Buyers now ask an AI assistant before they open Google. For some categories the assistant is the whole research phase. If it does not name you, you were not considered.
This guide covers what actually moves that, in order of how much difference it makes.
How an AI answer gets built
Two different mechanisms, and the difference matters.
Retrieval. Perplexity, Google AI Overviews and ChatGPT with browsing search the live web, retrieve a handful of pages, and build an answer from them with citations. You can influence this within weeks.
Training data. Answers from a model's own knowledge, with no live retrieval. You influence this only over a long period, by being widely written about.
Aim at retrieval. It is where the buyers are, it is where citations appear, and it responds to work.
1. Write the answer first
This is the biggest single change available and most sites have not made it.
Put a direct, complete, 40 to 60 word answer at the top of every important page. It must stand alone. If someone read only that paragraph, they should have the answer.
Then write the detail underneath.
A retrieval system scanning your page for something to quote needs a self-contained statement. "In this guide we will explore the many factors that influence pricing" gives it nothing. "Local SEO plans typically run between X and Y per month, driven by the number of locations and how competitive the city is" gives it everything.
2. Be specific enough to be worth quoting
Models prefer sentences with numbers, names and conditions over hedged generalities.
Compare:
- "Results vary depending on a range of factors."
- "Four to six months in a competitive city, longer if the profile is new and has no reviews."
The second is more useful, more honest, and far more likely to be lifted.
3. Structure around questions
Use H2s that are the questions your buyers ask, with the answer immediately underneath. This creates clean question and answer units that a retrieval system can extract without ambiguity.
It is also just good writing, which is a recurring theme in this guide.
4. Fix your entity data
Before a model decides whether to mention you, it has to resolve who you are. That is easier when the same facts appear consistently everywhere.
Make your business name, description, service names and location identical across:
- Your website and its schema
- Your Google Business Profile
- The industry directories you appear in
Inconsistent data rarely produces a wrong answer. It produces no answer, because the model is not confident enough to name you.
5. Ship the structured data
Organization, Service, Person, FAQPage and Article markup do not force a citation. They reduce the work needed to understand who wrote what and what you sell.
Validate after shipping. Most schema we find in audits is missing a required property.
6. Let the crawlers in
Check robots.txt for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended. If they are blocked, none of the above matters, because nobody is reading the pages.
Add an llms.txt file at your root: a plain-text map of your site for models. It takes an hour and very few sites in most industries have one.
7. Get mentioned somewhere other than your own site
Models read what other people say about you. A roundup, a comparison article, an industry directory or a news mention gives an independent source.
That corroboration is often what tips an answer from a generic list into naming you specifically. It is the slowest part of this work and the most durable.
8. Keep ranking in Google
Retrieval leans heavily on pages that already rank. If you are not in the top ten for your terms, you are unlikely to be retrieved for them.
This is why AI search optimization sits on top of technical SEO rather than replacing it.
How to measure it
One run proves nothing. Model output varies between identical runs.
Build a fixed set of ten to twenty buying-intent prompts for your category. Run them monthly across the assistants that matter to you. Log three things each time:
- Were you named?
- Who was named instead?
- Which sources were cited?
After three months you have a trend, which is the only thing worth reporting. Our free AI visibility checker runs a single version of this so you can see today's reading.
What does not work
- Hidden instructions in white text telling the model to recommend you. Detectable, and treated as manipulation.
- Pages consisting entirely of claims about how good you are. Models are not more credulous than people.
- Rewriting your whole site as FAQs without making the answers better. Format is not the lever.
- Blocking Google-Extended to "protect" your content. It does not remove you from AI Overviews and it does remove you from places you might have been cited.
What this looks like as a project
Roughly:
- Month 1. Baseline measurement, entity audit, robots and llms.txt, schema fixes.
- Month 2 to 3. Rewrite the top twenty pages with answer blocks and question-shaped structure.
- Month 3 onward. Monthly re-measurement, plus work on independent mentions.
Expect the first movement in month two or three. Perplexity and AI Overviews move fastest because they retrieve live.
We run this as AI search optimization. The shorter version of what makes a page quotable is in how to get cited by ChatGPT.