How to appear in answers from ChatGPT, Gemini and Perplexity

The principle: write so the answer can be cut out

A model composes an answer from fragments. It needs a piece of text that answers the question on its own — without the context of the whole page.

Compare two versions of the same thing:

The first text contains not a single quotable fact. The second contains four.

Six practical steps

1. Answer the question in the first paragraphDo not take three sentences to warm up. The opening lines of the page must contain a direct answer to its main question.
2. State facts explicitlyPrices, timing, warranties, conditions, limitations — in figures and words, not by implication.
3. Add structurePhrase headings as questions and keep answers short. Lists and tables are extracted better than solid text.
4. Add a questions-and-answers blockThis is the most convenient format for a model: question — short answer. Use the real questions customers ask.
5. Mark up the dataStructured markup of the organisation, services, prices and questions helps the machine understand what is what.
6. Work with mentionsModels rely on sources: directories, industry sites, reviews, articles. A company that only writes about itself looks weaker.

Which pages appear in answers more often

• Service pages with prices and scope of work — there is something to quote there.
• Comparisons and breakdowns — «how X differs from Y», «what to choose in this situation».
• Answers to specific questions — what people actually ask, not what is convenient to write.
• Pages with limitations — «who this is not for» builds trust with both people and models.
• Local information — addresses, service areas, opening hours, city specifics.

What not to do

• Trying to «optimise for AI» with keywords — this is not keyword search.
• Hiding text for robots: models read what a person reads.
• Writing walls of text without structure and hoping the model will sort it out.
• Promising what you do not have: the mismatch is exposed by cross-checking sources.
• Blocking AI crawlers and then expecting to be cited.

How to check whether you are named

1. Draw up a list of questions10–15 phrasings a customer might come to you with. As questions, not as keywords.
2. Ask the modelsChatGPT, Gemini, Perplexity. Where possible, in a mode with web search.
3. Record the resultWho is named, which sources are cited, whether you appear at all.
4. Study the winnersOpen the pages the model cited. It is almost always visible how they differ from yours.
5. Repeat once a quarterThat way you see the trend and the effect of your changes.

This simple procedure takes an hour and gives more understanding than any general speculation about the future of AI.

How to rewrite a service page for extraction

Let us look at an example. Before: «Our clinic offers modern cavity treatment using advanced materials and an individual approach to every patient».

After — the same meaning, but quotable:

• What it is. Cavity treatment — removal of damaged tissue and placement of a filling in a single appointment.
• What it costs. From $40 for a surface cavity, from $70 for a medium one, from $110 for a deep one.
• How long it takes. 40–60 minutes per tooth.
• What is included. Examination, anaesthesia, treatment, polishing, follow-up X-ray.
• Warranty. Two years on the filling if the aftercare instructions are followed.
• Who it is not for. Acute pulpitis requires different treatment.

The same amount of text, but the page now carries six facts, each of which can become part of an answer. And it is clearer to people too.

Working with sources outside your site

A model rarely relies on a single source. It cross-checks what it found in several places and prefers a consistent picture.

1. Directories and mapsThe same data in every service. A discrepancy in phone number or hours lowers trust in all sources at once.
2. Review platformsGenuine reviews give the model material to judge by. A company with no reviews looks insignificant.
3. Industry resourcesA mention in a trade publication or catalogue weighs more than ten links from throwaway sites.
4. Your own materialsAnalyses and articles that others link to are the most durable source.

Common mistakes

• Rewriting one page and expecting results across every query.
• Adding facts that contradict what is listed on the maps.
• Copying a competitor's questions block — the model sees a duplicate and prefers the original.
• Quoting «from» prices without an upper bound where the range is tenfold: such a fact is useless.
• Blocking AI crawlers and waiting for mentions.