuses
• Not blocking retrieval crawlers in robots.txt, which silently removes you from consideration

What Drives Parametric Visibility

This is slower and closer to brand building than to SEO. What appears to help: sustained mention across many independent sources, presence in the reference works models train heavily on, a clear and consistent description of what the company does, and longevity.

What does not help: anything that looks like manipulation. Content farms producing repetitive brand mentions are exactly what training data curation is designed to filter.

Recognise the lag. Influence exerted now affects models trained later. This is a multi-year exercise.

Testing Your Visibility

There is no analytics dashboard for this yet, so test manually and systematically.

Build a list of the 20 to 30 questions a prospective customer might ask an assistant in your category. Run them across ChatGPT, Perplexity, Copilot and Gemini. Record whether you are mentioned, whether the description is accurate, whether you are cited with a link, and who is named instead of you.

Repeat quarterly. The absolute results matter less than the direction of travel and than which competitors consistently appear.

Correcting Misrepresentation

If assistants describe you inaccurately, the cause is usually inaccurate or outdated source material on the web rather than a fault in the model.

Fix it at source: update your own site so the correct description is unambiguous, correct third-party profiles and directory listings, and get accurate coverage published where it will be retrieved. Retrieval-based assistants will reflect the correction within days. Parametric recall will not update until the next training cycle.

What You Can Control, and What You Cannot

Visibility in an assistant is not a ranking you can optimise in the way a search result is, and being precise about the boundary saves a great deal of wasted effort.

What you can influence: whether your content is crawlable by the relevant agents; whether it is structured so an answer can be extracted without the surrounding page; whether your claims are specific and checkable rather than hedged; and whether other sources that assistants rely on — reference sites, industry publications, community discussions — say accurate things about you.

What you cannot: which sources a given assistant draws on, how it weights them, whether it cites at all, and what it says about you when it does not cite. Different assistants make different choices and change them without notice.

The practical consequence is that the durable work is the same work that makes a page good: accurate, specific, structured and current. A team being sold assistant optimisation as a distinct practice with distinct tooling should ask which part is not already in their content brief.

Monitoring Without a Rank Tracker

There is no console for this, so monitoring is manual and should be treated as a scheduled task rather than a tool purchase.

Write down twenty questions a prospective customer would actually ask in your category, including your brand name and your competitors'.

Ask them across the assistants your audience uses, monthly, from a clean session so personalisation does not colour the result.

Record three things: whether you appeared, whether what was said was accurate, and who was cited instead.

That third column is the useful one. It tells you which sources the assistants trust in your category, and being accurately represented on those sources is more tractable than trying to influence the assistant directly.

Correct inaccuracies at the source. If an assistant repeats something wrong about your pricing or your product, the fix is usually a stale page somewhere — yours or a third party's — rather than anything to do with the assistant.

Sources

What each claim on this page rests on. Entries are typed so you can see which are primary.